Papers with perplexity
Identifying and Reducing Gender Bias in Word-Level Language Models (N19-3)
Copied to clipboard
| Challenge: | Existing discriminatory biases in training data can be amplified by models . text corpora exhibit socially problematic biase . |
| Approach: | They propose a metric to measure gender bias and a regularization loss term to minimize embeddings onto an embeddable subspace that encodes gender. |
| Outcome: | The proposed method reduces gender bias up to an optimal weight assigned to the loss term, and the model becomes unstable as the perplexity increases. |
Generating Benchmarks for Factuality Evaluation of Language Models (2024.eacl-long)
Copied to clipboard
Dor Muhlgay, Ori Ram, Inbal Magar, Yoav Levine, Nir Ratner, Yonatan Belinkov, Omri Abend, Kevin Leyton-Brown, Amnon Shashua, Yoav Shoham
| Challenge: | Existing methods for factuality evaluation of LLM generation focus on facts sampled from the LM itself and might under-represent domain specific or rare facts. |
| Approach: | They propose a method that transforms a factual corpus into a benchmark evaluating an LM's propensity to generate true facts from the corpus . |
| Outcome: | The proposed framework transforms a factual corpus of interest into a benchmark evaluating an LM's propensity to generate true facts from the corpus vs. similar but incorrect statements. |
Efficient Content-Based Sparse Attention with Routing Transformers (2021.tacl-1)
Copied to clipboard
| Challenge: | Self-attention suffers from quadratic computation and memory requirements with respect to sequence length . despite its effectiveness, self-attention models suffer from quadratic computation and a limited set of locations . |
| Approach: | They propose to learn dynamic sparse attention patterns that avoid allocating computation and memory to attend to content unrelated to the query of interest. |
| Outcome: | The proposed model outperforms similar sparse attention models on language modeling and image generation on Wikitext-103 . |
Knowledge Enhanced Contextual Word Representations (D19-1)
Copied to clipboard
Matthew E. Peters, Mark Neumann, Robert Logan, Roy Schwartz, Vidur Joshi, Sameer Singh, Noah A. Smith
| Challenge: | Existing methods to embed knowledge bases into large pre-training models do not contain any explicit grounding to real world entities and are difficult to recover factual knowledge. |
| Approach: | They propose a method to embed multiple knowledge bases (KBs) into large pretrained models with a Knowledge Attention and Recontextualization mechanism. |
| Outcome: | The proposed model improves perplexity, ability to recall facts and word sense disambiguation. |
Can Large Language Models Personalize Dialogues to Generational Styles? (2025.findings-emnlp)
Copied to clipboard
| Challenge: | a human evaluation reveals that annotators were able to most accurately identify the generation behind P-MultiWoZ dialogues, based only on a single query-reply pair. |
| Approach: | They create a personalized, generation-specific version of MultiWOZ 2.2 by prompting LLMs to generate personalized dialogue responses. |
| Outcome: | The proposed model is a personalized version of MultiWOZ 2.2 for Generation X, Y, and Z . it is validated by automatic and human evaluations to determine whether it reflects generational linguistic traits. |
Locally Typical Sampling (2023.tacl-1)
Copied to clipboard
| Challenge: | a discrepancy in probabilistic language generators has puzzled the language generation community for years . |
| Approach: | They propose a method for local typical sampling to enforce a conditional entropy criterion for probabilistic models. |
| Outcome: | The proposed method can provide new insights into why high-probability texts can be dull or repetitive. |
Reducing Sentiment Bias in Language Models via Counterfactual Evaluation (2020.findings-emnlp)
Copied to clipboard
Po-Sen Huang, Huan Zhang, Ray Jiang, Robert Stanforth, Johannes Welbl, Jack Rae, Vishal Maini, Dani Yogatama, Pushmeet Kohli
| Challenge: | Language modeling has advanced rapidly due to efficient model architectures and the availability of large text corpora. |
| Approach: | They propose to embed and regularize sentiment prediction-derived regularizations on the language model’s latent representations to reduce bias in the sentiment of generated text. |
| Outcome: | The proposed methods reduce bias in the sentiment of generated text by adopting individual and group fairness metrics from the fair machine learning literature. |
Large Scale Multi-Actor Generative Dialog Modeling (2020.acl-main)
Copied to clipboard
| Challenge: | Non-goal oriented dialog agents typically exhibit inconsistent personality across conversations or the average personality of all users. |
| Approach: | They propose a model that conditionally models past conversations to probabilistically model multi-turn conversations in the actor’s persona. |
| Outcome: | The proposed model improves perplexity on 1.7M held out Reddit conversations by 0.47 on scaling from 117M to 8.3B parameters. |
On the Limitations of Language-targeted Pruning: Investigating the Calibration Language Impact in Multilingual LLM Pruning (2026.tacl-1)
Copied to clipboard
| Challenge: | Recent advances in large language model pruning have shown high predictive performance in post-training settings. |
| Approach: | They conduct an empirical study on the performance and internal representation changes associated with pruning multilingual models for monolingual applications. |
| Outcome: | The proposed pruning methods retain perplexity and yield high signal-to-noise ratios, but not consistently improve downstream tasks. |
Modeling Text using the Continuous Space Topic Model with Pre-Trained Word Embeddings (2021.acl-srw)
Copied to clipboard
| Challenge: | Existing topic models that extract latent topics from text are based on latent topic and do not use intermediate variables such as latent subjects. |
| Approach: | They propose a model that extends the continuous space topic model (CSTM) they pre-train word embeddings which capture the semantics of words and plug them into the CSTM . |
| Outcome: | The proposed model performs better than the baseline model in terms of perplexity and convergence speed. |
Lexicon-Level Contrastive Visual-Grounding Improves Language Modeling (2024.findings-acl)
Copied to clipboard
| Challenge: | Neural language models (LMs) are trained on orders of magnitude more language data than human language learners receive, but without supervision from other sensory modalities that play a crucial role in human learning. |
| Approach: | They propose a grounded language learning procedure that leverages visual supervision to improve textual representations. |
| Outcome: | The proposed procedure outperforms standard language-only models in terms of learning efficiency in small and developmentally plausible data regimes and improves perplexity by around 5% on multiple language modeling tasks compared to other models trained on the same amount of text data. |
Semantics of the Unwritten: The Effect of End of Paragraph and Sequence Tokens on Text Generation with GPT2 (2021.acl-srw)
Copied to clipboard
| Challenge: | Experimental results show that pre-trained language model GPT2 can generate better continuations by learning to generate the in the fine-tuning stage. |
| Approach: | They conduct experiments on an English essay dataset using Chinese-GPT2 . they find that the model can generate better continuations by learning to generate the in the fine-tuning stage. |
| Outcome: | The pre-trained language model GPT2 can generate better continuations by learning to generate the in the fine-tuning stage. |
Different Time, Different Language: Revisiting the Bias Against Non-Native Speakers in GPT Detectors (2026.eacl-srw)
Copied to clipboard
| Challenge: | Prior work suggests that automated methods falsely flag essays from non-native speakers as generated due to their low perplexity extracted from an LLM, which is supposedly a key feature of the detectors. |
| Approach: | They propose to use a perplexity-based detector to detect essays from non-native speakers of Czech and a detector to examine the effects of different families. |
| Outcome: | The proposed methods are not biased against non-native speakers, but instead falsely flag essays from non-natural speakers as generated, compared to the English essays. |
Why Does Surprisal From Larger Transformer-Based Language Models Provide a Poorer Fit to Human Reading Times? (2023.tacl-1)
Copied to clipboard
| Challenge: | Existing studies have shown that larger pre-trained language models with more parameters and lower perplexity are less predictive of human reading times. |
| Approach: | They propose to use a transformer-based model with more parameters and lower perplexity to investigate why these models are less predictive of human reading times. |
| Outcome: | The results show that the larger models with more parameters and lower perplexity are less predictive of human reading times and eye-gaze durations collected during naturalistic reading. |
On Retrieval Augmentation and the Limitations of Language Model Training (2024.naacl-short)
Copied to clipboard
| Challenge: | Recent efforts to improve the performance of language models (LMs) have focused on scaling up model and training data size, though with steep accompanying energy and compute resource costs. |
| Approach: | They propose to augment a language model with k-nearest neighbors retrieval on its training data to reduce its perplexity. |
| Outcome: | The proposed model reduces storage costs by over 25x compared to traditional retrieval methods for GPT-2 and Mistral 7B . |
Learning Interpretable Latent Dialogue Actions With Less Supervision (2022.aacl-main)
Copied to clipboard
| Challenge: | supervised neural dialogue modeling requires a significant amount of work to obtain turn-level labels, usually with dialogue state annotation. |
| Approach: | They propose a novel architecture for explainable modeling of task-oriented dialogues with discrete latent variables to represent dialogue actions. |
| Outcome: | The proposed model outperforms previous approaches with less supervision in terms of perplexity and BLEU on three datasets. |
Sharp Nearby, Fuzzy Far Away: How Neural Language Models Use Context (P18-1)
Copied to clipboard
| Challenge: | Recent studies have shed light on the information encoded by long-term memory networks. |
| Approach: | They propose to use a neural caching model to model the role of context in an LSTM LM . they analyze the increase in perplexity when prior context words are shuffled, replaced, or dropped . |
| Outcome: | The proposed model is highly sensitive to the order of words within the most recent sentence, but ignores word order in the long-range context, suggesting the distant past is modeled only as a rough semantic field or topic. |
Dialogue Response Ranking Training with Large-Scale Human Feedback Data (2020.emnlp-main)
Copied to clipboard
| Challenge: | Existing open-domain dialog models can minimize the perplexity of target human responses . however, some human responses are more engaging than others, spawning more followup interactions . |
| Approach: | They train open-domain dialog models to minimize perplexity of target human responses . they use social media feedback data to train models to predict engaging dialog turns . |
| Outcome: | The proposed model outperforms existing models on 133M human feedback pairs . it also outperformed the conventional dialog perplexity baseline model . |
Cross-layer Attention Sharing for Pre-trained Large Language Models (2026.tacl-1)
Copied to clipboard
Yongyu Mu, Yuzhang Wu, Yuchun Fan, Chenglong Wang, Hengyu Li, Jiali Zeng, Qiaozhi He, Murun Yang, Fandong Meng, Jie Zhou, Tong Xiao, Jingbo Zhu
| Challenge: | Existing studies focus on compressing the Key-Value cache or grouping attention heads, while overlooking redundancy between layers. |
| Approach: | They propose a lightweight substitute for self-attention in well-trained LLMs that uses feed-forward networks to align attention heads between adjacent layers and low-rank matrices to approximate differences in layer-wise attention weights. |
| Outcome: | The proposed model reduces redundancy by sharing weights across layers while maintaining high response quality while reducing redundant calculations within 53% 84% of the total layers. |
A Trip Towards Fairness: Bias and De-Biasing in Large Language Models (2024.starsem-1)
Copied to clipboard
| Challenge: | a little or a large bias in CtB-LLMs may cause huge harm . LLaMA and OPT families have an important bias in gender, race, religion, and profession. |
| Approach: | They propose to debiase three families of Very Large-Language Models with LORA to reduce bias by 4.12 points in the normalized stereotype score. |
| Outcome: | The proposed model reduces bias up to 4.12 points in the normalized stereotype score. |
Reducing Gender Bias in Word-Level Language Models with a Gender-Equalizing Loss Function (P19-2)
Copied to clipboard
| Challenge: | Existing methods to reduce gender bias in natural language datasets are inadequate. |
| Approach: | They propose a loss function modification approach which equalizes the probabilities of male and female words in the output. |
| Outcome: | The proposed approach outperforms existing methods in several aspects, especially in reducing gender bias in occupation words. |
Relational Memory-Augmented Language Models (2022.tacl-1)
Copied to clipboard
| Challenge: | Existing language models rely on word correlation and are difficult to interpret . existing models often lack explicit representations for such information . |
| Approach: | They propose a memory-augmented approach to condition autoregressive language models on knowledge graphs. |
| Outcome: | The proposed model improves perplexity and bits per character in an autoregressive language model . it is complementary to token-based memory and enables causal interventions . |
Beyond Tokens: Concept-Level Training Objectives for LLMs (2026.eacl-short)
Copied to clipboard
| Challenge: | Large language models (LLMs) are trained with a surprisingly narrow objective: predicting the next token in a sequence. |
| Approach: | They propose a shift from token-level to concept-level prediction where concepts group multiple surface forms of the same idea. |
| Outcome: | The proposed model improves on human-level models on diverse NLP benchmarks. |
Deep Dirichlet Multinomial Regression (N18-1)
Copied to clipboard
| Challenge: | supervised topic models can incorporate arbitrary document-level features to inform topic priors, but their ability to model corpora is limited by the representation and selection of these features. |
| Approach: | They propose a generative topic model that simultaneously learns document feature representations and topics. |
| Outcome: | The proposed model outperforms DMR and LDA on three datasets and human subjects judge it more representative of associated document features. |
Director: Generator-Classifiers For Supervised Language Modeling (2022.aacl-main)
Copied to clipboard
| Challenge: | Current language models achieve low perplexity but their resulting generations still suffer from toxic responses, repetitiveness, and contradictions. |
| Approach: | They propose a new language model architecture that uses a language modeling and a classification head for each output token. |
| Outcome: | The proposed model outperforms existing model guiding approaches in terms of accuracy and efficiency. |
Transformer versus LSTM Language Models trained on Uncertain ASR Hypotheses in Limited Data Scenarios (2022.lrec-1)
Copied to clipboard
| Challenge: | Existing studies show that domain-specific LMs can only rely on limited in-domain speech data . a qualitative analysis reveals that Transformer LM can predict less frequent words . |
| Approach: | They propose a method to train Transformer LMs on ASR confusion networks . they find they are better at exploiting alternate uncertain ASR hypotheses . |
| Outcome: | The proposed method reduces perplexity by 3-6% on AMI scenarios but performs similar to LSTM LMs on Verbmobil conversational corpus. |
Surface-Based Retrieval Reduces Perplexity of Retrieval-Augmented Language Models (2023.acl-short)
Copied to clipboard
| Challenge: | Augmenting language models with a retrieval mechanism has been shown to improve performance while keeping the number of parameters low. |
| Approach: | They propose to augment language models with a retrieval mechanism by replacing semantic retrieval with BM25 . they find that the model's performance is better explained by surface-level similarities, they say . |
| Outcome: | The proposed method reduces perplexity and lowers the number of parameters while keeping the number low. |
Cascaded Head-colliding Attention (2021.acl-long)
Copied to clipboard
| Challenge: | Existing frameworks for natural language processing ignore interactions among different heads, which wastes the capacity of the model. |
| Approach: | They propose a model which explicitly models interactions between attention heads through a hierarchical variational distribution. |
| Outcome: | The proposed model outperforms the baseline model on Wikitext-103 and WMT14 EN-DE on language modeling and translation tasks. |
Human Language Modeling (2022.findings-acl)
Copied to clipboard
| Challenge: | Existing language modeling models treat text sequences as if they were created independently. |
| Approach: | They propose a hierarchical extension to the language modeling problem whereby a human-level exists to connect sequences of documents and capture the notion that human language is moderated by changing human states. |
| Outcome: | The proposed model outperforms the current state-of-the-art in terms of language modeling and fine-tuning for 4 downstream tasks spanning document- and user-levels. |
Entity Cloze By Date: What LMs Know About Unseen Entities (2022.findings-naacl)
Copied to clipboard
| Challenge: | Existing literature provides benchmarks to measure LMs' knowledge about entities . |
| Approach: | They propose a framework to analyze what language models can infer about new entities that did not exist when they were pretrained. |
| Outcome: | The proposed framework shows that models more informed about the entities achieve lower perplexity on this benchmark. |
Fire Burns, Sword Cuts: Commonsense Inductive Bias for Exploration in Text-based Games (2022.acl-short)
Copied to clipboard
| Challenge: | Existing RL agents are far away from solving text-based games due to their combinatorially large action spaces that hinders efficient exploration. |
| Approach: | They propose an exploration technique that injects external commonsense knowledge, via a pretrained language model, into the agent during training when the agent is the most uncertain about its next action. |
| Outcome: | The proposed method exhibits improvement on the collected game scores during the training in four out of nine games from Jericho. |
Why Exposure Bias Matters: An Imitation Learning Perspective of Error Accumulation in Language Generation (2022.findings-acl)
Copied to clipboard
| Challenge: | Current language generation models suffer from issues such as repetition, incoherence, and hallucinations . |
| Approach: | They propose to analyze exposure bias from an imitation learning perspective and prove it is a problem . they show that exposure bias leads to an accumulation of errors during generation . |
| Outcome: | The proposed model fails to capture errors during generation and poor generation quality. |
Where is the answer? An empirical study of positional bias for parametric knowledge extraction in language model (2025.naacl-long)
Copied to clipboard
| Challenge: | Language model (LM) stores diverse factual knowledge in their parameters, which is learned during self-supervised training on unlabeled documents. |
| Approach: | They investigate the issue of "perplexity curse" in the continued training of language model (LM) they find that all studied LMs suffer from positional bias in the training document . |
| Outcome: | The proposed model is able to extract information from multiple questions with diverse queries. |
Do Long-Range Language Models Actually Use Long-Range Context? (2021.emnlp-main)
Copied to clipboard
| Challenge: | Language models are generally trained on short, truncated input sequences, which limits their ability to use discourse-level information present in long-range context to improve their predictions. |
| Approach: | They analyze two long-range Transformer language models that accept 8K token inputs . they find that providing long-term context only improves their predictions on a small set of tokens - not sentence-level ones . |
| Outcome: | The proposed model improves on PG-19 with only 2K tokens and does not help at all for sentence-level prediction tasks. |
RE-FIN: Retrieval-based Enrichment for Financial data (2025.coling-industry)
Copied to clipboard
| Challenge: | Financial sentiment analysis (FSA) is a powerful tool to support business decision-making and perform financial forecasting. |
| Approach: | They propose a system that retrieves information from a knowledge base to enrich financial sentences, making them more knowledge-dense and explicit. |
| Outcome: | The proposed system generates propositions from the knowledge base and employs Retrieval-Augmented Generation (RAG) to augment the original text with relevant information. |
FitCF: A Framework for Automatic Feature Importance-guided Counterfactual Example Generation (2025.findings-acl)
Copied to clipboard
Qianli Wang, Nils Feldhus, Simon Ostermann, Luis Felipe Villa-Arenas, Sebastian Möller, Vera Schmitt
| Challenge: | Existing frameworks for counterfactual examples are lacking for many tasks. |
| Approach: | They propose a faithful approach for leveraging important words from feature attribution methods to generate counterfactual examples in a zero-shot setting. |
| Outcome: | The proposed framework outperforms state-of-the-art frameworks on many tasks. |
Self-Normalization Properties of Language Modeling (C18-1)
Copied to clipboard
| Challenge: | Existing methods to reduce run-times for language models with large word vocabularies are based on noise contrastive estimation (NCE) |
| Approach: | They propose to use noise-constrained noise-based models to approximate the normalized probability of a class without having to compute the partition function. |
| Outcome: | The proposed model outperforms softmax-based models in a variety of NLP tasks and is based on the noise-constrained noise-constant estimation properties. |
Provably Confidential Language Modelling (2022.naacl-main)
Copied to clipboard
| Challenge: | Existing methods to train language models without memorizing sensitive data are mismatched and can be difficult to screen and filter. |
| Approach: | They propose a method to train language generation models while protecting the confidential segments of training data. |
| Outcome: | The proposed method prevents unintended memorization by randomizing parts of the training process while protecting strong confidentiality. |
What Context Features Can Transformer Language Models Use? (2021.acl-long)
Copied to clipboard
| Challenge: | Recent studies show that transformer-based language models benefit from conditioning on contexts of hundreds to thousands of previous tokens. |
| Approach: | They propose to use lexical and structural information to ablate usable information in transformer language models. |
| Outcome: | The proposed model improves when conditioning on contexts of thousands of previous tokens. |
Can Machine Unlearning Reduce Social Bias in Language Models? (2024.emnlp-industry)
Copied to clipboard
Omkar Dige, Diljot Arneja, Tsz Fung Yau, Qixuan Zhang, Mohammad Bolandraftar, Xiaodan Zhu, Faiza Khattak
| Challenge: | Existing methods for mitigating bias in language models are expensive and time-consuming . comparative studies have not evaluated their respective advantages and disadvantages . |
| Approach: | They propose to use Partitioned Contrastive Gradient Unlearning and Negation via Task Vector to reduce social biases in open-source language models. |
| Outcome: | The proposed methods outperform PCGU and DPO in debiasing models . the proposed methods can be easily tuned to balance the trade-off between bias reduction and generation quality . |
Extending Neural Generative Conversational Model using External Knowledge Sources (D18-1)
Copied to clipboard
| Challenge: | Existing generative dialogue models lack coherence and are content poor . however, current models lack the capacity to handle large unstructured knowledge sources. |
| Approach: | They propose an architecture to incorporate unstructured knowledge sources to enhance the next utterance prediction in chit-chat type of generative dialogue models. |
| Outcome: | The proposed architecture improves the next utterance prediction in chit-chat type of generative dialogue models by incorporating external knowledge from Wikipedia summaries and the NELL knowledge base. |
Adaptation Approaches for Nearest Neighbor Language Models (2023.findings-acl)
Copied to clipboard
| Challenge: | Semi-parametric Nearest Neighbor Language Models (kNN-LMs) have produced impressive gains over purely parametric LMs, however, there has been little investigation into adapting such models for new domains. |
| Approach: | They propose to adapt kNN-LMs to expand neighborhood retrieval over an additional adaptation datastore and adapt the weights of retrieved neighbors using a learned Rescorer module. |
| Outcome: | The proposed approach outperforms purely parametric adaptation and zero-shot models and achieves perplexity improvements of 17.1% and 16% across domains. |
Why Generate When You Can Discriminate? A Novel Technique for Text Classification using Language Models (2024.findings-eacl)
Copied to clipboard
| Challenge: | Existing methods for text classification using autoregressive language models are limited . authors propose a novel technique for text classification using autoreregressives . |
| Approach: | They propose a two-step technique for text classification using autoregressive language models . they use a set of perplexity and log-likelihood based numeric features to elicit a text instance . |
| Outcome: | The proposed technique eliminates parameter updates in LMs and does not limit training examples . it is evaluated across 5 datasets and compares with multiple competent baselines . |
Inconsistent Tokenizations Cause Language Models to be Perplexed by Japanese Grammar (2025.acl-short)
Copied to clipboard
| Challenge: | Standard benchmarks for language models fail to capture nuanced capabilities such as the ability of language models to recognize and obey rare grammar points. |
| Approach: | They find that Weblab's uniformly bad tokenization is a possible root cause for its good performance . |
| Outcome: | The proposed model consistently assigns higher perplexity to ungrammatical psych predicate sentences than grammaticals. |
Are Pre-trained Language Models Useful for Model Ensemble in Chinese Grammatical Error Correction? (2023.acl-short)
Copied to clipboard
| Challenge: | Model ensemble has been in widespread use for Grammatical Error Correction (GEC), boosting model performance. |
| Approach: | They propose to use model ensembles computed by pre-trained language models to improve model performance. |
| Outcome: | The proposed ensembles do not improve but get worse after the PLM-based ensemble. |
Modeling Code-Switch Languages Using Bilingual Parallel Corpus (2020.acl-main)
Copied to clipboard
| Challenge: | Existing models for bilingual language modeling are limited due to lack of training data and syntactic structure. |
| Approach: | They propose a bilingual attention language model that performs language modeling objective with a quasi-translation objective to model the monolingual and cross-lingual sequential dependency. |
| Outcome: | The proposed model reduces the perplexity of 20.5% over the best-reported model. |
On the Efficacy of Sampling Adapters (2023.acl-long)
Copied to clipboard
| Challenge: | Using sampling adapters can improve the quality of the generated text. |
| Approach: | They propose a framework for understanding sampling adapters and propose 'sampling adapters' they argue that the shift enforced by them can be viewed as a trade-off between precision and recall . |
| Outcome: | The proposed framework can be used to improve the quality of language models by modifying their distributions to improve their precision and recall. |
In-Context Demonstration Selection with Cross Entropy Difference (2023.findings-emnlp)
Copied to clipboard
| Challenge: | Large language models (LLMs) can use in-context demonstrations to improve performance on zero-shot tasks. |
| Approach: | They propose a cross-entropy difference method for selecting in-context demonstrations that uses parameter efficient finetuning to train small models on training data. |
| Outcome: | The proposed method outperforms baseline selection methods on a mix-domain dataset and shows that the effectiveness of in-context demonstrations negatively correlates with the perplexity of the test example. |
Transformer Grammars: Augmenting Transformer Language Models with Syntactic Inductive Biases at Scale (2022.tacl-1)
Copied to clipboard
| Challenge: | a novel class of Transformer language models that combine expressive power, scalability, and strong performance of Transformers and recursive syntactic compositions. |
| Approach: | They introduce Transformer Grammars, a class of Transformer language models that combine expressive power and recursive syntactic compositions. |
| Outcome: | The proposed model outperforms strong baselines on sentence-level language modeling perplexity and syntax-sensitive language evaluation metrics. |
Dependency Transformer Grammars: Integrating Dependency Structures into Transformer Language Models (2024.acl-long)
Copied to clipboard
| Challenge: | Syntactic Transformer language models aim to achieve better generalization through simultaneously modeling syntax trees and sentences. |
| Approach: | They propose a class of Transformer language models with explicit dependency-based inductive bias. |
| Outcome: | Experiments show that the proposed models outperform constituency-based models on sentences annotated with dependency trees and achieve better generalization. |
Unstructured Minds, Predictable Machines: A Comparative Study of Narrative Cohesion in Human and LLM Stream-of-Consciousness Writing (2025.acl-srw)
Copied to clipboard
| Challenge: | Stream-of-consciousness narratives are a challenge for large language models (LLMs) authors examined differences between human and LLM-generated narratives to assess narrative coherence and personality expression. |
| Approach: | They generate SoC narratives by prompting LLMs with the first half of SoC-essays while either providing the models with the personality characteristics (Big Five) or omitting them. |
| Outcome: | The proposed models showed low similarity between LLM-generated continuations and original human texts, as measured by cosine similarity, perplexity, and BLEU scores. |
Benchmark Transparency: Measuring the Impact of Data on Evaluation (2024.naacl-long)
Copied to clipboard
| Challenge: | In this paper, we quantify the impact that data distribution has on the performance and evaluation of NLP models. |
| Approach: | They propose to use disproportional stratified sampling to measure the data distribution across 6 different dimensions to quantify model performance. |
| Outcome: | The proposed framework measures the data distribution across 6 different dimensions and shows that it is statistically significant and predicts model performance. |
How much do contextualized representations encode long-range context? (2025.findings-naacl)
Copied to clipboard
| Challenge: | Existing studies of contextualized representations focus on short sequences of tens to hundreds of tokens, whereas modern language models handle hundreds of thousands of token in a single context window. |
| Approach: | They use a perturbation setup and a metric to capture contextualization of long-range patterns from the perspective of representation geometry. |
| Outcome: | The proposed model can encode long-range contexts, but it's not fully recurrent, the authors say . their results suggest improvements in existing language models . |
Better Language Model with Hypernym Class Prediction (2022.acl-long)
Copied to clipboard
| Challenge: | Class-based language models (LMs) have been devised to address context sparsity in n-gram LMs for decades. |
| Approach: | They propose to use class-based prediction to improve generalization for rare words by annealing from predicting the class to token prediction during training. |
| Outcome: | The proposed model improves perplexity without sacrificing performance on rare words. |
Coherence-Aware Neural Topic Modeling (D18-1)
Copied to clipboard
| Challenge: | Topic models are evaluated for their ability to describe documents well (i.e. low perplexity) topic coherence is not optimized for and is only evaluated after training. |
| Approach: | They propose to incorporate a topic coherence objective into the training process by incorporating a coherency objective into a model. |
| Outcome: | The proposed model exhibits similar level of perplexity as baseline models but significantly higher topic coherence. |
Spurious Correlations in Reference-Free Evaluation of Text Generation (2022.acl-long)
Copied to clipboard
| Challenge: | Recent work suggests that reference-free evaluation metrics may rely on spurious correlations with human judgments. |
| Approach: | They propose to use model-based, reference-free evaluation metrics to evaluate natural language generation systems. |
| Outcome: | The proposed metrics achieve high correlations with human judgments, but they may not be robust enough to evaluate their efficacy and robustness. |
Exploring Forgetting in Large Language Model Pre-Training (2025.acl-long)
Copied to clipboard
| Challenge: | Existing research on task-level forgetting in LLMs has focused on pretraining . but, there is limited attention to finer-grained forgetting during training . |
| Approach: | They investigated the existence and measurement of forgetting in pre-training . they examined low-cost, straightforward methods to mitigate forgetting during the pre- training phase . |
| Outcome: | The proposed methods could be used to mitigate forgetting during the pre-training phase and offer insights into the dynamics of forgetting. |
Adversarial Decoding: Generating Readable Documents for Adversarial Objectives (2026.findings-eacl)
Copied to clipboard
| Challenge: | Existing methods for generating adversarial documents produce gibberish that is easy to detect and filter out. |
| Approach: | They propose a generic text generation technique that produces readable adversarial documents . they demonstrate that adversarials can be used for different objectives . |
| Outcome: | The proposed technique outperforms existing methods while producing readable documents for adversarial objectives. |
Form and Meaning in Intrinsic Multilingual Evaluations (2026.eacl-long)
Copied to clipboard
| Challenge: | Existing metrics for conditional language models are not universally comparable . a number of assumptions are made about their quality in multilingual settings . |
| Approach: | They make assumptions about the quality of conditional language models and their semantic meanings that are not consistent with current metrics. |
| Outcome: | The results show that current evaluation metrics are not universally comparable . they use six metrics on two multi-parallel corpora with mono- and multilingual models . |
Knowledge-Augmented Language Model and Its Application to Unsupervised Named-Entity Recognition (N19-1)
Copied to clipboard
| Challenge: | Current language models are unable to efficiently model entity names observed in text providing insufficient context. |
| Approach: | They propose to augment a traditional model with an external knowledge base to model entity names observed in text. |
| Outcome: | The proposed model improves on a Named Entity Recognition (NER) task by requiring no additional information such as named entity tags. |
Enhancing One-Shot Pruned Pre-trained Language Models through Sparse-Dense-Sparse Mechanism (2025.coling-main)
Copied to clipboard
Guanchen Li, Xiandong Zhao, Lian Liu, Zeping Li, Yixing Xu, Dong Li, Lu Tian, Jie He, Ashish Sirasao, Emad Barsoum
| Challenge: | Pre-trained language models (PLMs) are robust in contextual understanding but their considerable size incurs significant computational and storage costs. |
| Approach: | They propose a Sparse-Dense-Sparse pruning framework to prune PLMs . they prune less critical connections using conventional pruning methods . |
| Outcome: | The proposed pruning framework outperforms SparseGPT and Wanda under identical sparsity. |
Optimizing Packing and Shuffling Strategies for Enhanced Performance in Generative Language Models (2026.acl-srw)
Copied to clipboard
| Challenge: | Documents are typically concatenated to chunks of maximum sequence length (MSL) and shuffled in chunks (atom-size chunks). |
| Approach: | They propose to pack and shuff documents in chunks of tokens to prevent overfitting . they also propose to use padding to only include one document per chunk . |
| Outcome: | The proposed method reduces the risk of overfitting and improves generalizability. |
ValCAT: Variable-Length Contextualized Adversarial Transformations Using Encoder-Decoder Language Model (2022.naacl-main)
Copied to clipboard
| Challenge: | Existing word-level approaches to attack text are limited to a single word . existing methods ignore interactions between consecutive words, resulting in one-to-one attacks . |
| Approach: | They propose a black-box attack framework that misleads the language model by applying variable-length contextualized transformations to the original text. |
| Outcome: | The proposed framework outperforms existing methods on classification and inference tasks. |
On Localizing and Deleting Toxic Memories in Large Language Models (2025.findings-naacl)
Copied to clipboard
Anubrata Das, Manoj Kumar, Ninareh Mehrabi, Anil Ramakrishna, Anna Rumshisky, Kai-Wei Chang, Aram Galstyan, Morteza Ziyadi, Rahul Gupta
| Challenge: | Existing methods to reduce toxic generation in large language models are not fully understood. |
| Approach: | They propose to understand the mechanisms that drive toxic generation in large language models by using memory localization to reduce toxic generation. |
| Outcome: | The proposed method reduces toxic generation from 62.86% to 28.61%, but it also improves generation quality. |
Controlling the Extraction of Memorized Data from Large Language Models via Prompt-Tuning (2023.acl-short)
Copied to clipboard
Mustafa Ozdayi, Charith Peris, Jack FitzGerald, Christophe Dupuy, Jimit Majmudar, Haidar Khan, Rahil Parikh, Rahul Gupta
| Challenge: | Large Language Models memorize significant portions of training data, which poses privacy risk. |
| Approach: | They propose a prompt-tuning approach to control the extraction rates of memorized content in large language models. |
| Outcome: | The proposed techniques yield 9.3% increase in extraction rate compared to baseline model . the proposed defense achieves 97.7% reduction with a perplexity increase of 16.9% . |
Are Text Classifiers Xenophobic? A Country-Oriented Bias Detection Method with Least Confounding Variables (2024.lrec-main)
Copied to clipboard
| Challenge: | Existing methods for detecting biases are biased because of confounding variables . authors propose a method to detect the biased classifier on any type of unlabeled data . |
| Approach: | They propose a method to detect biases of a specific fine-tuned classifier on unlabeled data. |
| Outcome: | The proposed method detects biases on unlabeled data on named entity perturbations . it uses name-entity recognition on target-domain data and morphosynctactically different languages spoken in relation to countries of the target groups . |
Improved Language Modeling by Decoding the Past (P19-1)
Copied to clipboard
| Challenge: | Existing methods to improve language modeling performance are based on regularized LSTMs with a large number of parameters and training time. |
| Approach: | They propose a method that decodes the last token in context using the predicted distribution of the next token. |
| Outcome: | The proposed method improves perplexity on the Penn Treebank dataset by 1.8 points and 2.3 points on the WikiText-2 datasets. |
Language Modeling for Code-Mixing: The Role of Linguistic Theory based Synthetic Data (P18-1)
Copied to clipboard
| Challenge: | Code-mixed (CM) language training is a difficult problem because of lack of data and the increased confusability due to the presence of more than one language. |
| Approach: | They propose a computational technique for creating grammatically valid artificial CM data based on the Equivalence Constraint Theory. |
| Outcome: | The proposed method reduces the perplexity of the model and does not reduce the perceptibility of the models. |
ASTPrompter: Preference-Aligned Automated Language Model Red-Teaming to Generate Low-Perplexity Unsafe Prompts (2025.findings-emnlp)
Copied to clipboard
| Challenge: | Existing red-teaming approaches prioritize high attack success rate, resulting in high-perplexity prompts. |
| Approach: | a new method uses contrastive preference learning to train an attacker to maintain low perplexity while achieving a high attack success rate. |
| Outcome: | ASTPrompter achieves 5.1 times higher attack success rate on Llama-8.1B . low-perplexity attacks are more difficult to filter and more likely to arise during benign usage . |
Improving Neural Language Models by Segmenting, Attending, and Predicting the Future (P19-1)
Copied to clipboard
| Challenge: | Common language models typically predict the next word given a past context. |
| Approach: | They propose a method that aligns the given context and the following phrase . they define syntactic heights and phrase segmentation rules to enable it to learn . |
| Outcome: | The proposed model outperforms strong baseline models on Wikitext-103 dataset. |
Position-Aware Depth Decay Decoding (D3): Boosting Large Language Model Inference Efficiency (2025.findings-acl)
Copied to clipboard
| Challenge: | Recent dynamic computation methods show that not all components are required for inference, enabling a training-free pipeline. |
| Approach: | They propose a token-position aware layer skipping framework to save 1.5x times operations efficiently while maintaining performance. |
| Outcome: | The proposed algorithm achieves 1.5x speedup on large language models with no retraining and with comparable performance on the GSM8K and BBH benchmarks. |
CoCoa: An Encoder-Decoder Model for Controllable Code-switched Generation (2022.emnlp-main)
Copied to clipboard
| Challenge: | Generating code-switched text with fine-grained control on the degree of code-witching and the lexical choices used to convey formality has been well-explored. |
| Approach: | They propose to generate code-switched text with fine-grained control on the degree of code-changing and lexical choices used to convey formality. |
| Outcome: | The proposed model can be invoked at test-time to synthesize code-switched text faithful to syntactic and lexical attributes relevant to code-witching. |
A Systematic Assessment of Syntactic Generalization in Neural Language Models (2020.acl-main)
Copied to clipboard
| Challenge: | Existing work on syntactic knowledge models has not provided a clear picture of the properties required to produce proper syntaktic generalizations. |
| Approach: | They propose to evaluate syntactic knowledge of language models by varying model architectures . they find substantial differences in syntaktic generalization performance by model architecture . |
| Outcome: | The proposed model architectures outperform other architectures on a set of 34 English-language syntactic test suites. |
You should evaluate your language model on marginal likelihood over tokenisations (2021.emnlp-main)
Copied to clipboard
| Challenge: | Neural language models typically tokenise input text into sub-word units to achieve an open vocabulary. |
| Approach: | They propose that language models should be evaluated on their marginal likelihood over tokenisations instead. |
| Outcome: | The proposed approach is unsatisfactory and may bottleneck model out-of-domain performance. |
Retrofitting Structure-aware Transformer Language Model for End Tasks (2020.emnlp-main)
Copied to clipboard
| Challenge: | Experimental results show that structure-aware Transformer language model achieves improved perplexity, meanwhile inducing accurate syntactic phrases. |
| Approach: | They propose to exploit syntactic distance to encode phrasal constituency and dependency connection into Transformer language model and leverage it for structure integration. |
| Outcome: | The proposed model achieves significant improvements for both semantic- and syntactic-dependent tasks. |
Why Do Some Inputs Break Low-Bit LLM Quantization? (2025.emnlp-main)
Copied to clipboard
| Challenge: | Low-bit weight-only quantization reduces memory usage but disproportionately affects certain examples. |
| Approach: | They analyze quantization errors of 50 pairs of methods on large language models and test their hypothesis . |
| Outcome: | The proposed method reduces the memory footprint of large language models while maintaining reasonable performance across benchmarks. |
Unifying Human and Statistical Evaluation for Natural Language Generation (N19-1)
Copied to clipboard
| Challenge: | Human evaluation captures quality but fails to capture diversity . statistical evaluation fails to catch models that plagiarize from training set . |
| Approach: | They propose a framework which evaluates both diversity and quality based on the optimal error rate of predicting whether a sentence is human-generated. |
| Outcome: | The proposed framework evaluates diversity and quality on summarization and chit-chat dialogue. |
Large Language Models Have Intrinsic Meta-Cognition, but Need a Good Lens (2025.emnlp-main)
Copied to clipboard
| Challenge: | Existing studies have focused on the cognitive error detection capabilities of Large Language Models (LLMs), but few studies have examined the meta-cognitive abilities of LLMs. |
| Approach: | They propose an automated meta-cognition evaluation framework for evaluation of LLMs and a Markovian Intrinsic Reward Adjustment strategy to boost current lenses. |
| Outcome: | The proposed framework can be used to evaluate the meta-cognition abilities of LLMs and improve them. |
Attention-based Contextual Language Model Adaptation for Speech Recognition (2021.findings-acl)
Copied to clipboard
| Challenge: | Existing language models do not incorporate utterance level contextual information . however, for some domains like voice assistants, additional context provides a rich input signal . |
| Approach: | They propose a method for training neural speech recognition models on text and contextual data. |
| Outcome: | The proposed model reduces perplexity by 7.0% relative over a standard LM . it also improves perxicity by 2.8% relative to a state-of-the-art model for contextual LM. |
A Tale of Two Perplexities: Sensitivity of Neural Language Models to Lexical Retrieval Deficits in Dementia of the Alzheimer’s Type (2020.acl-main)
Copied to clipboard
| Challenge: | Recent studies show that cognitive manifestations of future dementia may appear as early as 18 years prior to clinical diagnosis . lack of clear diagnosis and prognosis, possibly for an Alzheimer's type, is a major limitation of current methods for identifying dementia-specific cognitive markers. |
| Approach: | They propose to interrogate neural LMs trained on participants with and without dementia by manipulating lexical frequency. |
| Outcome: | The proposed model improves upon the current state-of-the-art for models trained on transcripts of speech produced by healthy participants and those with dementia. |
LoRAPrune: Structured Pruning Meets Low-Rank Parameter-Efficient Fine-Tuning (2024.findings-acl)
Copied to clipboard
| Challenge: | Low-rank adaption (LoRA) is a low-level pruning method that can be expensive and slow to deploy. |
| Approach: | They propose a low-rank adaption pruning framework that provides an accurate structured pruned model in a memory-efficient manner. |
| Outcome: | The proposed pruning framework reduces perplexity and memory usage by 52.6% on LLaMA and T5 models while reducing memory usage. |
Unsupervised Domain Adaptation for Question Generation with DomainData Selection and Self-training (2022.findings-naacl)
Copied to clipboard
| Challenge: | Existing question generation models require large-scale and high-quality training data. |
| Approach: | They propose an unsupervised domain adaptation approach to combat the lack of training data and domain shift issue with domain data selection and self-training. |
| Outcome: | The proposed approach outperforms baselines on three large datasets with different domain similarities, using a transformer-based pre-trained QG model. |
PromptEHR: Conditional Electronic Healthcare Records Generation with Prompt Learning (2022.emnlp-main)
Copied to clipboard
| Challenge: | Existing methods for generating longitudinal multimodal EHRs are limited due to privacy concerns. |
| Approach: | They propose to generate longitudinal multimodal EHRs by unconditional generation or longitudinal inference . existing methods generate single-modal E HRs by conditional generation or by longitudinal inferment . |
| Outcome: | The proposed method is more flexible and controllable than existing methods and is more cost-effective than existing ones. |
ScalingFilter: Assessing Data Quality through Inverse Utilization of Scaling Laws (2024.emnlp-main)
Copied to clipboard
| Challenge: | Existing quality filtering methods rely on a high-quality dataset as reference . Existing methods introduce potential biases and compromise diversity . |
| Approach: | They propose a method that evaluates text quality based on the perplexity difference between two language models trained on the same data. |
| Outcome: | The proposed approach improves performance of pre-trained models without increasing training costs. |
Neural Models for Documents with Metadata (P18-1)
Copied to clipboard
| Challenge: | specialized models are often used to model text corpora without metadata . specialized algorithms are not widely used in the digital humanities and political science fields . |
| Approach: | They propose a general neural framework based on topic models to enable customization of metadata. |
| Outcome: | The proposed framework achieves strong performance with a manageable tradeoff between perplexity, coherence, and sparsity. |
Extending Input Contexts of Language Models through Training on Segmented Sequences (2024.findings-naacl)
Copied to clipboard
| Challenge: | Effectively training languages models on long sequences poses many technical challenges. |
| Approach: | They propose a method for extending positional embeddings by sub-sampling segments from long inputs while maintaining their original position. |
| Outcome: | The proposed method extends the input con-text size of pretrained models without any changes in the model's memory and memory costs. |
SynGhost: Invisible and Universal Task-agnostic Backdoor Attack via Syntactic Transfer (2025.findings-naacl)
Copied to clipboard
| Challenge: | Existing attacks are classified into end-to-end and pre-training types based on the attack phase . Existing backdoor attacks are based upon perplexity, fine-pruning, and maxEntropy. |
| Approach: | They propose an entropy-based poisoning filter that mitigates backdoor attacks . they propose an invisible and universal task-agnostic backdoor attack via syntactic transfer . |
| Outcome: | The proposed attack can transfer backdoors to various downstream tasks while preserving pre-trained language models' pre-training capabilities. |
Numeracy for Language Models: Evaluating and Improving their Ability to Predict Numbers (P18-1)
Copied to clipboard
| Challenge: | Numeracy is the ability to understand and work with numbers. |
| Approach: | They propose a neural architecture that uses a continuous probability density function to model numerals from an open vocabulary using hierarchical models. |
| Outcome: | The proposed model reduces errors by 18% and 54% on clinical and scientific datasets compared to the second best model for each dataset . |
TokAlign: Efficient Vocabulary Adaptation via Token Alignment (2025.acl-long)
Copied to clipboard
| Challenge: | Tokenization is a foundational step for Large Language Models (LLMs) but low compression rate of vanilla tokenizers decelerates training and inference process. |
| Approach: | They propose a method to replace the vocabulary of Large Language Models (LLMs) by learning a one-to-one mapping matrix for token IDs. |
| Outcome: | The proposed method significantly improves multilingual text compression rates and vocabulary initialization for Large Language Models. |
Skim-Attention: Learning to Focus via Document Layout (2021.findings-emnlp)
Copied to clipboard
| Challenge: | Existing approaches to document understanding have high computational and memory costs. |
| Approach: | They propose a new attention mechanism that takes advantage of the structure of a document and its layout. |
| Outcome: | The proposed attention mechanism obtains lower perplexity than previous studies while being more computationally efficient. |
Demystifying Mixed Outcomes of Self-Training: Pre-training Analyses on Non-Toy LLMs (2026.findings-eacl)
Copied to clipboard
| Challenge: | Recent studies on self-training report seemingly contradictory outcomes. |
| Approach: | They use OLMo-2 models as non-toy LLMs and perform multiple rounds of continual pre-training using self-generated text with different prompting strategies and data filtering. |
| Outcome: | The proposed model collapse is inherent to the training procedure itself, while self-improvement is likely owes its success to human-designed, strategic synthetic pipelines that inject external intelligence. |
Probabilistic Robustness for Data Filtering (2023.eacl-main)
Copied to clipboard
| Challenge: | Modern machine learning works with massive amounts of data on a range of tasks like language modeling, object detection, and data mining. |
| Approach: | They propose a probabilistic robustness rewarded data optimization approach to enhance the model's generalization power by selecting training data that optimizes probabilistic metrics. |
| Outcome: | The proposed approach achieves +17.2% increase of accuracy and -28.05 decrease of perplexity on unknown-domain test sets. |
Learning to Generate Word Representations using Subword Information (C18-1)
Copied to clipboard
| Challenge: | Existing word-based approaches to learning word representations are blind to subword information in words. |
| Approach: | They propose a character-based word representation approach to learn word representations from characters. |
| Outcome: | The proposed model outperforms baseline models that regard words as atomic units . the proposed model achieves 18.5% improvement on average in perplexity for morphologically rich languages . |
You can’t pick your neighbors, or can you? When and How to Rely on Retrieval in the kNN-LM (2022.findings-emnlp)
Copied to clipboard
| Challenge: | Recent studies have shown that retrieval-enhanced language models can improve perplexity by combining text from large external datastores with a k-nearest neighbors model. |
| Approach: | They propose a retrieval-enhanced language model that interpolates existing LMs with a k-nearest neighbors model and requires no additional training. |
| Outcome: | The proposed model improves on two English language modeling datasets and shows that it is most effective when items have high semantic similarity with the query. |
SoftDedup: an Efficient Data Reweighting Method for Speeding Up Language Model Pre-training (2024.acl-long)
Copied to clipboard
| Challenge: | Current methods focus on detecting and removing duplicates, which risks the loss of valuable information and neglects the varying degrees of duplication. |
| Approach: | They propose a method that maintains dataset integrity while selectively reducing the sampling weight of data with high commonness. |
| Outcome: | The proposed method significantly improves training efficiency on deduplicated datasets and improves downstream accuracy by 1.77%. |
Building Language Models for Text with Named Entities (P18-1)
Copied to clipboard
| Challenge: | Existing language models fail to predict the entity names due to their wide variations. |
| Approach: | They propose a language model which can learn the entity names by leveraging their entity type information. |
| Outcome: | The proposed model achieves 52.2% better perplexity in recipe generation and 22.06% on code generation than state-of-the-art language models. |
LM-Infinite: Zero-Shot Extreme Length Generalization for Large Language Models (2024.naacl-long)
Copied to clipboard
| Challenge: | Currently, large language models (LLMs) train on short text segments due to the computational overhead quadratic in the input lengths of their Transformer architectures. |
| Approach: | They propose a method that allows LLMs pre-trained with 2K or 4K-long segments to generalize to up to 200M length inputs while retaining perplexity. |
| Outcome: | The proposed method achieves 2.7 decoding speed up and 7.5 memory saving over the original model. |
Calibrating Beyond English: Language Diversity for Better Quantized Multilingual LLMs (2026.eacl-long)
Copied to clipboard
| Challenge: | Existing quantization methods typically use small, English-only calibration sets . however, their impact on multilingual models remains underexplored . |
| Approach: | They evaluate eight calibration settings across two quantizers on data from 10 different languages. |
| Outcome: | The results show that tailoring calibration sets to the evaluation language yields the largest improvements for individual languages, underscoring the importance of linguistic alignment. |
Wanda++: Pruning Large Language Models via Regional Gradients (2025.findings-acl)
Copied to clipboard
Yifan Yang, Kai Zhen, Bhavana Ganesh, Aram Galstyan, Goeric Huybrechts, Markus Müller, Jonas M. Kübler, Rupak Vignesh Swaminathan, Athanasios Mouchtaris, Sravan Babu Bodapati, Nathan Susanj, Zheng Zhang, Jack FitzGerald, Abhishek Kumar
| Challenge: | Existing pruning methods suffer from accuracy degradation without full-model sparsity-aware fine-tuning. |
| Approach: | They propose a pruning framework that uses decoder-block-level regional gradients to improve pruning accuracy. |
| Outcome: | The proposed pruning framework outperforms the state-of-the-art pruning frameworks by utilizing decoder-block-level regional gradients. |
ERNIE-Doc: A Retrospective Long-Document Modeling Transformer (2021.acl-long)
Copied to clipboard
| Challenge: | Existing models for document-level language pretraining are not suitable for long documents due to their quadratically increasing memory and time consumption. |
| Approach: | They propose a document-level language pretraining model based on Recurrence Transformers. |
| Outcome: | The proposed model outperforms existing models on language understanding tasks. |
Language Model-Driven Data Pruning Enables Efficient Active Learning (2026.findings-eacl)
Copied to clipboard
| Challenge: | Existing data pruning methods for active learning are expensive and time-consuming. |
| Approach: | They propose a plug-and-play data pruning strategy that leverages language models to prune the unlabeled pool. |
| Outcome: | The proposed pruning strategy outperforms existing pruning methods on translation, sentiment analysis, topic classification, and summarization tasks on diverse datasets. |
Adapting Language Models to Compress Contexts (2023.emnlp-main)
Copied to clipboard
| Challenge: | Transformer-based language models have a finite context window and expensive computational cost of processing long text documents. |
| Approach: | They propose to adapt pre-trained LMs into AutoCompressors to compress text into summary vectors . authors propose to use summary vector to speed up inference over long contexts based on a finite context window . |
| Outcome: | The proposed model can compress long contexts into summary vectors, which are accessible as soft prompts. |
Modeling Human Sentence Processing with Left-Corner Recurrent Neural Network Grammars (2021.emnlp-main)
Copied to clipboard
| Challenge: | Existing literature is agnostic about a parsing strategy of hierarchical models . a recent study showed that hierarchically model hierarchic structures capture grammatical dependencies much better than RNNs in targeted syntactic evaluations. |
| Approach: | They evaluated three LMs with head-final left-branching structures and Recurrent Neural Network Grammars with top-down and left-corner parsing strategies as hierarchical models. |
| Outcome: | The proposed model outperforms top-down and left-corner models against human reading times in Japanese. |
From Machine Translation to Code-Switching: Generating High-Quality Code-Switched Text (2021.acl-long)
Copied to clipboard
| Challenge: | a computational model for code-switching text is lacking in the corpus of real text. |
| Approach: | They propose a neural machine translation model to generate Hindi-English code-switched sentences using monolingual Hindi sentences. |
| Outcome: | The proposed model reduces perplexity on a language modeling task and improves on linguistic inference tasks. |
DiffusionBERT: Improving Generative Masked Language Models with Diffusion Models (2023.acl-long)
Copied to clipboard
| Challenge: | Existing generative masked language models have a shared training objective, i.e., denoising. |
| Approach: | They propose a noise schedule for the forward diffusion process that controls the degree of noise added at each step based on the information of each token. |
| Outcome: | The proposed model improves on existing models in terms of perplexity and BLEU score. |
Truncation Sampling as Language Model Desmoothing (2022.findings-emnlp)
Copied to clipboard
| Challenge: | Long samples of text from neural language models can be of poor quality. |
| Approach: | They propose to think of a neural language model as a mixture of k and a true distribution that avoids infinite perplexity. |
| Outcome: | The proposed methods generate more plausible long documents according to humans and break out of repetition. |
Improving Word Embedding Factorization for Compression Using Distilled Nonlinear Neural Decomposition (2020.findings-emnlp)
Copied to clipboard
| Challenge: | Word-embeddings are vital components of natural language processing (NLP) but they consume a lot of memory which poses a challenge for edge deployment. |
| Approach: | They propose an embedding compression method based on matrix decomposition and knowledge distillation that initializes weights of pre-trained word-embeddings and fine-tunes end-to-end. |
| Outcome: | The proposed method has higher BLEU score on translation and lower perplexity on language modeling compared to complex, difficult to tune methods. |
One-Shot Learning as Instruction Data Prospector for Large Language Models (2024.acl-long)
Copied to clipboard
Yunshui Li, Binyuan Hui, Xiaobo Xia, Jiaxi Yang, Min Yang, Lei Zhang, Shuzheng Si, Ling-Hao Chen, Junhao Liu, Tongliang Liu, Fei Huang, Yongbin Li
| Challenge: | Contemporary practices in instruction tuning often hinge on enlarging data scaling without a clear strategy for ensuring data quality. |
| Approach: | They propose a method that leverages one-shot learning to discern and select high-quality instruction data from extensive datasets. |
| Outcome: | Nuggets outperforms existing methods on MT-Bench and Alpaca-Eval benchmarks. |
BlockPruner: Fine-grained Pruning for Large Language Models (2025.findings-acl)
Copied to clipboard
| Challenge: | Large language models (LLMs) have significant computational and memory costs associated with training and inference. |
| Approach: | They propose a training-free structured pruning approach that targets redundancies in MHA and MLP blocks. |
| Outcome: | The proposed pruning approach achieves more granular and effective pruning compared to state-of-the-art pruning methods. |
CUE Vectors: Modular Training of Language Models Conditioned on Diverse Contextual Signals (2022.findings-acl)
Copied to clipboard
| Challenge: | Using contextual universal embeddings, we train neural language models on one type of contextual data and adapts to novel context types. |
| Approach: | They propose a framework to modularize the training of neural language models that use diverse forms of context by eliminating the need to jointly train context and within-sentence encoders. |
| Outcome: | The proposed framework trains LMs on one type of contextual data and adapts to novel context types. |
How Conservative are Language Models? Adapting to the Introduction of Gender-Neutral Pronouns (2022.naacl-main)
Copied to clipboard
| Challenge: | a recent study shows that gender-neutral pronouns are not associated with processing difficulties . linguistic scholars have observed how technology has altered the course of language evolution . |
| Approach: | They show that gender-neutral pronouns in Danish, English and Swedish are not associated with processing difficulties. |
| Outcome: | a new study shows that gender-neutral pronouns are not associated with human processing difficulties . the findings suggest that such conservativity in language models may limit widespread adoption . |
Improving Transformer Models by Reordering their Sublayers (2020.acl-main)
Copied to clipboard
| Challenge: | a sandwich transformer pattern is a new approach to multilayer transformers that can be used for different tasks. |
| Approach: | They propose a transformer ordering pattern that reorders sublayers in a sandwich transformer pattern . they generate random transformer models and train them with the language modeling objective . |
| Outcome: | The proposed pattern improves perplexity on multiple word-level and character-level language modeling benchmarks at no cost in parameters, memory, or training time. |
Transformer-XL: Attentive Language Models beyond a Fixed-Length Context (P19-1)
Copied to clipboard
| Challenge: | Term memory networks (RNNs) are difficult to optimize due to gradient vanishing and explosion. |
| Approach: | They propose a neural architecture Transformer-XL that enables learning dependency beyond a fixed length without disrupting temporal coherence. |
| Outcome: | The proposed method improves state-of-the-art performance on short and long sequences and generates coherent, novel text articles with thousands of tokens. |
Transparency at the Source: Evaluating and Interpreting Language Models With Access to the True Distribution (2023.findings-emnlp)
Copied to clipboard
| Challenge: | a new approach to train, evaluate and interpret neural language models uses artificial, language-like data. |
| Approach: | They propose a setup for training, evaluating and interpreting neural language models that uses artificial, language-like data. |
| Outcome: | The proposed model is based on a massive probabilistic grammar and a large natural language corpus, and provides complete control over the generative process. |
Do LLMs model human linguistic variation? A case study in Hindi-English Verb code-mixing (2026.findings-eacl)
Copied to clipboard
| Challenge: | Existing large language models (LLMs) do not reliably classify verb language preferences to match native speaker judgments. |
| Approach: | They investigate whether large language models (LLMs) model linguistic variation by comparing Hindi-English verb code-mixing with English verb karna. |
| Outcome: | The proposed models do not reliably classify verb language preferences to match native speaker judgments, but with specific supervision, some models do predict human preference to an extent. |
Instruction-tuned Language Models are Better Knowledge Learners (2024.acl-long)
Copied to clipboard
Zhengbao Jiang, Zhiqing Sun, Weijia Shi, Pedro Rodriguez, Chunting Zhou, Graham Neubig, Xi Lin, Wen-tau Yih, Srini Iyer
| Challenge: | Large language models store factual knowledge in parameters, but it can become outdated as the work evolves . pre-instruction-tuning improves ability of LLMs to absorb knowledge from new documents . |
| Approach: | They propose a method that instruction-tunes on questions prior to training on documents . they propose to use QA pairs to update factual knowledge of large language models . |
| Outcome: | The proposed method outperforms instruction-tuning on documents by 17.8%. |
Rethinking Creativity Evaluation: A Critical Analysis of Existing Creativity Evaluations (2026.eacl-long)
Copied to clipboard
| Challenge: | Creativity measures that distinguish creativity in one domain fail in others, and different metrics disagree on the same data points. |
| Approach: | They examine, analyze, and compare four representative creativity measures across the diverse creative domains, including creative writing, unconventional problem-solving, and research ideation. |
| Outcome: | The measures of creativity across creative domains are compared using a set of human-aligned examples and lack consistency across domains and metrics. |
Evaluating Approaches to Personalizing Language Models (2020.lrec-1)
Copied to clipboard
| Challenge: | a large amount of text is not available for training a user-specific language model, which suggests a need to personalize language models with only a small amount of data. |
| Approach: | They propose three approaches to personalize a language model that was trained on a large background corpus using a relatively small amount of text from an individual user. |
| Outcome: | The proposed techniques outperform language model adaptation based on demographic factors. |
Class-based LSTM Russian Language Model with Linguistic Information (2020.lrec-1)
Copied to clipboard
| Challenge: | LSTM models can be used in speech recognition systems at N-best or lattice rescoring stage. |
| Approach: | They propose to use word frequency and linguistic information to generate class-based LSTM Russian language models with various numbers of classes. |
| Outcome: | The proposed models outperform word-based models and word2vec models in terms of perplexity, training time, and word error rate. |
SLING: Sino Linguistic Evaluation of Large Language Models (2022.emnlp-main)
Copied to clipboard
| Challenge: | Using pre-trained language models, we find that the accuracy of LMs is far below human performance. |
| Approach: | They propose a benchmark of Sino LINGuistics which consists of 38K sentence pairs in Mandarin Chinese grouped into 9 high-level linguistic phenomena. |
| Outcome: | The proposed model performs better on local phenomena than hierarchical models and has a strong gender and number bias. |
An Analysis of the Utility of Explicit Negative Examples to Improve the Syntactic Abilities of Neural Language Models (2020.acl-main)
Copied to clipboard
| Challenge: | Neural language models are often trained on positive examples, but recent studies suggest they are not robust enough to handle complex syntactic constructions. |
| Approach: | They propose to use negative examples to boost models' robustness on English sentences with a negligible loss of perplexity. |
| Outcome: | The proposed model is robust to negative examples in English with negligible loss of perplexity . |
Evaluating Sparse Autoencoders for Monosemantic Representation (2026.findings-eacl)
Copied to clipboard
| Challenge: | Sparse autoencoders (SAEs) have been proposed to mitigate polysemanticity, where neurons activate for multiple unrelated concepts. |
| Approach: | They propose a sparse autoencoder to transform dense activations into sparser, more interpretable features by transforming them into sparses. |
| Outcome: | The proposed model reduces polysemanticity and achieves higher concept separability. |
Causal Distillation for Language Models (2022.naacl-main)
Copied to clipboard
Zhengxuan Wu, Atticus Geiger, Joshua Rozner, Elisa Kreiss, Hanson Lu, Thomas Icard, Christopher Potts, Noah Goodman
| Challenge: | Distillation efforts have led to language models that are more compact and efficient without serious drops in performance. |
| Approach: | They propose to augment distillation with a third objective that encourages the student model to imitate the causal dynamics of the teacher through a distillation interchange intervention training objective (DIITO). |
| Outcome: | The proposed method lowers perplexity on the WikiText-103M corpus and improves on the GLUE benchmark, SQuAD, and CoNLL-2003. |
Evaluating Language Models as Synthetic Data Generators (2025.acl-long)
Copied to clipboard
Seungone Kim, Juyoung Suk, Xiang Yue, Vijay Viswanathan, Seongyun Lee, Yizhong Wang, Kiril Gashteovski, Carolin Lawrence, Sean Welleck, Graham Neubig
| Challenge: | Prior studies have focused on developing effective data generation methods, but lack systematic comparison of different LMs as data generators in a unified setting. |
| Approach: | They propose to use a benchmark to compare language models' data generation abilities against a set of standardized settings and metrics. |
| Outcome: | The proposed benchmark provides standardized settings and metrics to evaluate LMs’ data generation abilities. |
The Butterfly Effect of Model Editing: Few Edits Can Trigger Large Language Models Collapse (2024.findings-acl)
Copied to clipboard
| Challenge: | Even a single edit can trigger model collapse, manifesting as significant performance degradation in various benchmark tasks. |
| Approach: | They propose to use perplexity as a surrogate metric to determine whether an edited model's performance is affected by a single edit. |
| Outcome: | The proposed method shows that even a single edit can cause model collapse, manifesting as significant performance degradation in various benchmark tasks. |
Narrative Text Generation with a Latent Discrete Plan (2020.findings-emnlp)
Copied to clipboard
| Challenge: | Prior work on story generation has focused on generating consistent stories via story outlines using keywords or key phrases. |
| Approach: | They propose a deep latent variable model that first samples a sequence of anchor words, one per sentence in the story, as part of its generative process. |
| Outcome: | The proposed model gets favorable scores when evaluated on perplexity, diversity, and control of story via discrete plan. |
Linguistic Frameworks Go Toe-to-Toe at Neuro-Symbolic Language Modeling (2022.naacl-main)
Copied to clipboard
| Challenge: | Existing models of language understanding are based on explicit representations of hierarchical structure, but there are good reasons to doubt that they can be said to understand language in any meaningful way. |
| Approach: | They examine whether syntactic and semantic graph representations can complement and improve neural language modeling. |
| Outcome: | The proposed model outperforms pretrained models on English WSJ in perplexity and other metrics. |
LexTempus: Enhancing Temporal Generalizability of Legal Language Models Through Dynamic Mixture of Experts (2025.acl-long)
Copied to clipboard
| Challenge: | a rapid evolution of legal concepts requires that legal language models adapt swiftly accounting for the temporal dynamics. |
| Approach: | They propose a dynamic mixture of experts model that explicitly models the temporal evolution of legal language in an online learning framework. |
| Outcome: | The proposed model can model the temporal evolution of legal language without forgetting past knowledge. |
Long-Tail Crisis in Nearest Neighbor Language Models (2025.findings-naacl)
Copied to clipboard
| Challenge: | Prior studies have shown that kNN-LM can retrieve long-tail contexts, leaving the model’s performance underexplored in estimating the probabilities of long-tailed target tokens. |
| Approach: | They investigate the behavior of kNN-LM on low-frequency tokens, examining prediction probability, retrieval accuracy, and token distribution in the datastore. |
| Outcome: | The proposed model improves the perplexity of given text by directly accessing a large datastore built from any text data during inference. |
Multi-source Neural Topic Modeling in Multi-view Embedding Spaces (2021.naacl-main)
Copied to clipboard
| Challenge: | Recent work has used pre-trained word embeddings to address data sparsity in short-text or small document collections. |
| Approach: | They propose a neural topic modeling framework using multi-view embedding spaces to improve topic quality and deal with polysemy. |
| Outcome: | The proposed framework improves topic quality and deal with polysemy. |
CItruS: Chunked Instruction-aware State Eviction for Long Sequence Modeling (2024.emnlp-main)
Copied to clipboard
| Challenge: | Recent studies have found that information relevant to the next token prediction task accumulates in the hidden representations of just a few tokens. |
| Approach: | They propose a method that integrates attention preferences useful for a downstream task into the eviction process of hidden states. |
| Outcome: | The proposed method performs better on comprehension and retrieval tasks while preserving language modeling perplexity. |
DeVisE: Towards the Behavioral Testing of Medical Large Language Models (2026.findings-eacl)
Copied to clipboard
| Challenge: | Existing evaluations of large language models do not reveal whether their outputs reflect genuine medical reasoning or superficial correlations. |
| Approach: | They propose a framework that probes fine-grained clinical understanding through controlled counterfactuals. |
| Outcome: | The proposed framework is based on demographic and vital signs data from the ICU discharge notes of patients in the intensive care unit (MIMIC-IV). |
PSC: Extending Context Window of Large Language Models via Phase Shift Calibration (2024.emnlp-main)
Copied to clipboard
| Challenge: | Large-scale language models (LLMs) have shown impressive results across a variety of tasks. |
| Approach: | They propose a module for calibrating the frequencies predefined by existing methods . they conducted extensive experiments across multiple models and tasks . |
| Outcome: | The proposed method reduces perplexity as the context window size is varied from 16k to 32k and up to 64k. |
Controlled Text Generation for Large Language Model with Dynamic Attribute Graphs (2024.findings-acl)
Copied to clipboard
| Challenge: | Controlled Text Generation (CTG) aims to produce texts that exhibit specific desired attributes. |
| Approach: | They propose a pluggable CTG framework for Large Language Models to control text . they use attribute scorers to evaluate attributes of sentences and construct dynamic attribute graphs . |
| Outcome: | The proposed framework achieves a peak improvement of 19.29% over baseline methods in two tasks. |
Code-switched Language Models Using Dual RNNs and Same-Source Pretraining (D18-1)
Copied to clipboard
| Challenge: | Using recurrent neural networks to build language models for code-switched text is an important problem with implications to downstream applications such as speech recognition and machine translation. |
| Approach: | They propose a novel recurrent neural network unit with dual components that focus on each language in the code-switched text separately and a generative model estimated using the training data. |
| Outcome: | The proposed techniques yield significant reductions in perplexity on Mandarin-English task and improve on baseline models. |
Reconsidering the Past: Optimizing Hidden States in Language Models (2021.findings-emnlp)
Copied to clipboard
| Challenge: | In this paper, we present Hidden-State Optimization (HSO) for language models at inference time. |
| Approach: | They propose a method that uses the log-probability gradient to update hidden states rather than the model parameters to improve the performance of transformer language models. |
| Outcome: | The proposed method improves performance of transformer language models at inference time. |
Neural Topic Model with Reinforcement Learning (D19-1)
Copied to clipboard
| Challenge: | Experimental results show superior performance on perplexity and topic coherence measures compared to state-of-the-art topic models. |
| Approach: | They propose to incorporate topic coherence measures as reward signals to guide the learning of a VAE-based topic model. |
| Outcome: | The proposed model is able to separating background words dynamically from topic words eliminating the pre-processing step of filtering infrequent and/or top frequent words, typically required for learning traditional topic models. |
Pseudo-label Data Construction Method and Syntax-enhanced Model for Chinese Semantic Error Recognition (2025.coling-main)
Copied to clipboard
| Challenge: | Existing research on Chinese text error recognition has focused on pre-trained models, but training them from scratch is time-consuming and laborious. |
| Approach: | They propose a method for Chinese Semantic Error Recognition that generates pseudo-labels for augmented samples based on perplexity and model respectively. |
| Outcome: | The proposed method surpasses existing models in Chinese text error recognition due to Chinese semantics' complexity. |
QUITO-X: A New Perspective on Context Compression from the Information Bottleneck Theory (2025.findings-emnlp)
Copied to clipboard
Yihang Wang, Xu Huang, Bowen Tian, Yueyang Su, Lei Yu, Huaming Liao, Yixing Fan, Jiafeng Guo, Xueqi Cheng
| Challenge: | Existing methods for compressing context by removing redundant tokens are inconsistent with the objective of retaining the most important tokens when conditioning on a given query. |
| Approach: | They propose a method that uses information bottleneck theory to compress context . they propose to remove redundant tokens using metrics such as self-information or perplexity . |
| Outcome: | The proposed method achieves a 25% increase in compression rate compared to the state-of-the-art . |
How to Train Long-Context Language Models (Effectively) (2025.acl-long)
Copied to clipboard
| Challenge: | a new study shows that language models can process extremely long contexts with minimal training. |
| Approach: | They use supervised fine-tuning and continued training to evaluate a language model's long-context capabilities. |
| Outcome: | The proposed model outperforms Llama-3.1-8B-Instruct on most long-context tasks . the model can process 512K tokens, one of the longest context windows of LMs . |
Dynamic Structured Neural Topic Model with Self-Attention Mechanism (2023.findings-acl)
Copied to clipboard
| Challenge: | Recent topic models that capture the time-series evolution of topics assume that topics evolve independently without interaction. |
| Approach: | They propose a dynamic structured neural topic model which captures topic dependencies while capturing their dependencies. |
| Outcome: | The proposed model outperforms a prior dynamic embedded topic model regarding perplexity and coherence while maintaining sufficient diversity across topics. |
Revisiting Syllables in Language Modelling and Their Application on Low-Resource Machine Translation (2022.coling-1)
Copied to clipboard
| Challenge: | Language modelling and machine translation tasks mostly use subword or character inputs, but syllables are rarely used. |
| Approach: | They explore the potential of syllables for open-vocabulary language modelling in 21 languages. |
| Outcome: | The proposed method outperforms characters and subwords in a non-related and low-resource language pair. |
Lessons Learned from GPT-SW3: Building the First Large-Scale Generative Language Model for Swedish (2022.lrec-1)
Copied to clipboard
Ariel Ekgren, Amaru Cuba Gyllensten, Evangelia Gogoulou, Alice Heiman, Severine Verlinden, Joey Öhman, Fredrik Carlsson, Magnus Sahlgren
| Challenge: | a prerequisite for building large-scale generative models for other languages is access to large amounts of high-quality text data and powerful computational resources. |
| Approach: | They present a 3.5 billion parameter autoregressive language model, trained on a 100 GB Swedish corpus. |
| Outcome: | The proposed model performs well on a 100 GB Swedish corpus and is competent in comparison with existing models of similar size. |
Divergent Token Metrics: Measuring degradation to prune away LLM components – and optimize quantization (2024.naacl-long)
Copied to clipboard
Björn Deiseroth, Max Meuer, Nikolas Gritsch, Constantin Eichenberg, Patrick Schramowski, Matthias Aßenmacher, Kristian Kersting
| Challenge: | Large Language Models (LLMs) have reshaped natural language processing with impressive capabilities, but their ever-increasing size has raised concerns about their effective deployment and the need for LLM compression. |
| Approach: | This study introduces the Divergent Token Metrics (DTMs) that measure token divergences that allow deeper insights into the subtleties of model compression. |
| Outcome: | The proposed measures can identify outliers and improve performance in the sparseness of the LLMs. |
Too Big to Fail: Larger Language Models are Disproportionately Resilient to Induction of Dementia-Related Linguistic Anomalies (2024.findings-acl)
Copied to clipboard
| Challenge: | Existing studies show that the attention mechanism in transformer-based NLMs may present an analogue to the notions of cognitive and brain reserve. |
| Approach: | They propose a bidirectional ablation method that masks attention heads to display degradation of similar magnitude to masking in smaller models. |
| Outcome: | The proposed method exhibits properties attributed to the concepts of cognitive and brain reserve in human brain studies. |
On the Effect of Pretraining Corpora on In-context Learning by a Large-scale Language Model (2022.naacl-main)
Copied to clipboard
Seongjin Shin, Sang-Woo Lee, Hwijeen Ahn, Sungdong Kim, HyoungSeok Kim, Boseop Kim, Kyunghyun Cho, Gichang Lee, Woomyoung Park, Jung-Woo Ha, Nako Sung
| Challenge: | Recent studies on large-scale in-context language models have reported successful in-const zero- and few-shot learning ability. |
| Approach: | They investigate the effects of the pretraining corpus on in-context learning in a Korean-centric model. |
| Outcome: | The study shows that pretraining corpus size does not determine in-context learning ability . the findings suggest that in-constext learning is not always competitive . |
Training Language Models with Memory Augmentation (2022.emnlp-main)
Copied to clipboard
| Challenge: | Existing methods for training memory-augmented language models only introduce mem-ories at testing time or represent them using a separately trained encoder. |
| Approach: | They propose a training approach that directly takes in-batch examples as accessible memory and new methods for memory construction and data batching that are used for adapting to different sets of memories at testing time. |
| Outcome: | The proposed approach reduces perplexity from 18.70 to 15.37 on multiple language modeling and machine translation benchmarks. |
ARCQuant: Boosting NVFP4 Quantization with Augmented Residual Channels for LLMs (2026.acl-long)
Copied to clipboard
| Challenge: | NVFP4 supports fine-grained block isolation, 4-bit quantization errors and mixed-precision approaches . ARCQuant boosts NVFO4 performance via Augmented Residual Channels . |
| Approach: | They propose a framework that boosts NVFP4 performance via Augmented Residual Channels. |
| Outcome: | ARCQuant boosts NVFP4 performance via Augmented Residual Channels . the proposed framework achieves state-of-the-art accuracy comparable to full-precision baselines compared to FP16 . |
Domain Private Transformers for Multi-Domain Dialog Systems (2023.findings-emnlp)
Copied to clipboard
| Challenge: | Large general purpose language models have demonstrated impressive performance across many different domains, but their outputs are not guaranteed to stay within the domain of a given input prompt. |
| Approach: | They propose to quantify how likely a conditional language model will leak across domains by defining domain privacy as a way to fine-tune a model's privacy. |
| Outcome: | The proposed method has comparable resiliency to methods adapted from recent literature on differentially private language models. |
SCVQ: Sparse-Compensated Vector Quantization for Large Language Models (2026.acl-long)
Copied to clipboard
| Challenge: | Existing vector quantization methods incur inference overhead due to massive codebook storage and intensive index lookups. |
| Approach: | They propose a framework for vector quantization that incorporates a salience-aware weighted K-means clustering scheme with symmetry constraints to reduce codebook size and indexing costs. |
| Outcome: | The proposed framework achieves a perplexity of 5.78 on WikiText-2 for LLaMA-2-7B at 2-bit quantization while delivering a 1.4 speedup over existing baselines. |
A Cognitive Regularizer for Language Modeling (2021.acl-long)
Copied to clipboard
| Challenge: | a uniform information density hypothesis is used to explain certain linguistic phenomena . a regularizer that encodes the UID hypothesis can be used for language training . |
| Approach: | They propose to augment the canonical MLE objective with a regularizer that encodes UID . they find that regularization consistently improves perplexity in language models . |
| Outcome: | The proposed hypothesis can be operationalized as an inductive bias for language modeling. |
Lower Perplexity is Not Always Human-Like (2021.acl-long)
Copied to clipboard
| Challenge: | Existing efforts to build human-like computational models have focused on English . a cross-lingual evaluation is needed to build such models, but current research has focused on Japanese . |
| Approach: | They re-examine an established generalization that lower perplexity is not always human-like in Japanese . they propose a cross-lingual evaluation to build human-type computational models . |
| Outcome: | The proposed model lacks universality and lower perplexity is not always human-like . the results suggest a cross-lingual evaluation will be necessary to build human-type models . |
Revisiting Simple Neural Probabilistic Language Models (2021.naacl-main)
Copied to clipboard
| Challenge: | Recent advances in language modeling have been driven not only by advances in neural architectures, but also through hardware and optimization improvements. |
| Approach: | They revisit the neural probabilistic language model (NPLM) of Bengio et al. (2003) which simply concatenates word embeddings within a fixed window and passes the result through a feed-forward network to predict the next word. |
| Outcome: | The proposed model performs better on word-level language model benchmarks than a baseline Transformer with short input contexts but struggles to handle long-term dependencies. |
LASS: A Novel and Economical Data Augmentation Framework Based on Language Models for Debiasing Opinion Summarization (2025.coling-main)
Copied to clipboard
| Challenge: | Existing methods to generate negative summaries are expensive and lack the capacity to generate large data sets. |
| Approach: | They propose a data augmentation framework based on LArge and Small language models for debiaSing opinion summarization that generates a small number of synthesized negative reviews by rewriting the positive text via a large language model. |
| Outcome: | The proposed framework can generate large numbers of negative reviews by rewriting the positive text using a large language model and training a disentangle reconstruction model based on the generated data. |
TemporalWiki: A Lifelong Benchmark for Training and Evaluating Ever-Evolving Language Models (2022.emnlp-main)
Copied to clipboard
Joel Jang, Seonghyeon Ye, Changho Lee, Sohee Yang, Joongbo Shin, Janghoon Han, Gyeonghun Kim, Minjoon Seo
| Challenge: | Language Models (LMs) become outdated as the world changes, a phenomenon called temporal misalignment. |
| Approach: | They propose a lifelong benchmark that utilizes the difference between consecutive snapshots of English Wikipedia and English Wikidata for training and evaluation. |
| Outcome: | The proposed benchmark can be trained on the difference between consecutive snapshots of English Wikipedia and English Wikidata for training and evaluation. |
Experimenting with Power Divergences for Language Modeling (D19-1)
Copied to clipboard
| Challenge: | Language models are an important component in many NLP tasks, where they provide prior knowledge on the language used. |
| Approach: | They propose to use power divergences to prioritize learning on frequent or rare words . they use a sample-based objective to approximate a softmax and noise-constrained estimate . |
| Outcome: | The proposed power divergences can be used to prioritize learning on the frequent or rare words and lead to general performance improvements. |
Shortformer: Better Language Modeling using Shorter Inputs (2021.acl-long)
Copied to clipboard
| Challenge: | Existing methods require computationally expensive relative position embeddings. |
| Approach: | They propose two methods that decrease input length to improve perplexity and perplexability. |
| Outcome: | The proposed methods speed up training by a factor of 1.65 and reduce memory usage. |
NACL: A General and Effective KV Cache Eviction Framework for LLM at Inference Time (2024.acl-long)
Copied to clipboard
Yilong Chen, Guoxia Wang, Junyuan Shang, Shiyao Cui, Zhenyu Zhang, Tingwen Liu, Shuohuan Wang, Yu Sun, Dianhai Yu, Hua Wu
| Challenge: | Large Language Models (LLMs) with extended context windows are expensive and infeasible on fixed memory hardware due to the surprisingly large memory consumption of KV Cache. |
| Approach: | They propose a general framework for long-context KV cache eviction that achieves more optimal and efficient evict in a single operation during the encoding phase. |
| Outcome: | The proposed framework improves performance on short- and long-text tasks by 80% and 76% respectively, reducing KV Cache by up to 5 with over 95% performance maintenance. |
Multilingual Nonce Dependency Treebanks: Understanding how Language Models Represent and Process Syntactic Structure (2024.naacl-long)
Copied to clipboard
| Challenge: | a number of studies have focused on making explicit the linguistic information encoded in language models (LMs) however, this method has been criticized for various reasons. |
| Approach: | They introduce a framework for creating nonce treebanks for multilingual UD corpora . they investigate word co-occurrence statistics and show how nonce data affects the performance of syntactic dependency probes. |
| Outcome: | The proposed framework satisfies syntactic argument structure and ensures grammaticality via language-specific rules. |
Can Data Diversity Enhance Learning Generalization? (2022.coling-1)
Copied to clipboard
| Challenge: | a diversity advanced actor-critical reinforcement learning framework is used to improve NLP generalization and accuracy. |
| Approach: | They introduce Diversity Advanced Actor-Critic reinforcement learning framework to improve NLP generalization and accuracy. |
| Outcome: | The proposed framework outperforms domain adaptation and generalization baselines without using any target domain knowledge. |
Language Modeling with Shared Grammar (P19-1)
Copied to clipboard
| Challenge: | Recent work on structure-aware models have shown promising results on language modeling, but how to incorporate structure knowledge on corpus without syntactic annotations remains an open problem. |
| Approach: | They propose a neural variational language model which enables the sharing of grammar knowledge among different corpora. |
| Outcome: | The proposed model converges significantly faster to lower perplexity on two popular benchmark datasets. |
Long Context is Not Long at All: A Prospector of Long-Dependency Data for Large Language Models (2024.acl-long)
Copied to clipboard
| Challenge: | Long-context modeling capabilities are important for large language models (LLMs) however, training LLMs with long context windows is insufficient since some samples do not exhibit strong semantic dependencies across long contexts. |
| Approach: | They propose a data mining framework ProLong that assigns each training sample with a long dependency score and ranks and filters them according to their results. |
| Outcome: | The proposed framework can rank and filter training samples that exhibit more powerful long-context modeling abilities. |
Data Curation Alone Can Stabilize In-context Learning (2023.acl-long)
Copied to clipboard
| Challenge: | In-context learning (ICL) is a new paradigm for few-shot learning with pretrainable large language models . however, randomly sampling examples from a training set leads to high variance in performance . |
| Approach: | They propose two methods to select training examples from a training set and then carefully curate them from corresponding subsets. |
| Outcome: | The proposed method improves accuracy over sampling from the entire training set. |
DOC-RAG: ASR Language Model Personalization with Domain-Distributed Co-occurrence Retrieval Augmentation (2024.lrec-main)
Copied to clipboard
| Challenge: | Extensive experiments on three user-specific speech-to-text tasks show that DOC-RAG significantly outperforms strong baselines with an 8-15% improvement in terms of perplexity and a 4-7% reduction in terms in terms . of Word Error Rates. |
| Approach: | They propose a domain-distributed co-occurrence augmentation approach to improve automatic speech recognition of rare word patterns in unseen domains by using n-gram co-existence distributions. |
| Outcome: | Experiments on three user-specific speech-to-text tasks show that DOC-RAG outperforms baselines with an 8-15% improvement in terms of perplexity and a 4-7% reduction in terms in terms . of Word Error Rates. |
Effects of sub-word segmentation on performance of transformer language models (2023.emnlp-main)
Copied to clipboard
| Challenge: | Language models are a fundamental task in natural language processing, but few studies focus on the effect of sub-word segmentation on the performance of models. |
| Approach: | They compare GPT and BERT models trained with statistical segmentation algorithm BPE to unsupervised morphological segmentation algorithms Morfessor and StateMorph. |
| Outcome: | The proposed model trains for several languages and compares them with two unsupervised morphological segmentation algorithms. |
TF-LM: TensorFlow-based Language Modeling Toolkit (L18-1)
Copied to clipboard
| Challenge: | Existing deep learning tools offer building blocks but training and building models takes time and knowledge. |
| Approach: | They propose to make available LSTM language models trained on Dutch texts and English benchmarks. |
| Outcome: | The proposed model can be used to test the perplexity, predict the next word(s), re-score hypotheses or generate debugging files for interpolation with n-gram models. |
Shall We Pretrain Autoregressive Language Models with Retrieval? A Comprehensive Study (2023.emnlp-main)
Copied to clipboard
Boxin Wang, Wei Ping, Peng Xu, Lawrence McAfee, Zihan Liu, Mohammad Shoeybi, Yi Dong, Oleksii Kuchaiev, Bo Li, Chaowei Xiao, Anima Anandkumar, Bryan Catanzaro
| Challenge: | a recent study shows that retrieval-augmented LMs can improve text generation quality and accuracy. |
| Approach: | They propose a model that reproduces RETRO parameters while retrieving a text corpus . they find RETRO outperforms GPT on text generation with less repetition . |
| Outcome: | The proposed model outperforms standard retrieval-augmented GPT and retrieval augmented GTP on text generation and accuracy tasks. |
Explaining and Improving Contrastive Decoding by Extrapolating the Probabilities of a Huge and Hypothetical LM (2024.emnlp-main)
Copied to clipboard
| Challenge: | Contrastive decoding (CD) improves the next-token distribution of a large expert language model (LM) using a small amateur LM. |
| Approach: | They propose a new unsupervised decoding method called Asymptotic Probability Decoding (APD) that extrapolates the probability curves from the LMs of different sizes to infer the asymptototic probabilities from an infinitely large LM. |
| Outcome: | The proposed method improves the next-token distribution of a large expert language model using a small amateur LM. |
GRASPrune: Global Gating for Budgeted Structured Pruning of Large Language Models (2026.acl-long)
Copied to clipboard
| Challenge: | Large language models are expensive to serve because dense FFN blocks, multi-head attention, and KV caches dominate memory. |
| Approach: | They propose a global budgeted structured pruning framework that prunes FFN channels and attention KV head groups under a single global parameter budget. |
| Outcome: | The proposed model removes 50% of parameters and achieves 12.18 perplexity on WikiText-2 while maintaining competitive average zero-shot accuracy on five downstream benchmarks. |
Pyramidal Recurrent Unit for Language Modeling (D18-1)
Copied to clipboard
| Challenge: | Long short term memory units are powerful tools for language modeling, but their performance can be limited by the number of parameters. |
| Approach: | They propose a pyramidal recurrent unit which enables learning representations in high dimensional space with more generalization power and fewer parameters. |
| Outcome: | The proposed model outperforms existing models with different gating mechanisms and transformations on word-level language modeling tasks. |
Modeling Online Discourse with Coupled Distributed Topics (D18-1)
Copied to clipboard
| Challenge: | a topic model that incorporates structural relationships connecting documents in socially generated corpora is of limited application in the sciences. |
| Approach: | They propose a topic model that incorporates structural relationships connecting documents in socially generated corpora, such as online forums. |
| Outcome: | The proposed model captures discursive interactions along observed reply links and integrates latent distributed representations in a deep architecture. |
MADial-Bench: Towards Real-world Evaluation of Memory-Augmented Dialogue Generation (2025.naacl-long)
Copied to clipboard
| Challenge: | Existing evaluation metrics for memory-augmented dialogue systems lack practical value . current evaluation methods only consider passive memory retrieval while ignoring diverse memory recall with rich triggering factors. |
| Approach: | They propose to use long-term memory to create human-like dialogues using chatbots. |
| Outcome: | The proposed benchmark covers memory retrieval and memory recognition tasks with both passive and proactive memory recall data. |
Dual Fixed-Size Ordinally Forgetting Encoding (FOFE) for Competitive Neural Language Models (D18-1)
Copied to clipboard
| Challenge: | In this paper, we propose a new approach to employ the fixed-size ordinally-forgetting encoding (FOFE) in neural languages modelling, called dual-FOFE. |
| Approach: | They propose a new approach to employ the fixed-size ordinally-forgetting encoding (FOFE) in neural languages modelling, called dual-FOFE. |
| Outcome: | The proposed method significantly reduces the complexity and improves perplexity by 10% over the original FOFE model. |
CartesianMoE: Boosting Knowledge Sharing among Experts via Cartesian Product Routing in Mixture-of-Experts (2025.naacl-long)
Copied to clipboard
Zhenpeng Su, Xing W, Zijia Lin, Yizhe Xiong, Minxuan Lv, Guangyuan Ma, Hui Chen, Songlin Hu, Guiguang Ding
| Challenge: | Large language models (LLMs) have been attracting much attention due to their impressive performance in all kinds of downstream tasks. |
| Approach: | They propose a mix-of-experts model that allows the model size to grow without raising training costs. |
| Outcome: | The proposed model outperforms existing models in perplexity and robustness tests. |
Like a Baby: Visually Situated Neural Language Acquisition (P19-1)
Copied to clipboard
| Challenge: | A multi-modal neural architecture outperforms its equivalent trained on language alone with a 2% decrease in perplexity . |
| Approach: | They propose to use visual context to train neural language models to perform next-word prediction. |
| Outcome: | The proposed model outperforms its equivalent trained on language with 2% decrease in perplexity even when no visual context is available at test. |
DB-LLM: Accurate Dual-Binarization for Efficient LLMs (2024.findings-acl)
Copied to clipboard
Hong Chen, Chengtao Lv, Liang Ding, Haotong Qin, Xiabin Zhou, Yifu Ding, Xuebo Liu, Min Zhang, Jinyang Guo, Xianglong Liu, Dacheng Tao
| Challenge: | Existing methods for ultra-low bit quantization cause severe accuracy drops . a novel Dual-Binarization method is proposed for efficient Large Language Models . |
| Approach: | They propose a Dual-Binarization method that takes 2-bit-width and binarization into account . they propose DB-LLM, which uses a 2-bit binarized weighted model to represent weights efficiently . |
| Outcome: | The proposed method surpasses the current State-of-the-Art in ultra-low bit quantization and achieves 20% reduction in computational consumption compared to the SOTA method under the same bit-width. |
Temperature-scaling surprisal estimates improve fit to human reading times – but does it do so for the “right reasons”? (2024.acl-long)
Copied to clipboard
| Challenge: | a wide body of evidence shows that human language processing difficulty is predicted by the information-theoretic measure surprisal, a word’s negative log probability in context. |
| Approach: | They propose to use large language models to predict the surprisal of a word's negative log probability in context to test their predictive power. |
| Outcome: | The proposed model can be significantly more accurate than humans because it has more data. |
Learning Dynamic Contextualised Word Embeddings via Template-based Temporal Adaptation (2023.acl-long)
Copied to clipboard
| Challenge: | Existing methods for learning dynamic contextualised word embeddings do not capture temporal semantic variations of words. |
| Approach: | They propose a method for learning DCWEs by time-adapting a pretrained Masked Language Model using time-sensitive templates. |
| Outcome: | The proposed method significantly reduces the perplexity of test sentences in C2 outperforming the current state-of-the-art. |
How poor is the stimulus? Evaluating hierarchical generalization in neural networks trained on child-directed speech (2023.acl-long)
Copied to clipboard
| Challenge: | LSTMs and Transformers perform well at capturing the surface statistics of child-directed speech, but both model types generalize in a way consistent with an incorrect linear rule than the correct hierarchical rule. |
| Approach: | They train LSTMs and Transformers on text from the CHILDES corpus and evaluate what they learn about English yes/no questions. |
| Outcome: | The proposed models perform well at capturing the surface statistics of child-directed speech, but generalize more consistent with an incorrect linear rule than the correct hierarchical rule. |
Reverse-Engineering the Reader (2024.emnlp-main)
Copied to clipboard
| Challenge: | Existing studies have sought to determine to what extent language models can serve as useful models of human cognition by aligning them to human psychometric data. |
| Approach: | They propose a method to fine-tune a language model to implicitly optimize parameters of a linear regressor that directly predicts humans’ reading times of in-context linguistic units. |
| Outcome: | The proposed technique improves language models’ psychometric predictive power but also its perplexity on held-out test data. |
Bit-Flip Error Resilience in LLMs: A Comprehensive Analysis and Defense Framework (2025.emnlp-main)
Copied to clipboard
Yuhang Chen, Zhen Tan, Ajay Kumar Jaiswal, Huaizhi Qu, Xinyu Zhao, Qi Lin, Yu Cheng, Andrew Kwong, Zhichao Cao, Tianlong Chen
| Challenge: | Bit-flip errors (BFEs) are hardware faults where individual bits in memory or processing units are unintentionally flipped. |
| Approach: | They propose a novel defense strategy to mitigate bit-flip errors (BFEs) they propose bfe protection and a self-correction mechanism to minimize performance degradation . |
| Outcome: | The proposed defense strategy minimizes performance degradation while significantly improving robustness against BFEs. |
Neural Mixed Counting Models for Dispersed Topic Discovery (2020.acl-main)
Copied to clipboard
| Challenge: | Existing methods for inference of parameter parameters are time-consuming and difficult to use. |
| Approach: | They propose two efficient neural mixed counting models that use the negative binomial distribution as the prior for dispersed topic discovery. |
| Outcome: | The proposed models outperform state-of-the-art models in terms of perplexity and topic coherence on real-world datasets. |
UPER: Boosting Multi-Document Summarization with an Unsupervised Prompt-based Extractor (2022.coling-1)
Copied to clipboard
| Challenge: | Multi-Document Summarization (MDS) uses the extract-then-abstract paradigm, which extracts a relatively short meta-document and then feeds it into the deep neural networks to generate an abstract. |
| Approach: | They propose to use pre-trained language models to calculate document and keyword’s perplexity to boost other metrics for evaluating a document’s salience. |
| Outcome: | The proposed method can be applied as a plug-in to boost other metrics for evaluating a document’s salience, thus improving the subsequent abstract generation. |
Softmax Bottleneck Makes Language Models Unable to Represent Multi-mode Word Distributions (2022.acl-long)
Copied to clipboard
| Challenge: | Existing language models can't predict the probability distribution over next words or masked words because of a single hidden state. |
| Approach: | They propose a multi-facet softmax to address the limitations of softmax . they propose LMs that can predict the next word distribution better as their sizes grow . |
| Outcome: | The proposed model achieves two-fold improvements in the perplexity of GPT-2 and BERT. |
Hit the Sweet Spot! Span-Level Ensemble for Large Language Models (2025.coling-main)
Copied to clipboard
| Challenge: | a recent study focused on sample-level and token-level ensembles, which hinder dynamic correction and enhancement of outputs during the generation process. |
| Approach: | They propose a span-level ensemble method that balances real-time adjustments and accurate ensemble decisions. |
| Outcome: | The proposed method improves performance across language generation tasks significantly. |
I Learn Better If You Speak My Language: Understanding the Superior Performance of Fine-Tuning Large Language Models with LLM-Generated Responses (2024.emnlp-main)
Copied to clipboard
| Challenge: | Recent research has demonstrated that a large language model (LLM) can generate training data for another LLM, or for creating supplementary training materials, such as rationales. |
| Approach: | They conduct an in-depth investigation to understand why fine-tuning an LLM with responses generated by a LLM often yields better results than using responses generated from humans. |
| Outcome: | The proposed approach can be used to transfer knowledge from a larger model to a smaller one, or for creating supplementary training materials, such as rationales. |
Monotonic Paraphrasing Improves Generalization of Language Model Prompting (2024.findings-emnlp)
Copied to clipboard
| Challenge: | Large language models (LLMs) have demonstrated remarkable proficiency in zero-shot decision making and instruction following. |
| Approach: | They propose an end-to-end decoding strategy that paraphrases given prompts or instructions into their lower perplexity counterparts based on an ensemble of a paraphrase LM for prompt rewriting, and a target LM that constrains the generation for lower perxity. |
| Outcome: | The proposed method can efficiently paraphrase the original prompt without altering its semantic meaning while decreasing the perplexity of each generation as calculated by the target LM. |
Change Is the Only Constant: Dynamic LLM Slicing based on Layer Redundancy (2024.findings-emnlp)
Copied to clipboard
| Challenge: | Using dynamic slicing, large language models can be used to reduce computational burden and improve performance. |
| Approach: | They propose a dynamic layer-specific pruning approach that leverages the newly proposed Layer Redundancy score to prune parts of individual layers based on redundancy. |
| Outcome: | The proposed method maintains and enhances model performance over the SliceGPT baseline. |
Towards Automatic Evaluation of Dialog Systems: A Model-Free Off-Policy Evaluation Approach (2021.emnlp-main)
Copied to clipboard
| Challenge: | Existing methods for evaluation of dialog systems are expensive and not scalable . a framework for estimating human evaluation scores is proposed to bridge this gap . |
| Approach: | They propose a framework for estimating human evaluation scores based on off-policy evaluation . they use language quality metrics for single-turn response generation given a fixed context . |
| Outcome: | The proposed framework outperforms existing methods in terms of correlation with human evaluation scores. |
Exploiting Syntactic Structure for Better Language Modeling: A Syntactic Distance Approach (2020.acl-main)
Copied to clipboard
| Challenge: | incorporating syntactic structure into language models has been a challenge since the 1990s. |
| Approach: | They propose to use syntactic information to integrate syntastic structure into neural language models by providing ground truth parse trees as additional training signals. |
| Outcome: | The proposed model achieves lower perplexity and better quality when ground truth parse trees are provided as training signals. |
Interpolated Spectral NGram Language Models (P19-1)
Copied to clipboard
| Challenge: | Spectral methods for learning weighted non-deterministic automata have nice theoretical and algorithmic properties, but they have been challenging to get competitive results on language modeling tasks. |
| Approach: | They propose a method for scaling up spectral learning and use interpolated predictions that are optimized to maximize perplexity. |
| Outcome: | The proposed method matches state-of-the-art ngram models while being very fast to train. |
Spontaneous gestures encoded by hand positions improve language models: An Information-Theoretic motivated study (2023.findings-acl)
Copied to clipboard
| Challenge: | a key missing step is to explore whether the nonverbal information can be quantified. |
| Approach: | They explore whether incorporating gesture representations can improve the language model’s performance . they also examine whether spontaneous gestures demonstrate entropy rate constancy (ERC) . |
| Outcome: | The proposed model improves the performance of the mixed-modal language models against monologue video data. |
Exploring the Value of Personalized Word Embeddings (2020.coling-main)
Copied to clipboard
| Challenge: | a subset of words belonging to specific psycholinguistic categories vary more in their representations across users . combining generic and personalized word embeddings yields the best performance . |
| Approach: | They propose personalized word embeddings and compare their performance to generic ones . they show that personalized word representations can be leveraged for improved performance . |
| Outcome: | The proposed model can be used for authorship attribution. |
Improving Neural Story Generation by Targeted Common Sense Grounding (D19-1)
Copied to clipboard
| Challenge: | Recent advances in language modeling have yielded thematic and stylistic coherence in story generation through large scale pretraining of Transformer models. |
| Approach: | They propose a multi-task learning scheme to achieve better common sense reasoning in language models by leveraging auxiliary training signals from datasets designed to provide common sense grounding. |
| Outcome: | The proposed model achieves improved common sense reasoning and state-of-the-art perplexity on the WritingPrompts dataset. |
The language of prompting: What linguistic properties make a prompt successful? (2023.findings-emnlp)
Copied to clipboard
| Challenge: | Recent studies show that pretraining and instruction-tuned LLMs can achieve impressive performance on a multitude of tasks. |
| Approach: | They propose to use a standard for prompting research to better understand linguistic properties of LLMs. |
| Outcome: | The proposed standard would improve the performance of pre-trained and instruction-tuned LLMs on a multitude of tasks. |
From Long to Lean: Performance-aware and Adaptive Chain-of-Thought Compression via Multi-round Refinement (2025.emnlp-main)
Copied to clipboard
| Challenge: | Chain-of-Thought reasoning introduces significant inference latency due to its verbosity. |
| Approach: | They propose a framework that leverages token elasticity phenomenon to progressively compress CoTs via multiround refinement. |
| Outcome: | The proposed method achieves an average accuracy improvement of 5.6% over state-of-the-art baselines while reducing CoT length by an average of 47 tokens and significantly lowering latency. |
Efficiently Selecting Response Generation Strategies for Synthetic Data Construction by Self-Aligned Perplexity (2025.findings-emnlp)
Copied to clipboard
| Challenge: | Using a small sample of data, we find that perplexity is suboptimal in characterizing “familiarity” . |
| Approach: | They propose a method that assesses a small subset of generated data to estimate suitability for a specific target LLM. |
| Outcome: | The proposed method assesses a small subset of generated data to estimate suitability for a specific target LLM. |
Generating Classical Chinese Poems from Vernacular Chinese (D19-1)
Copied to clipboard
| Challenge: | Existing models for classical Chinese poetry generation only allow users to use keywords to interfere with the meaning of generated poems. |
| Approach: | They propose a model to generate classical Chinese poems from vernacular . their model uses unsupervised machine translation to generate Chinese poems . human evaluation shows it can generate high-quality poems comparable to amateur poems - authors . |
| Outcome: | The proposed model improves the perplexity and BLEU of the proposed model compared with typical models and human evaluation shows it generates high-quality poems comparable to amateur poems. |
TEMA: Token Embeddings Mapping for Enriching Low-Resource Language Models (2024.emnlp-main)
Copied to clipboard
| Challenge: | Low-resource languages, that is, languages that do not have a massive amount of text, risk being almost excluded from the possibility of having good NLP applications. |
| Approach: | They propose an algorithm that maps token embeddings of a richly pre-trained model to a poorly trained model and creates a more complex model. |
| Outcome: | The proposed model reduces perplexity and is competitive or better for the most semantic tasks. |
Automatic Document Selection for Efficient Encoder Pretraining (2022.emnlp-main)
Copied to clipboard
| Challenge: | Pretraining language models is expensive and data-intensive, but can it be improved? Several studies have found that directly pretraining on task data is more effective . |
| Approach: | They propose to automatically identify smaller yet domain-representative subsets by pretraining a model on a target domain. |
| Outcome: | The proposed method outperforms random selection on perplexity and downstream tasks with 20x less data and 3x fewer training iterations and 2x less estimated cloud compute cost. |
Dual Debiasing for Noisy In-Context Learning for Text Generation (2025.findings-acl)
Copied to clipboard
| Challenge: | Existing methods detect noisy annotations by ranking local perplexities, but this assumption breaks down when the noise ratio is high and many demonstrations are flawed. |
| Approach: | They propose a method that uses synthesized neighbors to explicitly correct perplexity estimates, yielding a robust Sample Cleanliness Score. |
| Outcome: | The proposed method is comparable to a fully clean demonstration corpus and performs well even with noise ratios as high as 0.8. |
Good-Enough Compositional Data Augmentation (2020.acl-main)
Copied to clipboard
| Challenge: | a proposed data augmentation protocol provides a compositional inductive bias in conditional and unconditional sequence models. |
| Approach: | They propose a data augmentation protocol that provides a compositional inductive bias in conditional and unconditional sequence models by replacing discontinuous fragments with other fragments that appear in at least one similar environment. |
| Outcome: | The proposed protocol reduces error rate by 87% on diagnostic tasks and 16% on semantic parsing tasks. |
Demystifying Prompts in Language Models via Perplexity Estimation (2023.findings-emnlp)
Copied to clipboard
| Challenge: | Language models can be prompted to perform a wide variety of tasks with zero- and few-shot learning. |
| Approach: | They propose a method to automatically extend a small seed set of manually written prompts by paraphrasing with GPT3 and backtranslation. |
| Outcome: | The proposed method extends a small seed set of manually written prompts by paraphrasing with GPT3 and backtranslation. |
Improving Low Compute Language Modeling with In-Domain Embedding Initialisation (2020.emnlp-main)
Copied to clipboard
| Challenge: | Existing approaches to train language models on in-domain data are limited. |
| Approach: | They propose to initialise and freeze in-domain embeddings to provide a useful representation of rare words in English . they find that the standard configuration is not optimal when rare words are present . |
| Outcome: | The proposed approach improves language modeling by providing a useful representation of rare words in English. |
Hints on the data for language modeling of synthetic languages with transformers (2023.acl-long)
Copied to clipboard
| Challenge: | Language Models (LMs) are becoming more useful for providing representations for NLP applications. |
| Approach: | They evaluated whether the critical amount of data varies for different morphological typologies . they found that the size of the vocabulary due to morphology is directly correlated with LM perplexity . |
| Outcome: | The proposed method reduces perplexity by more than a half for a polysynthetic language like Quechua . |
Debiasing the Fine-Grained Classification Task in LLMs with Bias-Aware PEFT (2025.acl-long)
Copied to clipboard
| Challenge: | Existing methods to mitigate label biases such as retraining, post-hoc adjustment, and parameter-efficient fine-tuning fail to address prediction propensity and discriminative ability biase. |
| Approach: | They propose a bias-aware optimization framework that incorporates two distinct label balance constraints with a PEFT strategy targeting an intermediate layer to mitigate this issue. |
| Outcome: | The proposed approach outperforms or matches the performance of full-parameter fine-tuning and LoRA, achieving superior results with lower perplexity. |
A Comprehensive Evaluation of Quantization Strategies for Large Language Models (2024.findings-acl)
Copied to clipboard
| Challenge: | Quantization studies have focused on instruction-tuned LLMs, leaving their performance on other benchmarks unclear. |
| Approach: | They propose a framework to evaluate quantized large language models using four dimensions . they propose to reduce the bits needed for model weights or activations with minimal performance loss . |
| Outcome: | The proposed framework can retain comparable performance to non-quantized LLMs on most benchmarks. |
SampleMix: A Sample-wise Pre-training Data Mixing Strategy by Coordinating Data Quality and Diversity (2025.findings-emnlp)
Copied to clipboard
Xiangyu Xi, Deyang Kong, Jian Yang, Jiawei Yang, Zhengyu Chen, Wei Wang, Jingang Wang, Xunliang Cai, Shikun Zhang, Wei Ye
| Challenge: | Existing methods for pretraining data mixing for large language models neglect significant inter-domain overlaps and commonalities, failing to control the global diversity of the constructed training dataset. |
| Approach: | They propose a sample-wise data mixture approach that performs global cross-domain sampling by systematically evaluating the quality and diversity of each sample. |
| Outcome: | The proposed method exceeds existing domain-based methods in multiple downstream tasks and perplexity assessments. |
Propagation and Pitfalls: Reasoning-based Assessment of Knowledge Editing through Counterfactual Tasks (2024.findings-acl)
Copied to clipboard
| Challenge: | Existing knowledge editing methods struggle to effectively propagate updates to interconnected facts, limiting the performance of reasoning tasks based on these updated facts. |
| Approach: | They propose a reasoning-based benchmark, ReCoE, which covers six common reasoning schemes in the real world. |
| Outcome: | The proposed reasoning-based benchmark shows that current models struggle to propagate updated knowledge within reasoning schemes. |
Assigning Distinct Roles to Quantized and Low-Rank Matrices Toward Optimal Weight Decomposition (2025.findings-acl)
Copied to clipboard
| Challenge: | Existing joint optimization methods prioritize one component at the expense of the other, resulting in suboptimal decompositions that fail to leverage each component’s unique strengths. |
| Approach: | They introduce Outlier-Driven Low-Rank Initialization (ODLRI) which assigns low-rank components the specific role of capturing activation-sensitive weights. |
| Outcome: | Experiments on Llama2 (7B, 13B, 70B, and Mistral-7B) and Llma3-8B show that incorporating ODLRI into the joint optimization framework reduces activation-aware error, minimizes quantization scale, and improves perplexity and zero-shot accuracy in low-bit settings. |
PersonaLM: Language Model Personalization via Domain-distributed Span Aggregated K-Nearest N-gram Retrieval Augmentation (2023.findings-emnlp)
Copied to clipboard
| Challenge: | Existing language modeling tools for automatic speech recognition (ASR) are difficult to personalize. |
| Approach: | They propose a domain-distributed Span-Aggregated K-nearest N-gram retrieval augmentation to improve language modeling for automatic speech recognition (ASR) personalization. |
| Outcome: | The proposed model outperforms baselines on Wikitext-103, UserLibri, and ASAP datasets with a 10-16% improvement in perplexity and a 5-8% reduction in word error rates. |
Rephrasing the Web: A Recipe for Compute and Data-Efficient Language Modeling (2024.acl-long)
Copied to clipboard
| Challenge: | Large language model pre-training is infeasible due to the large compute costs and duration associated with pre- training and the impending scarcity of high-quality data on the web. |
| Approach: | They propose to use an off-the-shelf instruction-tuned model prompted to paraphrase documents on the web in specific styles such as “like Wikipedia” or in “question-answer format” to jointly pre-train LLMs on real and synthetic rephrases. |
| Outcome: | The proposed model speeds up pre-training by 3x on the C4 dataset, and improves perplexity by 50% on average across different subsets of the Pile. |
Minority Positive Sampling for Switching Points - an Anecdote for the Code-Mixing Language Modeling (2020.lrec-1)
Copied to clipboard
| Challenge: | Multilingual people code-mix using English phonetic typing and insertion of anglicisms in their native language. |
| Approach: | They propose to use minority positive sampling to selectively induce more sample to achieve better performance. |
| Outcome: | The proposed model performs better than other models, but switching points are the main challenge . |
Noisy Pair Corrector for Dense Retrieval (2023.findings-emnlp)
Copied to clipboard
| Challenge: | Existing dense retrieval models assume that query-document pairs are exactly matched, resulting in mismatched-pair noise. |
| Approach: | They propose a novel approach to train an effective model with mismatched-pair noise. |
| Outcome: | The proposed model performs well on natural question and triviaQA, code-search benchmarks and SO-DS. |
Training Trajectories of Language Models Across Scales (2023.acl-long)
Copied to clipboard
Mengzhou Xia, Mikel Artetxe, Chunting Zhou, Xi Victoria Lin, Ramakanth Pasunuru, Danqi Chen, Luke Zettlemoyer, Veselin Stoyanov
| Challenge: | Scaling up language models has led to unprecedented performance gains, but little is understood about how the training dynamics change as models get larger. |
| Approach: | They analyze the training checkpoints of different-sized OPT models on next-token prediction, sequence-level generation and downstream tasks. |
| Outcome: | The results show that language models of different sizes learn more during training . small models halt at hallucinations, larger ones learn to assign lower probabilities . |
CritiQ: Mining Data Quality Criteria from Human Preferences (2025.acl-long)
Copied to clipboard
Honglin Guo, Kai Lv, Qipeng Guo, Tianyi Liang, Zhiheng Xi, Demin Song, Qiuyinzhe Zhang, Yu Sun, Kai Chen, Xipeng Qiu, Tao Gui
| Challenge: | Existing methods to train language models rely on manual design, perplexity, or careful prompt engineering. |
| Approach: | They propose a method that automatically mines criteria from human preferences for data quality with only 30 human-annotated pairs and performs efficient data selection. |
| Outcome: | The proposed method improves on human-annotated test sets and shows high accuracy on code, math, and logic domains. |
Evaluating distillation methods for data-efficient syntax learning (2025.findings-emnlp)
Copied to clipboard
| Challenge: | knowledge distillation (KD) targeting attention should selectively accelerate syntax acquisition, a study finds . logit-based KD dramatically improves data-efficiency, attention-based one provides minimal benefit even for syntactic tasks. |
| Approach: | a study predicts that knowledge distillation targeting attention should selectively accelerate syntax acquisition . a systolic analysis of student models compared to logit-based knowledge distillations . |
| Outcome: | a new study shows that knowledge distillation (KD) targeting attention accelerates syntax acquisition . the hypothesis is tested on syntactic benchmarks and perplexity. |
SliceMoE: Routing Embedding Slices Instead of Tokens for Fine-Grained and Balanced Transformer Scaling (2025.emnlp-main)
Copied to clipboard
| Challenge: | Token-level routing assigns an entire semantic spectrum to each expert, creating capacity bottlenecks, load-balancing pathologies, and limited specialisation. |
| Approach: | They propose an architecture that routes contiguous slices of a token’s hidden vector and a lightweight shared router predicts the top-k experts. |
| Outcome: | The proposed architecture achieves 1.7x faster inference than dense baselines, 12–18% lower perplexity than parameter-matched token-MoE, and improved expert balance. |
SABER: Uncovering Vulnerabilities in Safety Alignment via Cross-Layer Residual Connection (2025.emnlp-main)
Copied to clipboard
| Challenge: | Large Language Models (LLMs) with safe-alignment training are vulnerable to jailbreak attacks, causing malicious users to generate harmful outputs. |
| Approach: | They propose a safe-alignment jailbreak method that bypasses the middle-to-late layers of large language models by a residual connection. |
| Outcome: | The proposed method improves by 51% over the best performing baseline GCG on HarmBench test set. |
BOSE: A Systematic Evaluation Method Optimized for Base Models (2025.findings-acl)
Copied to clipboard
| Challenge: | Existing evaluation methods for large language models (LLMs) are inadequate to provide solid conclusions for key experiments such as data ablation and scaling law. |
| Approach: | They propose a method specifically designed to optimize the evaluation of base models by incorporating two innovations: In-Context Light-instruction Prompt and Blank-ppl for multi-choice tasks with candidate options. |
| Outcome: | The proposed method significantly improves stability and consistency of evaluations during pre-training and consistency between base and instruct models. |
Calibration-Aware Policy Optimization for Reasoning LLMs (2026.acl-long)
Copied to clipboard
| Challenge: | Existing approaches to model calibration are limited or sacrifice gains in reasoning accuracy. |
| Approach: | They propose a method that improves calibration by 15% while boosting accuracy by 5% . they propose GRPO-style algorithms that misalign uncertainty-agnostic advantage estimation . |
| Outcome: | The proposed approach improves calibration by 15% while achieving comparable to or better than GRPO on multiple mathematical reasoning benchmarks. |
Diffusion Guided Language Modeling (2024.findings-acl)
Copied to clipboard
| Challenge: | Existing guidance methods for text generation are prone to decoding errors and degrade performance. |
| Approach: | They propose a model that steers an auto-regressive language model to generate text with desired properties. |
| Outcome: | The proposed model outperforms existing guidance methods on a wide range of benchmark data sets. |
Beyond Perplexity: Multi-dimensional Safety Evaluation of LLM Compression (2024.findings-emnlp)
Copied to clipboard
| Challenge: | Prior work on compression prioritizes preserving perplexity, which is analogous to training loss. |
| Approach: | They examine the impact of model compression along four dimensions: degeneration harm, representational harm, dialect bias, and language modeling and downstream task performance. |
| Outcome: | The proposed compression methods can lead to unexpected consequences, the authors show . quantization preserves bias while pruning degrades quickly. |
InstOptima: Evolutionary Multi-objective Instruction Optimization via Large Language Model-based Instruction Operators (2023.findings-emnlp)
Copied to clipboard
| Challenge: | Existing studies focus on automating instruction generation but do not consider other objectives that impact instruction quality. |
| Approach: | They propose an approach that treats instruction generation as an evolutionary multi-objective optimization problem. |
| Outcome: | The proposed approach improves fine-tuning performance and the generation of high-quality instructions. |
SAGE: Sign-Adaptive Gradient for Memory-Efficient LLM Optimization (2026.findings-acl)
Copied to clipboard
| Challenge: | Existing methods to train LLMs consume memory equivalent to twice the model size, resulting in a hybrid design that reverts to AdamW and negates the memory gains. |
| Approach: | They propose a new, memory-efficient O(d) adaptive scale that replaces AdamW in a hybrid structure that combines a Lion-style update direction with a memory-saving adaptive scale. |
| Outcome: | The proposed model outperforms existing methods on LLMs up to 1.3B parameters while significantly reducing optimizer state memory. |
kNN-LM Does Not Improve Open-ended Text Generation (2023.emnlp-main)
Copied to clipboard
| Challenge: | Interpolation-based retrieval-augmented language models (LMs) are a subtype of retrieval augmented language model that computes the probability of the next token by interpolating between the softmax distribution of the original LM and a token distribution formed by retrieving over an external datastore. |
| Approach: | They propose to interpolate the predicted distribution of the next word with a distribution formed from the most relevant retrievals for a given prefix. |
| Outcome: | The proposed methods do not exhibit improvements in open-ended generation quality, as measured by automatic evaluation metrics and human evaluations. |
Towards A Unified View of Sparse Feed-Forward Network in Pretraining Large Language Model (2023.emnlp-main)
Copied to clipboard
| Challenge: | Large and sparse feed-forward layers (S-FFN) have proven effective in scaling up the model size for pretraining large language models. |
| Approach: | They compare S-FFN architectures for language modeling and compare their performance and efficiency . they found a simpler selection method that selects blocks through their mean aggregated hidden states . |
| Outcome: | The proposed model size and selection method achieve lower perplexity in language model pretraining compared to existing MoE architectures. |
KurTail : Kurtosis-based LLM Quantization (2025.findings-emnlp)
Copied to clipboard
| Challenge: | Outliers in quantization of large language models make uniform quantization schemes less effective . a new PTQ scheme that leverages Kurtosis-based rotation mitigates outliers . |
| Approach: | They propose a new post-training quantization scheme that leverages Kurtosis-based rotation to mitigate outliers in the activations of large language models. |
| Outcome: | The proposed method outperforms existing quantization methods with 13.3% boost in MMLU accuracy and 15.5% boost in Wiki perplexity. |
Stepwise Perplexity-Guided Refinement for Efficient Chain-of-Thought Reasoning in Large Language Models (2025.findings-acl)
Copied to clipboard
Yingqian Cui, Pengfei He, Jingying Zeng, Hui Liu, Xianfeng Tang, Zhenwei Dai, Yan Han, Chen Luo, Jing Huang, Zhen Li, Suhang Wang, Yue Xing, Jiliang Tang, Qi He
| Challenge: | Chain-of-Thought (CoT) reasoning has improved the performance of large language models (LLMs) however, the detailed reasoning process in CoT often incurs long generation times and high computational costs due to the inclusion of unnecessary steps. |
| Approach: | They propose a method to identify critical reasoning steps using perplexity as a measure of their importance. |
| Outcome: | The proposed method achieves a better balance between reasoning accuracy and efficiency of CoT. |
GTA: Gated Toxicity Avoidance for LM Performance Preservation (2023.findings-emnlp)
Copied to clipboard
| Challenge: | Existing Controllable Text Generation methods that generate toxic text can negatively impact the performance of the language model. |
| Approach: | They propose a gated Toxicity Avoidance method that can be applied to any CTG method and evaluate its effectiveness. |
| Outcome: | The proposed method reduces toxicity and preserves performance while preserving the language model's generation performance. |
Nirvana: A Specialized Generalist Model With Task-Aware Memory Mechanism (2026.acl-long)
Copied to clipboard
Yuhua Jiang, Shuang Cheng, Yihao Liu, Ermo Hua, Che Jiang, Weigao Sun, Yu Cheng, Feifei Gao, Biqing Qi, Bowen Zhou
| Challenge: | Existing models lack task-guided specialized memory mechanisms . specialized generalist models excel at general language tasks but struggle in specialized domains. |
| Approach: | They propose a specialized generalist model with specialized memory and updater that can optimize for specialized domains. |
| Outcome: | The proposed model matches or surpasses baselines on general benchmarks and achieves lowest perplexity across specialized domains. |
Perplexity-Aware Data Scaling Law: Perplexity Landscapes Predict Performance for Continual Pre-training (2026.acl-long)
Copied to clipboard
| Challenge: | Large language models (LLMs) have impressive capabilities across a wide range of domains, but their generalpurpose pre-training objectives often leave them illsuited for specialized applications such as healthcare. |
| Approach: | They propose a perplexity-aware data scaling law that establishes a predictive relationship between the perplexities of domain-specific data and the test loss. |
| Outcome: | Experiments on medical and general-domain benchmarks show that the proposed scaling law consistently identifies near-optimal training subsets with significantly reduced data consumption. |
ATQ: Activation Transformation forWeight-Activation Quantization of Large Language Models (2024.findings-emnlp)
Copied to clipboard
| Challenge: | Quantization methods are available to solve the problem of high computational and storage costs for Large language models. |
| Approach: | They propose an INT8 weight-activation quantization method that can achieve lossless accuracy. |
| Outcome: | The proposed method can achieve lossless accuracy on OPT and LLaMA families. |
CARVQ: Corrective Adaptor with Group Residual Vector Quantization for LLM Embedding Compression (2025.findings-emnlp)
Copied to clipboard
Dayin Gou, Sanghyun Byun, Nilesh Malpeddi, Gabrielle De Micheli, Prathamesh Vaste, Jacob Song, Woo Seong Chung
| Challenge: | Large Language Models typically rely on a large number of parameters for token embedding, leading to substantial storage requirements and memory footprints. |
| Approach: | They propose a corrective Adaptor with group Residual Vector Quantization that can be used to compress the embedding layer without requiring specialized hardware. |
| Outcome: | The proposed corrective adaptor can achieve lower average bitwidth-per-parameter while maintaining reasonable perplexity and accuracy compared to scalar quantization. |
How to Compute the Probability of a Word (2024.emnlp-main)
Copied to clipboard
| Challenge: | Language models estimate a probability distribution over strings in a natural language . many recent linguistic studies have been incorrectly computing word probabilities . |
| Approach: | They propose to use the correct method to compute word probabilities . they highlight issues when relying on models that use end-of-word tokenisers . |
| Outcome: | Empirically, correcting the widespread bug affects measured outcomes in sentences and lexical optimisation analyses. |
Who is better at math, Jenny or Jingzhen? Uncovering Stereotypes in Large Language Models (2024.emnlp-main)
Copied to clipboard
| Challenge: | Existing research on stereotypes in large language models is limited and focuses on African Ameri- F. |
| Approach: | They propose to use global bias to probe a set of large language models via perplexity to determine how certain stereotypes are represented in the model's internal representations. |
| Outcome: | The proposed model amplifys harmful stereotypes and shows that the demographic groups associated with stereotypes remain consistent across model likelihoods and outputs. |
Semformer: Transformer Language Models with Semantic Planning (2024.emnlp-main)
Copied to clipboard
| Challenge: | Neural language models (LLMs) employ teacher forcing to predict tokens based on preceding ground truth tokens. |
| Approach: | They propose a method for training a Transformer language model that explicitly models the semantic planning of response. |
| Outcome: | The proposed method exhibits near-perfect performance and mitigates shortcut learning. |
Are Data Augmentation Methods in Named Entity Recognition Applicable for Uncertainty Estimation? (2024.emnlp-main)
Copied to clipboard
| Challenge: | Named Entity Recognition (NER) is a key task in NLP to find mentions of named entities and classify them into predefined categories. |
| Approach: | They investigated the impact of data augmentation on confidence calibration and uncertainty estimation in Named Entity Recognition (NER) tasks. |
| Outcome: | The data augmentation improves calibration and uncertainty in cross-genre and cross-lingual setting, especially in-domain setting. |
Beyond Fixed-Length Calibration for Post-Training Compression of LLMs (2025.findings-emnlp)
Copied to clipboard
| Challenge: | a recent study has demonstrated that the sequence length of calibration data plays a crucial role in the effectiveness of post-training compression methods. |
| Approach: | They propose a calibration technique that applies masking along the sequence axis to normalized hidden states. |
| Outcome: | The proposed method improves perplexity and zero-shot downstream tasks performance. |
RiOT: Efficient Prompt Refinement with Residual Optimization Tree (2025.acl-long)
Copied to clipboard
| Challenge: | Existing methods for automatic prompt optimization face two challenges: lack of diversity and semantic drift. |
| Approach: | They propose a framework for automatic prompt optimization that iteratively refines prompts through text gradients and selects the best prompt using perplexity. |
| Outcome: | The proposed framework outperforms existing prompt optimization methods and manual prompting on commonsense, mathematical, logical, temporal, and semantic reasoning benchmarks. |
VRPO: Rethinking Value Modeling for Robust RL under Noisy Supervision in LLM Post-Training (2026.acl-long)
Copied to clipboard
Dingwei Zhu, Shihan Dou, Zhiheng Xi, Senjie Jin, Guoqiang Zhang, Jiazheng Zhang, Junjie Ye, Mingxu Chai, Enyu Zhou, Ming Zhang, Yuhui Wang, Caishuang Huang, Chenhao Huang, Yunke Zhang, Yuran Wang, Tao Gui, Qi Zhang, Xipeng Qiu, Xuanjing Huang
| Challenge: | Reinforcement Learning (RL) in real-world environments often suffers from ambiguous or incomplete supervision. |
| Approach: | They propose a framework that enhances value modeling for robust RL in LLM post-training by integrating auxiliary losses guided by entropy and perplexity from a frozen language model and variational information bottleneck. |
| Outcome: | The proposed framework outperforms baselines on multi-turn dialogue, math reasoning, and science QA with rule-based and model-based rewards. |
Adaptation Odyssey in LLMs: Why Does Additional Pretraining Sometimes Fail to Improve? (2024.emnlp-main)
Copied to clipboard
| Challenge: | In the last decade, the generalization and adaptation abilities of deep learning models were evaluated on fixed training and test distributions. |
| Approach: | They propose to train large language models on unlabeled text corpora and train them online. |
| Outcome: | The proposed model training on a text domain could degrade its perplexity on the test portion of the same domain. |
Massive Supervised Fine-tuning Experiments Reveal How Data, Layer, and Training Factors Shape LLM Alignment Quality (2025.emnlp-main)
Copied to clipboard
| Challenge: | Recent advances in large language models (LLMs) have greatly improved natural language understanding and generation. |
| Approach: | They train a wide range of base models on a variety of datasets including code generation, mathematical reasoning, and general-domain tasks. |
| Outcome: | The results show that training–task synergies persist across all models while others vary substantially, emphasizing the importance of model-specific strategies. |
Mixture of Weight-shared Heterogeneous Group Attention Experts for Dynamic Token-wise KV Optimization (2025.emnlp-main)
Copied to clipboard
| Challenge: | Existing methods for token-level KV optimization and grouping of tokens are inefficient and strain compute and storage resources. |
| Approach: | They propose a mixture-of-expert approach that dynamically optimizes token-wise computation and memory allocation by a token-based expert-choice routing mechanism guided by learned importance scores. |
| Outcome: | The proposed approach retains all tokens while adaptively routing them to specialized experts with varying KV group sizes, balancing granularity and efficiency. |
Scalable Data Ablation Approximations for Language Models through Modular Training and Merging (2024.emnlp-main)
Copied to clipboard
Clara Na, Ian Magnusson, Ananya Harsh Jha, Tom Sherborne, Emma Strubell, Jesse Dodge, Pradeep Dasigi
| Challenge: | Training data compositions for Large Language Models (LLMs) can significantly affect their downstream performance. |
| Approach: | They propose a method which trains individual models on subsets of a training corpus and reuses them across evaluations of combinations of subset. |
| Outcome: | The proposed method improves training efficiency by scaling only linearly with respect to new data. |
Reference-less Analysis of Context Specificity in Translation with Personalised Language Models (2024.lrec-main)
Copied to clipboard
Sebastian Vincent, Rowanne Sumner, Alice Dowek, Charlotte Prescott, Emily Preston, Chris Bayliss, Chris Oakley, Carolina Scarton
| Challenge: | Conventional approaches to NLP tasks build models in a one-size-fits-all fashion disregarding the context of the processed text. |
| Approach: | They build LMs which leverage rich contextual information to reduce perplexity by up to 6.5% compared to a non-contextual model. |
| Outcome: | The proposed models reduce perplexity by up to 6.5% compared to non-contextual models and generalise well to a scenario with no speaker-specific data. |
Reflections & Resonance: Two-Agent Partnership for Advancing LLM-based Story Annotation (2024.lrec-main)
Copied to clipboard
| Challenge: | Existing methods for story annotation require a meticulous and resourceintensive effort, but the advent of advanced computational tools like GPT-4 can streamline the process and mitigate common limitations. |
| Approach: | They propose a multi-agent system that generates tailored prompts for a large language model and provides feedback to refine the initial prompts. |
| Outcome: | The proposed system significantly improves the model's reconstruction accuracy and confidence, demonstrating that dynamic interaction between agents significantly boosts the annotation process's precision and efficiency. |
Inhibitory Attacks on Backdoor-based Fingerprinting for Large Language Models (2026.acl-long)
Copied to clipboard
| Challenge: | Backdoor-based LLM fingerprinting is a promising solution for intellectual property protection . however, the vulnerability of existing LLMs for the ensemble scenario is unexplored . |
| Approach: | They propose two new fingerprinting attack methods to assess the robustness of LLM fingerprinting by token filter attack and sentence verification attack. |
| Outcome: | The proposed methods inhibit the fingerprint response while maintaining ensemble performance. |
MOSAIC: Multiple Observers Spotting AI Content (2025.findings-acl)
Copied to clipboard
| Challenge: | Large Language Models (LLMs) have made it easier for all to produce harmful, toxic, faked or forged content. |
| Approach: | They propose to use large language models to automatically discriminate from human-written texts by comparing their probability distributions over a document to see if they can detect forged or harmful content. |
| Outcome: | The proposed approach harnesses each model’s capabilities, leading to strong detection performance on a variety of domains. |
Scansion-based Lyrics Generation (2024.lrec-main)
Copied to clipboard
| Challenge: | a new method for generating lyrics for Mandarin songs is based on scansion . the number of syllables required is variable due to the number and number of notes . |
| Approach: | They propose a method to generate Mandarin lyrics with a good match between melody and tonal contour. |
| Outcome: | The proposed system outperforms all other systems in lyric-melody fit and uses proxies for quantifying creativity. |
ABLE: Personalized Disability Support with Politeness and Empathy Integration (2024.emnlp-main)
Copied to clipboard
| Challenge: | Adaptive, Bespoke, Listen and Empathetic is a conversational support system for physical disabilities that tracks user personas and provides personalized support according to user person preferences. |
| Approach: | They propose a conversational support system that tracks user personas and integrates politeness and empathy levels into responses to ensure that support interactions are tailored to each user's characteristics and preferences. |
| Outcome: | The proposed system is based on a conversational dataset enriched with user profile annotations and tested on 84 users with physical disabilities. |
Sequence Reducible Holdout Loss for Language Model Pretraining (2024.lrec-main)
Copied to clipboard
| Challenge: | Data selection techniques have shown empirical benefits in reducing the number of gradient steps to train neural models. |
| Approach: | They propose to modify an existing data selection technique to adapt it to the sequence losses typical in language modeling. |
| Outcome: | The proposed technique reduces the number of steps required to train neural models by 4.3% and improves generalization ability on out of domain datasets. |
Attention Speaks Volumes: Localizing and Mitigating Bias in Language Models (2025.acl-long)
Copied to clipboard
| Challenge: | ambiguous comparative prompting frameworks generate biased content when confronted with ambiguously prompts . ambiguity in LLMs can manifest in various forms, including reinforcing stereotypes, displaying gender bias, or demonstrating preferential treatment towards specific demographic groups. |
| Approach: | They propose a technique to localize bias to specific layers of an LLM by analyzing attention scores and then reduce bias by scaling attention in these biased layers. |
| Outcome: | The proposed technique reduces bias by scaling attention in biased layers. |
LLM-Codec: Neural Audio Codec Meets Language Model Objectives (2026.findings-acl)
Copied to clipboard
| Challenge: | Neural audio codecs are optimized for waveform reconstruction rather than autoregressive prediction. |
| Approach: | They propose to augment codec training with language-model-facing objectives while keeping both codec and LLM architectures unchanged. |
| Outcome: | The proposed model improves speech coherence and predictability by preserving the semantic alignment between audio and text representations. |
DRPruning: Efficient Large Language Model Pruning through Distributionally Robust Optimization (2025.acl-long)
Copied to clipboard
| Challenge: | Structured pruning reduces model size but often causes uneven degradation across domains, leading to biased performance. |
| Approach: | They propose a method that dynamically adjusts the data distribution during training to restore balanced performance across heterogeneous and multi-tasking data. |
| Outcome: | Experiments in monolingual and multilingual settings show that the proposed method surpasses similarly sized models in pruning and continued pretraining over perplexity, downstream tasks, and instruction tuning. |
GiLT: Augmenting Transformer Language Models with Dependency Graphs (2026.acl-long)
Copied to clipboard
| Challenge: | Recent work focuses on syntactic tree structures of languages, in particular constituency tree structures. |
| Approach: | They propose a Graph-Infused Layers Transformer Language Model which leverages dependency graphs to augment Transformer language models. |
| Outcome: | The proposed model achieves better syntactic generalization while maintaining competitive perplexity compared with baseline models. |
How Does DPO Reduce Toxicity? A Mechanistic Neuron-Level Analysis (2025.emnlp-main)
Copied to clipboard
| Challenge: | Direct Preference Optimization (DPO) is a popular choice of safety fine-tuning algorithms, but prior explanations of its effects only account for dampened toxic neurons in the MLP layers. |
| Approach: | They analysed four language models and found that toxic neurons only account for 2.5% to 24% of DPO’s effects across models. |
| Outcome: | The proposed method outperforms DPO in reducing toxicity while preserving perplexity, without requiring any weight updates. |
Context-aware Biases for Length Extrapolation (2025.emnlp-main)
Copied to clipboard
| Challenge: | Existing methods for Relative Positional Encoding (RPE) lack the capacity to adapt to different input contexts. |
| Approach: | They propose an additive RPE method that learns token-specific, context-aware biases for each attention head in transformers by dynamically adjusting positional biase based on the input sequence. |
| Outcome: | The proposed method significantly improves the extrapolation performance of existing RPE methods on the fineWeb-Edu-10B and WikiText-103 datasets. |
Emergent Abilities of Large Language Models under Continued Pre-training for Language Adaptation (2025.acl-long)
Copied to clipboard
| Challenge: | Existing large language models are notoriously English-centric, and their performance has been reported to drop significantly in lessresourced languages. |
| Approach: | They propose a language-agnostic benchmark for in-context learning that reveals catastrophic forgetting early on CPT when English is not included. |
| Outcome: | The proposed method does not impact validation perplexity but is critical for emergence of downstream capabilities in the target language. |
Quantized but Deceptive? A Multi-Dimensional Truthfulness Evaluation of Quantized LLMs (2025.emnlp-main)
Copied to clipboard
| Challenge: | Quantization enables efficient deployment of large language models in resource-constrained environments . but impact on truthfulness remains largely unexplored . |
| Approach: | They propose a framework to assess the truthfulness of quantized large language models . they find quantized models retain internally truthful representations but produce false outputs . |
| Outcome: | The framework assesses the truthfulness of quantized models across three dimensions . it finds that quantized model models retain internally truthful representations but are more susceptible to false outputs . |
Auto-Stega: An Agent-Driven System for Lifelong Strategy Evolution in LLM-Based Text Steganography (2026.findings-acl)
Copied to clipboard
| Challenge: | prevailing methods rely on hand-crafted or pre-specified strategies and struggle to balance efficiency, imperceptibility, and security, particularly at high embedding rates. |
| Approach: | They propose an agent-driven self-evolving framework that is the first to realize self-changing steganographic strategies by automatically discovering, composing, and adapting strategies at inference time. |
| Outcome: | The proposed framework achieves 42.2% perplexity and 1.6% anti-steganalysis performance over SOTA methods at high embedding rates. |
From Local to Global: Revisiting Structured Pruning Paradigms for Large Language Models (2026.acl-long)
Copied to clipboard
Ziyan Wang, Enmao Diao, Qi Le, Pu Wang, Minwoo Lee, Shu-ping Yeh, Evgeny Stupachenko, Hao Feng, Li Yang
| Challenge: | Structured pruning is a practical approach to deploying large language models (LLMs) but it fails to capitalize on modest task-specific calibration signals, causing limited downstream gains. |
| Approach: | They propose a method that removes attention heads and MLP channels using loss-based important scores . they use perplexity for language modeling and a margin-based objective for decision-style tasks . |
| Outcome: | The proposed method lowers perplexity and improves accuracy at higher sparsity . it also stabilizes accuracy and mitigates perxity collapse without fine-tuning . |
MPCG: Multi-Round Persona-Conditioned Generation for Modeling the Evolution of Misinformation with LLMs (2025.emnlp-main)
Copied to clipboard
| Challenge: | Misinformation evolves as it spreads, shifting in language, framing, and moral emphasis to adapt to new audiences. |
| Approach: | They propose a multi-round, persona-conditioned framework that simulates how claims are iteratively reinterpreted by agents with distinct ideological perspectives. |
| Outcome: | The proposed framework generates persona-specific claims across multiple rounds . it is based on an uncensored large language model and is scalable to multiple tasks . |
Uncertainty-Aware Routing for Principled Alignment with MoE Dynamics (2026.acl-long)
Copied to clipboard
Yilong Chen, Junyuan Shang, Yuchen Feng, Zhenyu Zhang, Naibin Gu, Ziqi Wang, Tingwen Liu, Shuohuan Wang, Yu Sun, Hua Wu, Haifeng Wang
| Challenge: | Mixture-of-Experts (MoE) is a cornerstone for scaling LLMs, yet its training dynamics remain poorly understood, often leading to sub-optimal specialization. |
| Approach: | They propose to use Helmholtz Free Energy and Router Entropy to study the MoE lifecycle and identify a universal Three-Stage Phase Transition . |
| Outcome: | The proposed model reduces perplexity and improves expert distinctiveness, offering a principled path toward thermodynamically aligned computation. |
HTMuon: Improving Muon via Heavy-Tailed Spectral Correction (2026.findings-acl)
Copied to clipboard
| Challenge: | Muon’s orthogonalized update rule suppresses the emergence of heavy-tailed weight spectra and over-emphasizes the training along noise-dominated directions. |
| Approach: | They propose to preserve Muon's ability to capture parameter interdependencies while producing heavier-tailed updates and inducing heavier-tail weight spectra. |
| Outcome: | The proposed algorithm suppresses the emergence of heavy-tailed weight spectra and over-emphasizes training along noise-dominated directions. |
Resonant Context Anchoring: Decoupling Attention Routing and Signal Gain at Inference Time (2026.findings-acl)
Copied to clipboard
Mingkuan Zhao, Yide Gao, Wentao Hu, Suquan Chen, Tianchen Huang, Zhenhua An, Zetao Chang, Xiayu Sun, Yuheng Min
| Challenge: | Existing mitigation strategies rely on suppressing specific neuron activations or employing computationally expensive contrastive decoding mechanisms, which often result in increased perplexity or significantly elevated inference latency. |
| Approach: | They propose a lightweight inference-time intervention method grounded in the perspective of residual stream signal dynamics to resolve the signal attenuation of external evidence during its propagation through deep networks. |
| Outcome: | The proposed method improves contextual faithfulness across multiple factual consistency and strong knowledge-conflict tasks while maintaining the model’s general language understanding capabilities. |
Red-Bandit: Test-Time Adaptation for LLM Red-Teaming via Bandit-Guided LoRA Experts (2026.acl-long)
Copied to clipboard
| Challenge: | Existing approaches to audit Large Language Models (LLMs) lack mechanisms to efficiently adapt to model-specific vulnerabilities at inference. |
| Approach: | They propose a red-teaming framework that adapts online to identify and exploit model failure modes under distinct attack styles. |
| Outcome: | The proposed framework outperforms state-of-the-art methods on AdvBench and HarmBench, while generating more human-readable adversarial prompts (lower perplexity). |
SharVeT: Similarity-aware Parameter Sharing with Vector-based Tuning for Efficient LLM Compression (2026.acl-long)
Copied to clipboard
| Challenge: | Existing methods for parameter sharing rely on naive grouping and fail to correct sharing-induced discrepancies. |
| Approach: | They propose a parameter sharing framework that performs similarity-based grouping to ensure accurate sharing and allocates parameters adaptively to preserve diversity within each group. |
| Outcome: | The proposed framework outperforms existing methods, achieving 32.1% lower perplexity and 23.3% higher few-shot reasoning accuracy. |