Papers with finetuning
BitFit: Simple Parameter-efficient Fine-tuning for Transformer-based Masked Language-models (2022.acl-short)
Copied to clipboard
| Challenge: | Large pre-trained models are expensive to train and deploy . large data sets make finetuning expensive to deploy - a new paradigm . |
| Approach: | They propose a sparse-finetuning method where only bias-terms are modified . they show that applying BitFit on pre-trained BERT models is competitive . |
| Outcome: | The proposed method is competitive with sparse-finetuning methods on small-to-medium training data. |
Quantized Side Tuning: Fast and Memory-Efficient Tuning of Quantized Large Language Models (2024.acl-long)
Copied to clipboard
Zhengxin Zhang, Dan Zhao, Xupeng Miao, Gabriele Oliaro, Zhihao Zhang, Qing Li, Yong Jiang, Zhihao Jia
| Challenge: | Existing methods to finetun large language models (LLMs) only update a small number of trainable parameters, or attempt to reduce the memory footprint during the training phase of the finetune process. |
| Approach: | They propose quantized side tuing (QST) which quantizes an LLM’s model weights into 4-bit to reduce the memory footprint of the original weights. |
| Outcome: | The proposed method reduces the memory footprint of the model weights, optimizer states, and intermediate activations while reducing the memory requirements. |
On Training Instance Selection for Few-Shot Neural Text Generation (2021.acl-short)
Copied to clipboard
| Challenge: | Pretraining large neural networks with a language modeling objective has led to dramatic improvements in text generation. |
| Approach: | They propose a selection strategy to select few-shot training instances based on unlabeled data to identify the most worthwhile data points that should be annotated under some budget of labeling cost. |
| Outcome: | The proposed strategy outperforms random sampling on three text generation tasks. |
P-Tuning: Prompt Tuning Can Be Comparable to Fine-tuning Across Scales and Tasks (2022.acl-short)
Copied to clipboard
| Challenge: | Existing methods of prompt tuning cannot handle hard sequence labeling tasks. |
| Approach: | They propose to optimize prompt tuning to tune continuous prompts with a frozen language model. |
| Outcome: | The proposed method matches finetuning with prompt tuning while having only 0.1%-3% tuned parameters. |
LMFlow: An Extensible Toolkit for Finetuning and Inference of Large Foundation Models (2024.naacl-demo)
Copied to clipboard
| Challenge: | Foundation models have demonstrated a great ability to achieve general human-level intelligence far beyond traditional approaches. |
| Approach: | They propose a toolkit to simplify the finetuning of general foundation models. |
| Outcome: | The proposed toolkit simplifies the domain- and task-aware finetuning of general foundation models with limited computing resources. |
Knowledge Distillation Transfer Sets and their Impact on Downstream NLU Tasks (2022.emnlp-industry)
Copied to clipboard
| Challenge: | Domain Classification (DC) and Intent Classification/Named Entity Recognition (ICNER) are the most common methods for reducing teacher-student knowledge into manageable sizes for low-latency downstream applications. |
| Approach: | They investigate whether distillation from a generic LM benefits downstream tasks . a domain classification and a task-specific data set are used to fine tune the model . |
| Outcome: | The proposed model improves across tasks and test sets when only task-specific data is used. |
LMdiff: A Visual Diff Tool to Compare Language Models (2021.emnlp-demo)
Copied to clipboard
| Challenge: | LMdiff visually compares probability distributions of two different language models . notably absent from the range of available tools are those that aim to compare distributions produced by different models. |
| Approach: | They propose a tool that visually compares probability distributions of two different language models that differ through finetuning, distillation, or simply training with different parameter sizes. |
| Outcome: | The proposed tool allows the generation of hypotheses about model behavior by investigating text instances token by token and further assists in choosing interesting text instances from large corpora. |
CoTAR: Chain-of-Thought Attribution Reasoning with Multi-level Granularity (2024.findings-emnlp)
Copied to clipboard
| Challenge: | State-of-the-art QA systems employ Large Language Models (LLMs) however, these models tend to hallucinate information in their responses. |
| Approach: | They propose an attribution-oriented Chain-of-Thought reasoning method to enhance attributions. |
| Outcome: | The proposed method outperforms existing models on context enhanced question-answering datasets and shows that it can be used to improve accuracy. |
Infusing Finetuning with Semantic Dependencies (2021.tacl-1)
Copied to clipboard
| Challenge: | Several diagnostics help to localize the benefits of our approach. |
| Approach: | They apply convolutional graph encoders to integrate semantic parses into task-specific finetuning. |
| Outcome: | The proposed approach yields benefits to natural language understanding (NLU) tasks in the GLUE benchmark. |
Learning Which Features Matter: RoBERTa Acquires a Preference for Linguistic Generalizations (Eventually) (2020.emnlp-main)
Copied to clipboard
| Challenge: | Pretraining on self-supervised linguistic tasks is effective for learning features helpful for language understanding, but it requires more data to learn to prefer linguistic generalizations over surface ones. |
| Approach: | They propose a set of 20 ambiguous binary classification tasks to test whether a pretrained model prefers linguistic or surface generalizations. |
| Outcome: | The proposed model can learn to represent linguistic features with little pretraining data, but requires far more data to learn to prefer linguistic generalizations over surface ones. |
Unleashing the Unseen: Harnessing Benign Datasets for Jailbreaking Large Language Models (2026.findings-eacl)
Copied to clipboard
| Challenge: | Despite significant efforts in safety alignment, large language models (LLMs) such as GPT-4 and LLaMA 3 remain vulnerable to jailbreak attacks that can induce harmful behaviors. |
| Approach: | They propose a feature extraction method to extract sample-agnostic features from benign datasets in the form of adversarial suffixes and propose 'suffix maybe features' they show that adversarials generated from jailbreak attacks may contain meaningful features, i.e. appending the same suffix to different prompts results in responses exhibiting specific characteristics. |
| Outcome: | The proposed method extracts sample-agnostic features from benign datasets and shows that they may contain meaningful features. |
Domain-agnostic Question-Answering with Adversarial Training (D19-58)
Copied to clipboard
| Challenge: | Adapting models to new domain without finetuning is a challenging problem in deep learning. |
| Approach: | They propose an adversarial training framework for domain generalization in Question Answering task using a conventional QA model and a discriminator. |
| Outcome: | The proposed model outperforms the baseline model on Question Answering (QA) task. |
Enhancing Transformers with Gradient Boosted Decision Trees for NLI Fine-Tuning (2021.findings-acl)
Copied to clipboard
| Challenge: | Recent advances in transfer learning have brought significant improvements to many natural language processing tasks. |
| Approach: | They propose a method of fitting a GBDT head on the features computed during finetuning to increase performance without additional computation by the neural network. |
| Outcome: | The proposed method improves on several NLI datasets using a strong baseline model (RoBERTa-large) with MNLI pretraining. |
Exploring Semantics in Pretrained Language Model Attention (2024.starsem-1)
Copied to clipboard
| Challenge: | Abstract Meaning Representations (AMRs) encode the semantics of sentences in the form of graphs. |
| Approach: | They propose to use attention heads of two LMs to detect semantic relations encoded in AMRs. |
| Outcome: | The proposed models detect semantic relations without fine tuning, using both unsupervised and supervised learning techniques. |
Prompt Tuning for Unified Multimodal Pretrained Models (2023.findings-acl)
Copied to clipboard
| Challenge: | Prompt tuning has demonstrated success in natural language pretraining and even vision pretraining. |
| Approach: | They propose to apply prompt tuning to a unified sequence-to-sequence pretrained model by adding a sequence of learnable embeddings to each layer and finetuning the pretrained models on downstream tasks. |
| Outcome: | The proposed method outperforms other parameter-efficient tuning methods on multimodal models and is robust against adversarial attacks. |
GradSafe: Detecting Jailbreak Prompts for LLMs via Safety-Critical Gradient Analysis (2024.acl-long)
Copied to clipboard
| Challenge: | Existing methods for detecting jailbreak prompts are primarily online moderation APIs or finetuned LLMs. |
| Approach: | They propose a method which scrutinizes the gradients of safety-critical parameters in large LLMs to detect jailbreak prompts. |
| Outcome: | The proposed method outperforms Llama Guard in detecting jailbreak prompts despite extensive finetuning with a large dataset. |
MEAL: Stable and Active Learning for Few-Shot Prompting (2023.findings-emnlp)
Copied to clipboard
| Challenge: | Existing methods for few-shot classification have high variance across different sets of few shots and finetuning runs. |
| Approach: | They propose novel ensembling methods that significantly reduce run variability and introduce a new active learning criterion for *data selection*. |
| Outcome: | The proposed method significantly reduces run variability and improves performance on five tasks. |
PACIT: Unlocking the Power of Examples for Better In-Context Instruction Tuning (2024.findings-acl)
Copied to clipboard
| Challenge: | In-context instruction tuning is a method that fine tunes with supervised instruction data to enhance the instruction following ability of large language models. |
| Approach: | They propose an in-context instruction tuning method that incorporates positive and negative examples into the prompt for better performance. |
| Outcome: | The proposed method outperforms baseline instruction tuning methods on in-domain and out-domain tasks up to 9.16 and 3.14 average ROUGE-L scores. |
Self-Supervised Meta-Learning for Few-Shot Natural Language Classification Tasks (2020.emnlp-main)
Copied to clipboard
| Challenge: | Existing methods for supervised meta-learning require many training tasks to generalize . cloze-style objectives can be used to generate a large, rich, meta-training task distribution from unlabeled text. |
| Approach: | They propose a self-supervised approach to generate a large, rich, meta-learning task distribution from unlabeled text. |
| Outcome: | The proposed approach generates a large, rich, meta-learning task distribution from unlabeled text. |
Prompterator: Iterate Efficiently towards More Effective Prompts (2023.emnlp-demo)
Copied to clipboard
| Challenge: | Large Language Models (LLMs) use a process known as prompting to solve arbitrary language tasks. prompting is a non-trivial task that requires experimentation in order to arrive at a prompt that solves a specific task. |
| Approach: | They propose a tool that helps users iterate over different potential prompts and choose the best performing one based on human feedback. |
| Outcome: | The proposed tool is open source and easily extensible. |
Building Multilingual Machine Translation Systems That Serve Arbitrary XY Translations (2022.naacl-main)
Copied to clipboard
| Challenge: | Multilingual Neural Machine Translation (MNMT) systems are often limited to many-to-one directions and suffer from poor performance in one-to one directions. |
| Approach: | They propose to build multilingual machine translation systems that serve arbitrary X-Y directions while leveraging multilinguality with a two-stage training strategy of pretraining and finetuning. |
| Outcome: | The proposed system outperforms baseline bilingual models and pivot translation models in most directions without the need for architecture change or extra data collection. |
LLMs Know More About Numbers than They Can Say (2026.eacl-short)
Copied to clipboard
| Challenge: | Large language models (LLMs) are increasingly used in mathematical, scientific, financial and engineering domains. |
| Approach: | They probe the hidden states of several smaller open-source LLMs to find out how big they are . |
| Outcome: | The proposed model improves verbalized accuracy by 3.22% over base models. |
Efficient Low-Resource Language Models Using Tokenizer Transfer (2026.eacl-srw)
Copied to clipboard
| Challenge: | Tokenizer transfer allows training a model for low-resource languages without full retraining . a study of pre-trained tokenizers shows that they are more efficient than traditional training methods. |
| Approach: | They evaluate tokenizer transfer on models trained on language-specific corpora, Orthogonal Mapping Pursuit and Fast Vocabulary Transfer. |
| Outcome: | The proposed model adapts to a pre-trained model without full retraining and improves cross-lingual applicability. |
Probing Critical Learning Dynamics of PLMs for Hate Speech Detection (2024.findings-eacl)
Copied to clipboard
| Challenge: | Existing studies on pretrained language models (PLMs) for hate speech detection have not investigated how their performance is affected by pretraining and finetuning. |
| Approach: | They propose to compare pretrained language models, evaluate their seed robustness, finetuning settings, and the impact of pretraining data collection time. |
| Outcome: | The proposed models show that they are more robust than other models and that they have a better chance of performing better than domain-specific models. |
How to Dissect a Muppet: The Structure of Transformer Embedding Spaces (2022.tacl-1)
Copied to clipboard
| Challenge: | Pretrained embeddings based on the Transformer architecture have taken the NLP community by storm . a novel decomposition of Transformer output embeddables is demonstrated . |
| Approach: | They propose to decompose Transformer output embeddings into a sum of vector factors . they show multi-head attentions and feed-forwards are not equally useful in downstream applications . |
| Outcome: | The proposed method outperforms recurrent architectures on a wide variety of tasks. |
ALERT: Adapt Language Models to Reasoning Tasks (2023.acl-long)
Copied to clipboard
Ping Yu, Tianlu Wang, Olga Golovneva, Badr AlKhamissi, Siddharth Verma, Zhijing Jin, Gargi Ghosh, Mona Diab, Asli Celikyilmaz
| Challenge: | Large language models have shown increasing in-context learning capabilities with scaling up the model and data sizes. |
| Approach: | They propose a benchmark and suite of analyses to evaluate reasoning skills of large language models. |
| Outcome: | The proposed model compares pre-trained and fine-tuned models on tasks that require reasoning skills to solve. |
LNN-EL: A Neuro-Symbolic Approach to Short-text Entity Linking (2021.acl-long)
Copied to clipboard
Hang Jiang, Sairam Gurajada, Qiuhao Lu, Sumit Neelam, Lucian Popa, Prithviraj Sen, Yunyao Li, Alexander Gray
| Challenge: | Existing work deals with EL in the context of longer text, such as a sentence. |
| Approach: | They propose a neuro-symbolic approach that uses interpretable rules based on first-order logic to achieve better performance with black-box neural approaches. |
| Outcome: | The proposed approach achieves better performance than heuristics-based approaches on short-text EL . it can easily blend existing rule templates with multiple types of features, and even with scores resulting from previous EL methods. |
Multilingual Speech Translation from Efficient Finetuning of Pretrained Models (2021.acl-long)
Copied to clipboard
Xian Li, Changhan Wang, Yun Tang, Chau Tran, Yuqing Tang, Juan Pino, Alexei Baevski, Alexis Conneau, Michael Auli
| Challenge: | Recent advances in text pretraining and finetuning have improved multitasking applications significantly. |
| Approach: | They propose a minimalistic LNA finetuning approach to build multilingual speech-to-text translation using a pretrained speech encoder and text decoder. |
| Outcome: | The proposed approach surpasses the cascaded ST benchmark for 36 translation directions on the large-scale multilingual ST benchmark CoVoST 2. |
Language-Independent Representations Improve Zero-Shot Summarization (2024.naacl-short)
Copied to clipboard
| Challenge: | Pretrained models can be fine tuned on downstream generation tasks, but they can fail in zero-shot conditions. |
| Approach: | They propose query-key finetuning to decouple task-specific knowledge from pretrained models . they propose a variant that more directly enforces language-agnostic representations . |
| Outcome: | The proposed model decouples task-specific knowledge from pretrained language generation abilities. |
On Curriculum Learning for Commonsense Reasoning (2022.naacl-main)
Copied to clipboard
| Challenge: | Recent research suggests that data order can have a significant impact on the performance of finetuned models for natural language understanding. |
| Approach: | They use paced curriculum learning to rank data and sample training mini-batches with increasing levels of difficulty during finetuning. |
| Outcome: | The proposed model improves performance for socialIQA, CosmosQA, CODAH, HellaSwag, WinoGrande in both tuning settings. |
Meta-Learning the Difference: Preparing Large Language Models for Efficient Adaptation (2022.tacl-1)
Copied to clipboard
| Challenge: | Large pretrained language models are often domain- or task-adapted via finetuning or prompting. |
| Approach: | They propose to use domain-adaptive pretraining to prepare large pretrained language models for domain- or task-adaptation by learning to learn the difference between general and adapted PLMs. |
| Outcome: | Experiments on few-shot dialogue completion, low-resource abstractive summarization, and multi-domain language modeling show improvements in adaptation time and performance over finetuning or preparation via domain-adaptive pretraining. |
StarFlow: Generating Structured Workflow Outputs From Sketch Images (2026.eacl-long)
Copied to clipboard
Patrice Bechard, Chao Wang, Amirhossein Abaskohi, Juan A. Rodriguez, Christopher Pal, David Vazquez, Spandana Gella, Sai Rajeswar, Perouz Taslakian
| Challenge: | Despite being widely used, building workflows can be complex, often requiring manual configuration through low-code platforms or visual programming tools. |
| Approach: | They propose a framework for generating structured workflow outputs from sketches using vision-language models to automate the process. |
| Outcome: | The proposed framework outperforms large vision-language models in the task of generating structured workflow outputs from sketches and diagrams. |
NoisyTune: A Little Noise Can Help You Finetune Pretrained Language Models Better (2022.acl-short)
Copied to clipboard
| Challenge: | Existing methods for finetuning pretrained language models (PLMs) have risks in overfitting the pretraining tasks and data, which may lead to suboptimal performance. |
| Approach: | They propose a method which adds noise to parameters of PLMs before fine-tuning. |
| Outcome: | The proposed method can be used on GLUE English and XTREME multilingual benchmarks. |
AmazonQAC: A Large-Scale, Naturalistic Query Autocomplete Dataset (2024.emnlp-industry)
Copied to clipboard
| Challenge: | Existing systems that provide a graphical representation of QAC are limited in their ability to provide real-time data. |
| Approach: | They introduce a new QAC dataset sourced from Amazon Search logs . they assess Prefix Trees, semantic retrieval, and Large Language Models with and without finetuning . |
| Outcome: | The proposed system can predict search terms based on user-typed prefixes . the proposed system achieves only half of what is theoretically possible on the test data . |
Diverse Demonstrations Improve In-context Compositional Generalization (2023.acl-long)
Copied to clipboard
| Challenge: | In-context learning has shown great success in i.i.d semantic parsing splits . however, in compositional generalization, selecting similar demonstrations is insufficient . |
| Approach: | They propose a method to select diverse demonstrations that collectively cover all the structures required in the output program and encourage the model to generalize to new structures from these demonstrations. |
| Outcome: | The proposed method improves performance across three compositional generalization datasets and finetuning. |
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. |
Analyzing Large Language Models’ Capability in Location Prediction (2024.lrec-main)
Copied to clipboard
| Challenge: | Large language models (LLMs) are underutilized in the field of location prediction due to the sparsity of geotagged tweets. |
| Approach: | They present experimental results with four large language models in various instruction finetuning and exemplar settings and analyze whether taking into account the context is beneficial. |
| Outcome: | The proposed model is able to predict location in a variety of settings, including fine tuning and exemplar settings, and it is compared with the best model in the literature. |
Planning with Learned Entity Prompts for Abstractive Summarization (2021.tacl-1)
Copied to clipboard
| Challenge: | a simple but flexible mechanism is used to ground the generation of abstractive summaries. |
| Approach: | They propose a mechanism to learn an intermediate plan to ground the generation of abstractive summaries. |
| Outcome: | The proposed model outperforms state-of-the-art methods for faithfulness on CNN and BillSum. |
SharPT: Shared Latent Space Prompt Tuning (2023.findings-eacl)
Copied to clipboard
| Challenge: | Prompt tuning is an efficient method for adapting large language models, but it is difficult and expensive to identify the source task that provides optimal prompts. |
| Approach: | They propose to learn a shared latent space which captures a set of basis skills from a mixture of source tasks and then transfer them to target tasks. |
| Outcome: | The proposed method outperforms previous methods on NLI, sentence completion, QA, conference resolution, word sense disambiguation and on various model scales. |
Translate to Disambiguate: Zero-shot Multilingual Word Sense Disambiguation with Pretrained Language Models (2024.eacl-long)
Copied to clipboard
| Challenge: | Pretrained language models learn cross-lingual knowledge and perform well on diverse tasks when finetuned. |
| Approach: | They propose a zero-shot prompting approach that captures cross-lingual word sense with a contextual prompt. |
| Outcome: | The proposed approach outperforms baselines on recall in many evaluation languages without additional training or finetuning. |
Toward Informal Language Processing: Knowledge of Slang in Large Language Models (2024.naacl-long)
Copied to clipboard
| Challenge: | Recent advances in large language models (LLMs) have offered a strong potential for natural language systems to process informal language. |
| Approach: | They propose to use movie subtitles to evaluate slang in large language models . they find that smaller LLMs finetuned on the dataset achieve comparable performance . |
| Outcome: | The proposed dataset can be used to evaluate LLMs on slang detection and identification of regional and historical sources for interpretive insights. |
Knowledge is a Region in Weight Space for Fine-tuned Language Models (2023.findings-emnlp)
Copied to clipboard
| Challenge: | Specifically, language models that have been finetuned on the same dataset form a tight cluster in the weight space, while models finetuning on different datasets from the same underlying task form s looser clusters. |
| Approach: | They analyze the weight space and the underlying loss landscape of different models to determine how they are interconnected and how they perform. |
| Outcome: | The proposed method improves accuracy on 11 out of 12 datasets by starting from the center of the region, if not more, than using the pretrained model. |
Analyzing the Forgetting Problem in Pretrain-Finetuning of Open-domain Dialogue Response Models (2021.eacl-main)
Copied to clipboard
| Challenge: | a large-scale unsupervised pretraining has been shown to greatly boost the performance of natural language processing models. |
| Approach: | They propose an intuitive finetuning strategy to regularize the finetune process . they propose a mix-review strategy to alleviate the forgetting problem . |
| Outcome: | The proposed strategy regularizes the finetuning process, and the forgetting problem is alleviated . the proposed strategy also improves the performance of the resulting model . |
Grounded Compositional Outputs for Adaptive Language Modeling (2020.emnlp-main)
Copied to clipboard
| Challenge: | Language models are a key component of natural language processing, but their size is a problem because they are typically trained with a closed output vocabulary derived from the training data. |
| Approach: | They propose a fully compositional output embedding layer for language models that is grounded in semantically related words and free-text definitions. |
| Outcome: | The proposed model outperforms state-of-the-art methods and adaptation approaches on cross-domain modeling and cross-learning tasks. |
Efficient Data Generation for Source-grounded Information-seeking Dialogs: A Use Case for Meeting Transcripts (2024.findings-emnlp)
Copied to clipboard
| Challenge: | Existing methods for automating data generation with Large Language Models (LLMs) are difficult, and we propose a semi-automatic approach to generate dialogs with attributions. |
| Approach: | They propose a semi-automatic approach to generate dialog queries and responses with Large Language Models followed by human verification and identification of attribution spans. |
| Outcome: | The proposed approach improves the quality of the response generation and attribution quality of MISeD datasets while reducing time and effort. |
Learning Label Hierarchy with Supervised Contrastive Learning (2024.findings-eacl)
Copied to clipboard
| Challenge: | Existing approaches to supervised contrastive learning treat each class as independent and therefore consider all classes to be equally important. |
| Approach: | They propose a family of Label-Aware SCL methods that incorporate hierarchical information to SCL by leveraging similarities between classes. |
| Outcome: | The proposed method outperforms baseline supervised approaches on three datasets. |
ASR-EC Benchmark: Evaluating Large Language Models on Chinese ASR Error Correction (2025.emnlp-industry)
Copied to clipboard
| Challenge: | Automatic Speech Recognition (ASR) systems have a substantial number of erroneous recognition due to environmental noise, ambiguity, etc. |
| Approach: | They use a benchmark dataset to analyze ASR errors in the Chinese language . they then apply large language models to correct ASR error correction . |
| Outcome: | The proposed method is based on a dataset of ASR errors in the Chinese language . it shows prompting is not effective for ASR error correction . |
Tokenization Matters: Navigating Data-Scarce Tokenization for Gender Inclusive Language Technologies (2024.findings-naacl)
Copied to clipboard
Anaelia Ovalle, Ninareh Mehrabi, Palash Goyal, Jwala Dhamala, Kai-Wei Chang, Richard Zemel, Aram Galstyan, Yuval Pinter, Rahul Gupta
| Challenge: | a recent study documented the harmful limitations of gender binary-centric large language models . data scarcity is a known culprit, but the precise mechanisms through which scarcity affects this behavior remain underexplored. |
| Approach: | They propose to use BPE tokenization to enforce consistent tokenization across gendered pronouns to improve neopronoun proficiency. |
| Outcome: | The proposed methods outperform finetuning with standard BPE, and improve neopronoun proficiency. |
Unsupervised Pronoun Resolution via Masked Noun-Phrase Prediction (2021.acl-short)
Copied to clipboard
| Challenge: | Masked Noun-Phrase Prediction (MNPP) is a pre-training strategy to tackle pronoun resolution in a fully unsupervised setting. |
| Approach: | They propose a pre-training strategy to tackle pronoun resolution in an unsupervised setting by fine-tuning a large pre-trained model on a human-labeled dataset and then transferring to a smaller dataset such as Winograd Schema Challenge (WSC). |
| Outcome: | The proposed method outperforms all previous unsupervised methods on all datasets by large margins. |
Conditional Language Policy: A General Framework For Steerable Multi-Objective Finetuning (2024.findings-emnlp)
Copied to clipboard
Kaiwen Wang, Rahul Kidambi, Ryan Sullivan, Alekh Agarwal, Christoph Dann, Andrea Michi, Marco Gelmi, Yunxuan Li, Raghav Gupta, Kumar Dubey, Alexandre Rame, Johan Ferret, Geoffrey Cideron, Le Hou, Hongkun Yu, Amr Ahmed, Aranyak Mehta, Leonard Hussenot, Olivier Bachem, Edouard Leurent
| Challenge: | Existing approaches for multi-objective Reinforcement Learning (RL) are difficult due to plurality of preferences and applications. |
| Approach: | They propose a framework for finetuning language models on multiple objectives using conditional language policy. |
| Outcome: | The proposed framework outperforms and Pareto-dominates existing approaches for multi-objective Reinforcement Learning (RL) it does not require training or maintaining multiple models to achieve different trade-offs between the objectives. |
KD-VLP: Improving End-to-End Vision-and-Language Pretraining with Object Knowledge Distillation (2022.findings-naacl)
Copied to clipboard
| Challenge: | Existing vision-and-language pretraining approaches rely on external object detectors to encode images in a multi-modal transformer framework. |
| Approach: | They propose an object-aware end-to-end VLP framework which feeds image grid features from CNNs into the Transformer and learns the multi-modal representations jointly. |
| Outcome: | The proposed framework achieves competitive or superior performances on vision-language tasks. |
“Chère maison” or “maison chère”? Transformer-based prediction of adjective placement in French (2023.findings-eacl)
Copied to clipboard
| Challenge: | Prior work has shown that transformer-based language models are insensitive to permutated word order, but this is not the case with pretraining. |
| Approach: | They conduct experiments to assess whether transformer-based language models are able to learn the adjective position in noun phrases in French. |
| Outcome: | The proposed model is weaker with complex structures and fixed expressions, but favors context and global syntactic roles. |
MolXPT: Wrapping Molecules with Text for Generative Pre-training (2023.acl-short)
Copied to clipboard
| Challenge: | Experimental results show that Generative pre-trained Transformers (GPT) have great success in natural language processing. |
| Approach: | They propose a unified language model of text and molecules pre-trained on SMILES wrapped by text. |
| Outcome: | The proposed model outperforms strong baselines of molecular property prediction on MoleculeNet and performs comparably to the best model in text-molecule translation while using less than half of its parameters. |
Hierarchical and Dynamic Prompt Compression for Efficient Zero-shot API Usage (2024.findings-eacl)
Copied to clipboard
| Challenge: | Long prompts present a significant challenge for LLM-based systems that learn to use unseen APIs directly in-context from their documentation. |
| Approach: | They propose a method that learns to compress a prompt into a few "gist token" activations during finetuning but is ineffective in compressing API documentation. |
| Outcome: | The proposed model keeps the simplicity, efficiency, and large compression factor (20x on SGD) of the gist token approach while achieving significantly better accuracy. |
Cross-Lingual Transfer of Debiasing and Detoxification in Multilingual LLMs: An Extensive Investigation (2025.findings-acl)
Copied to clipboard
| Challenge: | Prior work has shown that finetuning on specialized datasets can mitigate this behavior, and doing so in English can transfer to other languages. |
| Approach: | They propose to fine tune generative large language models to provide safe responses to harmful user input and to use direct preference optimization to mitigate toxicity. |
| Outcome: | The proposed models show that finetuning on specialized datasets reduces biases but also produces fluent and diverse text in non-English languages. |
Is This LLM Library Learning? Evaluation Must Account For Compute and Behaviour (2026.eacl-long)
Copied to clipboard
| Challenge: | Recent advances in the coding, reasoning, and tool-use ability of LLMs have raised the possibility of library learning with LLM. |
| Approach: | They propose to use reusable and composable functions and tools to create reusable, composesable code and tools that can be reused by modifying relevant examples. |
| Outcome: | The proposed system fails to consistently outperform the baseline model and does not correct for the difference in computational cost. |
Revisiting Parameter-Efficient Tuning: Are We Really There Yet? (2022.emnlp-main)
Copied to clipboard
| Challenge: | Pretrained language models (PLMs) are used as backbones to be combined with additional parameters and finetuned on downstream tasks in an end-to-end manner. |
| Approach: | They propose to use a fraction of parameters to tune pretrained language models (PLMs) this is the first comprehensive investigation into the training and evaluation of PETuning methods. |
| Outcome: | The proposed methods have been validated and tested with a rigorous evaluation protocol and have shown that they are unstable and inconsistent. |
LLM The Genius Paradox: A Linguistic and Math Expert’s Struggle with Simple Word-based Counting Problems (2025.naacl-long)
Copied to clipboard
| Challenge: | Existing conjectures about the reason for deficiency of LLMs in simple word-based counting problems are invalid. |
| Approach: | They propose to evaluate model transferability from specialized LLMs to simple counting tasks by comparing their results to popular conjectures . |
| Outcome: | The proposed model evaluations show that engaging reasoning is the most robust and efficient way to help LLMs better perceive tasks with more accurate responses. |
Compensate Quantization Errors: Make Weights Hierarchical to Compensate Each Other (2024.findings-naacl)
Copied to clipboard
| Challenge: | Emergent Large Language Models (LLMs) use extraordinary performance and powerful deduction capacity to discern from traditional language models. |
| Approach: | They propose a method that uses weights to compensate quantization error and learnable singular value incremental (LSI) LSI is a technique that helps weights compensate each other conditioned on activation. |
| Outcome: | The proposed method achieves state-of-the-art performance in diverse quantization settings, no matter in weight-only, weight-activation or extremely low bit scenarios. |
Masking as an Efficient Alternative to Finetuning for Pretrained Language Models (2020.emnlp-main)
Copied to clipboard
| Challenge: | Extensive evaluations of masking BERT, RoBERTa, and DistilBERT on eleven diverse NLP tasks show that our binary masked language models encode information necessary for solving downstream tasks. |
| Approach: | They propose an efficient method of utilizing pretrained language models where selective binary masks are learned instead of finetuning. |
| Outcome: | Extensive evaluations of masking BERT, RoBERTa, and DistilBERT on eleven diverse NLP tasks show that the proposed method yields comparable performance to finetuning, but has a much smaller memory footprint when multiple tasks need to be solved. |
A Pretrainer’s Guide to Training Data: Measuring the Effects of Data Age, Domain Coverage, Quality, & Toxicity (2024.naacl-long)
Copied to clipboard
Shayne Longpre, Gregory Yauney, Emily Reif, Katherine Lee, Adam Roberts, Barret Zoph, Denny Zhou, Jason Wei, Kevin Robinson, David Mimno, Daphne Ippolito
| Challenge: | a large number of pretraining data design practices are under-documented, authors say . authors: strong performance of modern language models depends on selfsupervised pretraining . |
| Approach: | They propose to pretrain models on data curated at different collection times . they find temporal shift between evaluation data and pretraining data leads to performance degradation . |
| Outcome: | The results validate, quantify, and expose many undocumented intuitions about text pretraining . authors say this practice has outperformed other models in the field . |
Crossing Linguistic Horizons: Finetuning and Comprehensive Evaluation of Vietnamese Large Language Models (2024.findings-naacl)
Copied to clipboard
| Challenge: | Existing open-source LLMs exhibit limited effectiveness in processing Vietnamese . lack of systematic benchmark datasets and metrics tailored for Vietnamese LLM evaluation exacerbates these issues. |
| Approach: | They propose to fine tune LLMs specifically for Vietnamese and develop a framework for evaluation . they find that larger models introduce more biases and uncalibrated outputs . |
| Outcome: | The proposed framework finetunes LLMs specifically for Vietnamese and provides a framework for evaluation . |
Attending via both Fine-tuning and Compressing (2021.findings-acl)
Copied to clipboard
| Challenge: | Existing studies show that attention mechanisms can improve models' interpretation, but they are not explicable. |
| Approach: | They propose a framework consisting of a learner and a compressor to purify attention scores . they propose to fine-tune and compress the attention mechanism to obtain a more faithful explanation . |
| Outcome: | The proposed framework improves performance and interpretability on eight benchmark datasets. |
Confronting LLMs with Traditional ML: Rethinking the Fairness of Large Language Models in Tabular Classifications (2024.naacl-long)
Copied to clipboard
| Challenge: | Recent studies suggest using large language models to make tabular classifications . however, LLMs have been shown to exhibit harmful social biases based on stereotypes and inequalities present in society. |
| Approach: | They propose to use large language models to make tabular classifications . they show that LLMs inherit biases from their training data . |
| Outcome: | The proposed models exhibit harmful biases that reflect stereotypes and inequalities in society. |
Memory-efficient NLLB-200: Language-specific Expert Pruning of a Massively Multilingual Machine Translation Model (2023.acl-long)
Copied to clipboard
| Challenge: | Neural Machine Translation models are based on a Mixture of Experts architecture and can be pruned to remove up to 80% of experts without further finetuning. |
| Approach: | They propose a pruning method that removes up to 80% of experts without further finetuning and with a negligible loss in translation quality. |
| Outcome: | The proposed pruning method removes up to 80% of experts without further finetuning and with a negligible loss in translation quality. |
DAMP: Doubly Aligned Multilingual Parser for Task-Oriented Dialogue (2023.acl-long)
Copied to clipboard
| Challenge: | Existing studies show that multilingual models are less robust for semantic parsing compared to other tasks. |
| Approach: | They propose a constrained optimization technique to optimize multilingual parsing systems for multilingual use. |
| Outcome: | The proposed technique outperforms XLM-R and mT5-Large on three benchmarks and significantly outperformed other models. |
Long Context Question Answering via Supervised Contrastive Learning (2022.naacl-main)
Copied to clipboard
| Challenge: | Long-context question answering tasks often require identifying evidence spans (e.g., sentences) prior work showed that jointly training models to perform evidence extraction and question answering is important for achieving high performance. |
| Approach: | They propose a method for equipping long-context QA models with an additional sequence-level objective for better identification of the supporting evidence. |
| Outcome: | The proposed method exhibits consistent improvements on three different strong long-context transformer models, across two challenging question answering benchmarks – HotpotQA and QAsper. |
Entity Tracking in Language Models (2023.acl-long)
Copied to clipboard
| Challenge: | Existing studies on the ability of large language models to track discourse entities have not been conducted. |
| Approach: | They propose to investigate whether large language models can track entities . they first investigate whether Flan-T5, GPT-3 and GPT-3.5 can track the state of entities based on an English description of the initial state and a series of state-changing operations. |
| Outcome: | The proposed task investigates whether language models can track entities based on language descriptions and state-changing operations. |
Making Pretrained Language Models Good Long-tailed Learners (2022.emnlp-main)
Copied to clipboard
| Challenge: | Prompt-tuning has shown appealing performance in few-shot classification . however, it is less promising in long-tailed classification due to long tail . |
| Approach: | They propose to use prompt-tuning to make pretrained language models at least good long-tailed learners by bridging the gap between prompt- and commonly used finetun. |
| Outcome: | The proposed method makes pretrained language models at least good long-tailed learners, bridging the gap between prompt-tuning and finetunation. |
M2PT: Multimodal Prompt Tuning for Zero-shot Instruction Learning (2024.emnlp-main)
Copied to clipboard
Taowen Wang, Yiyang Liu, James Liang, Junhan Zhao, Yiming Cui, Yuning Mao, Shaoliang Nie, Jiahao Liu, Fuli Feng, Zenglin Xu, Cheng Han, Lifu Huang, Qifan Wang, Dongfang Liu
| Challenge: | Multimodal Large Language Models (MLLMs) exhibit remarkable performance across a wide range of domains. |
| Approach: | They propose a multimodal prompt tuning approach for efficient instruction tuning of MLLMs. |
| Outcome: | The proposed approach shows superior performance on multimodal evaluation datasets compared to state-of-the-art methods. |
Med-MoE: Mixture of Domain-Specific Experts for Lightweight Medical Vision-Language Models (2024.findings-emnlp)
Copied to clipboard
| Challenge: | Recent advances in multimodal large language models have seen remarkable progress for medical decision-making, however, they are designated for specific classification or generative tasks and require model training or finetuning on large-scale datasets with sizeable parameters and tremendous computing. |
| Approach: | They propose a framework that tackles discriminative and generative multimodal medical tasks using multimodal alignment, instruction tuning and routing. |
| Outcome: | The proposed model can achieve superior performance to or on par with state-of-the-art baselines while only requiring 30%-50% of activated model parameters. |
Cutting Down on Prompts and Parameters: Simple Few-Shot Learning with Language Models (2022.findings-acl)
Copied to clipboard
| Challenge: | Prompting language models (LMs) with training examples and task descriptions has been seen as critical to recent successes in few-shot learning. |
| Approach: | They propose to fine tune masked language models with training examples and task descriptions to reduce prompt engineering by using null prompts. |
| Outcome: | The proposed prompts can be used to improve few-shot learning by finetuning only the bias terms while updating only 0.1% of the parameters. |
Advancing Beyond Identification: Multi-bit Watermark for Large Language Models (2024.naacl-long)
Copied to clipboard
| Challenge: | Existing methods to detect machine-generated text focus on detection, but some misuses require tracing the adversary user for counteracting them. |
| Approach: | They propose a method for embedding traceable multi-bit information during language model generation. |
| Outcome: | The proposed method outperforms existing methods in terms of robustness and latency while maintaining text quality. |
FIREBALL: A Dataset of Dungeons and Dragons Actual-Play with Structured Game State Information (2023.acl-long)
Copied to clipboard
| Challenge: | Recent work shows that large language models that have access to state information can generate higher quality game turns than LLMs that use dialog history alone. |
| Approach: | They present a dataset of game play sessions from real D&D gameplay on Discord with true game state info. |
| Outcome: | The proposed model can generate executable Avrae commands, especially after fine tuning. |
Modular and Parameter-Efficient Multimodal Fusion with Prompting (2022.findings-acl)
Copied to clipboard
| Challenge: | Recent research has made impressive progress in large-scale multimodal pre-training. |
| Approach: | They propose to use prompt vectors to align multimodal modalities by pretraining text inputs with prompts or embedding vectors. |
| Outcome: | The proposed method achieves comparable performance to several other multimodal fusion methods in low-resource settings. |
ToxiGen: A Large-Scale Machine-Generated Dataset for Adversarial and Implicit Hate Speech Detection (2022.acl-long)
Copied to clipboard
| Challenge: | Toxic language detection systems often falsely flag text that contains minority group mentions as toxic . this over-reliance on spurious correlations also causes systems to struggle with detecting implicitly toxic language. |
| Approach: | They develop a machine-generated dataset of toxic and benign statements about 13 minority groups that generates subtly toxic and harmless text with a massive pretrained language model. |
| Outcome: | The proposed method can detect toxic and benign statements on a large scale . it can also detect hate speech on 94.5% of the toxic examples . |
Continual Mixed-Language Pre-Training for Extremely Low-Resource Neural Machine Translation (2021.findings-acl)
Copied to clipboard
| Challenge: | a lack of data in low-resource languages has limited the performance of a multilingual pre-trained model. |
| Approach: | They propose a continuous pre-training framework to adapt mBART to unseen languages . they construct noisy mixed-language text from the monolingual corpus of the target language . |
| Outcome: | The proposed framework improves finetuning performance on low-resource translation pairs . the proposed framework also improves on translation pairs where both languages are seen . |
Masked Language Model Scoring (2020.acl-main)
Copied to clipboard
| Challenge: | Pretrained masked language models require finetuning for most tasks. |
| Approach: | They evaluate pretrained masked language models out of the box via their pseudo-log-likelihood scores (PLLs) they attribute this success to PLL’s unsupervised expression of linguistic acceptability without a left-to-right bias, greatly improving on scores from GPT-2 . |
| Outcome: | The proposed model outperforms autoregressive language models in a variety of tasks. |
F-Actor: Controllable Conversational Behavior in Full-Duplex Models (2026.findings-acl)
Copied to clipboard
| Challenge: | Current spoken conversational systems lack customization capabilities, limiting their naturalness and usability. |
| Approach: | They propose an instruction-following full-duplex conversational speech model that can be trained efficiently under typical academic resource constraints. |
| Outcome: | The proposed model requires just 2,000 hours of data to be trained under typical academic resource constraints. |
Explanation-based Finetuning Makes Models More Robust to Spurious Cues (2023.acl-long)
Copied to clipboard
Josh Magnus Ludan, Yixuan Meng, Tai Nguyen, Saurabh Shah, Qing Lyu, Marianna Apidianaki, Chris Callison-Burch
| Challenge: | Large Language Models (LLMs) learn correlations between labels and features that are irrelevant to the task, leading to poor generalization on out-of-distribution data. |
| Approach: | They propose an explanation-based approach to fine tune large language models to generate a free-text explanation supporting their answer. |
| Outcome: | The proposed model is more robust against spurious cues in terms of accuracy drop across four classification tasks: ComVE (+1.2), CREAK (+9.1), e-SNLI (+5.4), and SBIC (+6.5). |
MiLoRA: Harnessing Minor Singular Components for Parameter-Efficient LLM Finetuning (2025.naacl-long)
Copied to clipboard
| Challenge: | Efficient finetuning of large language models (LLMs) aims to adapt the LLMs with reduced computational and memory costs. |
| Approach: | They propose a simple yet effective method that initializes low-rank matrices with Gaussian distribution and zero values while keeping the original weight matrics frozen. |
| Outcome: | The proposed approach only updates the minor components of the weight matrix while keeping the principal singular components frozen. |
Context Length Extension via Generalized Extrapolation Scale (2024.findings-acl)
Copied to clipboard
| Challenge: | Existing work on extrapolating positional embedding (RoPE) has limited results in the application of long context language models. |
| Approach: | They propose a set of parameterized extrapolation functions applied to each layer and attention head to adaptively adjust its extrapolations scales. |
| Outcome: | The proposed model achieves stable extrapolation on 64k contexts by training on 16k length text. |
Multi-Objective Linguistic Control of Large Language Models (2024.findings-acl)
Copied to clipboard
| Challenge: | Existing Large language models prefer to generate verbose responses due to the length bias, which may increase unnecessary reading complexity. |
| Approach: | They propose to use off-the-shelf data to fine tune multiple linguistic complexities of LLM outputs to improve multi-complexity controllability and improve the quality of the responses. |
| Outcome: | The proposed method improves multi-complexity controllability significantly and retains or enhances the quality of the responses as a side benefit. |
2INER: Instructive and In-Context Learning on Few-Shot Named Entity Recognition (2023.findings-emnlp)
Copied to clipboard
| Challenge: | Named Entity Recognition (NER) tasks are a fundamental task of natural language processing (NLP). |
| Approach: | They propose a text-to-text framework for Few-Shot Named Entity Recognition (NER) that employs instruction finetuning and auxiliary tasks to enhance the model's understanding of entity types in the overall semantic context of a sentence. |
| Outcome: | The proposed framework outperforms existing Few-Shot NER methods and remains competitive with state-of-the-art NER algorithms. |
Narrate Dialogues for Better Summarization (2022.findings-emnlp)
Copied to clipboard
| Challenge: | Recent work on dialogue summarization models focuses on generating concise summaries for multi-party dialogues. |
| Approach: | They propose several ways to convert dialogue into a third-person narrative style . they propose to use narration as a valuable annotation for LLMs . |
| Outcome: | Empirical results show that the proposed approach achieves higher scores on ROUGE and a factual correctness metric. |
Steering Safely or Off a Cliff? Rethinking Specificity and Robustness in Inference-Time Interventions (2026.eacl-long)
Copied to clipboard
| Challenge: | Existing studies have shown that model steering can preserve fluency and unrelated abilities, but it fails to preserve robustness specificity. |
| Approach: | They propose a framework that distinguishes three dimensions of specificity: general, control, and robustness. |
| Outcome: | The proposed framework distinguishes three dimensions of specificity: general (preserving fluency and unrelated abilities), control (preserving related control properties), and robustness (preserving control properties under distribution shifts). |
Retrieval Enhanced Model for Commonsense Generation (2021.findings-acl)
Copied to clipboard
| Challenge: | Existing frameworks for commonsense generation are lacking for pre-trained models. |
| Approach: | They propose a framework that uses concept matching to retrieve prototype sentences and trainable sentence retriever to enhance pre-training and fine-tuning. |
| Outcome: | The proposed framework achieves state-of-the-art on the large-scale Common-Gen benchmark. |
Neural Pipeline for Zero-Shot Data-to-Text Generation (2022.acl-long)
Copied to clipboard
| Challenge: | In data-to-text generation, training on in-domain data leads to overfitting and repeating training data noise. |
| Approach: | They propose to train pretrained language models on general-domain text-based operations by transforming single-item descriptions with modules trained on ordering, aggregation, and paragraph compression. |
| Outcome: | The proposed approach enables D2T generation from RDF triples in zero-shot settings. |
PALT: Parameter-Lite Transfer of Language Models for Knowledge Graph Completion (2022.findings-emnlp)
Copied to clipboard
| Challenge: | Pretrained language models (LMs) are a powerful transfer learning approach for knowledge graph (KG) completion. |
| Approach: | They propose a parameter-lite transfer learning approach for pretrained language models for knowledge graph (KG) completion. |
| Outcome: | The proposed model outperforms the state-of-the-art models on a knowledge graph completion benchmark by tuning 1% of the parameters. |
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. |
Concept-skill Transferability-based Data Selection for Large Vision-Language Models (2024.emnlp-main)
Copied to clipboard
| Challenge: | Large Vision-Language Models (LVLMs) require instruction tuning on extensive data . training on large VL datasets can be prohibitively expensive . |
| Approach: | They propose a data selection technique that uses a small model as a reference model to select training data for efficient finetuning of a target LVLM. |
| Outcome: | The proposed method achieves superior performance and data selection efficiency against 8 strong baselines on two distinct datasets: LLaVA-1.5 and Vision-Flan. |
Aligning to Constraints for Data-Efficient Language Model Customization (2025.findings-naacl)
Copied to clipboard
Fei Wang, Chao Shang, Shuai Wang, Sarthak Jain, Qiang Ning, Bonan Min, Vittorio Castelli, Yassine Benajiba, Dan Roth
| Challenge: | General-purpose language models (LMs) are aligned to diverse user intents, but fall short when it comes to specific applications. |
| Approach: | They propose a framework that uses constraints to automatically produce supervision signals for user alignment with constraints. |
| Outcome: | The proposed framework can produce supervision signals for user alignment with constraints. |
Chain-of-Rank: Enhancing Large Language Models for Domain-Specific RAG in Edge Device (2025.findings-naacl)
Copied to clipboard
| Challenge: | Retrieval-augmented generation (RAG) is valuable in specialized domains where precision is critical. |
| Approach: | They propose a chain-of-rank algorithm which allows LLMs to access a target domain early via finetuning. |
| Outcome: | The proposed method achieves state-of-the-art in benchmarks and analyzes its efficacy. |
Event Detection with a Context-Aware Encoder and LoRA for Improved Performance on Long-Tailed Classes (2026.findings-eacl)
Copied to clipboard
| Challenge: | Decoder-only models dominate the event detection literature, but their unidirectional attention mechanism has been a roadblock in getting strong performance on embedding. |
| Approach: | They propose to use Macro-F1 as a more representative measure of a model’s ability across the long-tail of event types to improve their models' performance. |
| Outcome: | The proposed model improves on the decoder-only models, showing that low-rank Adaptation can be an effective tool to enhance LLMs’ performance on long-tailed event classes. |
Evaluating Parameter Efficient Learning for Generation (2022.emnlp-main)
Copied to clipboard
Peng Xu, Mostofa Patwary, Shrimai Prabhumoye, Virginia Adams, Ryan Prenger, Wei Ping, Nayeon Lee, Mohammad Shoeybi, Bryan Catanzaro
| Challenge: | Parameter efficient learning methods (PERMs) are gaining attention for their ability to adapt to a downstream task. |
| Approach: | They propose to use parameter efficient learning methods to improve model adaptation . they compare in-domain evaluations and generalizations to unseen domains and new datasets . |
| Outcome: | The proposed method outperforms finetuning and PERMs in in-domain evaluations. |
DeepInsert: Early Layer Bypass for Efficient and Performant Multimodal Understanding (2026.eacl-long)
Copied to clipboard
Moulik Choraria, Xinbo Wu, Akhil Bhimaraju, Nitesh Sekhar, Yue Wu, Xu Zhang, Prateek Singhal, Lav R. Varshney
| Challenge: | Recent work shows that hyperscaling of data and parameter count in LLMs is yielding diminishing improvement when weighed against training costs. |
| Approach: | They propose to insert multimodal tokens directly into the middle of the model to bypass the early layers. |
| Outcome: | The proposed method reduces training and inference costs while preserving performance. |
AutoPrompt: Eliciting Knowledge from Language Models with Automatically Generated Prompts (2020.emnlp-main)
Copied to clipboard
| Challenge: | Pretrained language models have been successful when finetuned to downstream tasks . however, it is difficult to determine whether the knowledge that finetuning LMs contain is learned during the pretraining or the finetailing process. |
| Approach: | They propose a method to create prompts for a diverse set of tasks using a gradient-guided search. |
| Outcome: | The proposed method performs sentiment analysis and natural language inference without additional parameters and finetuning. |
MultiTabQA: Generating Tabular Answers for Multi-Table Question Answering (2023.acl-long)
Copied to clipboard
| Challenge: | Recent tabular question answering models only answer questions over a single table . multi-table operations often result in tabular outputs . |
| Approach: | They propose a model that answers questions over multiple tables and generalizes to generate tabular answers. |
| Outcome: | The proposed model outperforms state-of-the-art single table QA models on a multi-table QA setting. |
Look Before You Leap: A Lookahead Reasoning Quality Gate for Speculative Decoding (2026.eacl-long)
Copied to clipboard
| Challenge: | Unlike token-level likelihood search, which is myopic and often rewards verbosity, our approach works at an intermediate granularity. |
| Approach: | They propose a lookahead quality gate for speculative decoding that accepts the longest reliable prefix of each k-token lookaheaded draft. |
| Outcome: | The proposed method improves accuracy over baselines while achieving 2.6-7.9 faster generation on math and science benchmarks. |
The Unreasonable Effectiveness of Easy Training Data for Hard Tasks (2024.acl-long)
Copied to clipboard
| Challenge: | Existing pretrained language models perform well on hard data, but hard data is noisier and costlier to collect. |
| Approach: | They propose to use in-context learning, linear classifier heads, and QLoRA to show that pretrained language models generalize relatively well from easy to hard data. |
| Outcome: | The proposed model generalizes well from easy to hard data even better than oracle models finetuned on hard data. |
Finding Memo: Extractive Memorization in Constrained Sequence Generation Tasks (2022.findings-emnlp)
Copied to clipboard
| Challenge: | Memorization presents a challenge for constrained Natural Language Generation tasks . previous studies focused on counterfactual memorization, linking it to hallucinations . |
| Approach: | They propose an algorithm for extractive memorization in constrained sequence generation tasks . they propose to elicit non-memorized translations of memorized samples from the same model . |
| Outcome: | The proposed algorithm could be leveraged to mitigate memorization in the model through finetuning. |
Discovering Low-rank Subspaces for Language-agnostic Multilingual Representations (2022.emnlp-main)
Copied to clipboard
| Challenge: | Existing studies show that pre-trained ML-LMs can achieve zero-shot cross-lingual transfer without explicit cross-linguistic supervision. |
| Approach: | They propose a method to remove language-specific factors from multilingual embedding spaces by using a single value decomposition method with multiple monolingual corpora as input. |
| Outcome: | The proposed method can boost language agnosticism without finetuning . Empirical results show that it consistently leads to improvements over existing models. |
Learning Semantic Structure through First-Order-Logic Translation (2024.findings-emnlp)
Copied to clipboard
| Challenge: | a recent study shows that transformer-based language models can confuse which predicates apply to which objects . a this is a crucial building block of semantic structure, but if an LM mixes up which objects have which property, it makes errors in reasoning . |
| Approach: | They propose to use transformer-based language models to learn predicate argument structure from simple sentences. |
| Outcome: | The proposed model can learn predicate argument structure from simple sentences. |
CitRet: A Hybrid Model for Cited Text Span Retrieval (2022.coling-1)
Copied to clipboard
| Challenge: | Current methods for citing text span retrieval (CTSR) rely on pre-trained off-the-shelf deep learning models like SciBERT. |
| Approach: | They propose a hybrid model for cited text span retrieval that leverages unique semantic and syntactic structural characteristics of scientific documents. |
| Outcome: | The proposed model improves state-of-the-art by 15% on the CLSciSumm shared tasks. |
Key ingredients for effective zero-shot cross-lingual knowledge transfer in generative tasks (2024.naacl-long)
Copied to clipboard
| Challenge: | Existing studies have focused on zero-shot cross-lingual transfer . mBERT, mBART and mT5 provide high-quality representations for texts in various languages . |
| Approach: | They propose to use mBART and NLLB-200 to finetune a multilingual pretrained language model on input-output pairs in one language and use it to make task predictions for inputs in other languages. |
| Outcome: | The proposed approach significantly reduces generation in the wrong language with full finetuning and can be competitive in some cases. |
Byte-Level Grammatical Error Correction Using Synthetic and Curated Corpora (2023.acl-long)
Copied to clipboard
Svanhvít Lilja Ingólfsdóttir, Petur Ragnarsson, Haukur Jónsson, Haukur Simonarson, Vilhjalmur Thorsteinsson, Vésteinn Snæbjarnarson
| Challenge: | Spelling mistakes due to typos and rushed writing, nonstandard punctuation and spelling, and grammatical and stylistic issues are common to almost everyone who writes any kind of text. |
| Approach: | They propose to use a common subword unit vocabulary and byte-level encoding to fine tune two subword-level models and one byte level model on hand-corrected error corpora. |
| Outcome: | The proposed model improves accuracy for spelling and grammatical errors and more complex errors. |
Unsupervised Paraphrasing with Pretrained Language Models (2021.emnlp-main)
Copied to clipboard
| Challenge: | Paraphrase generation has benefited from recent advances in the design of training objectives and model architectures, but previous studies focused on supervised methods that require a large amount of labeled data that is costly to collect. |
| Approach: | They propose a transfer learning approach that enables pre-trained language models to generate high-quality paraphrases in an unsupervised setting. |
| Outcome: | The proposed model performs state-of-the-art on the Quora Question Pair and ParaNMT datasets and is robust to domain shift between the two datasets. |
A Semantic-Aware Layer-Freezing Approach to Computation-Efficient Fine-Tuning of Language Models (2025.findings-acl)
Copied to clipboard
| Challenge: | Existing work on how to finetune but neglects the issue of where to fine-tune language models is expensive. |
| Approach: | They propose to use transition traces of latent representation to compute deviations (or loss) and then estimate the gain of each layer in reducing deviation (or gain). |
| Outcome: | The proposed approach outperforms baseline methods and is cost-benefit balanced. |
MaxMatch-Dropout: Subword Regularization for WordPiece (2022.coling-1)
Copied to clipboard
| Challenge: | Existing subword regularization methods are specialized to a particular tokenizer type. |
| Approach: | They propose a subword regularization method for WordPiece that uses a maximum matching algorithm for tokenization. |
| Outcome: | The proposed method improves the performance of text classification and machine translation tasks as well as other subword regularization methods. |
Calibrating Factual Knowledge in Pretrained Language Models (2022.findings-emnlp)
Copied to clipboard
| Challenge: | Existing studies show that Pretrained Language Models can store factual knowledge, but facts stored in PLMs are not always correct. |
| Approach: | They propose a lightweight method to calibrate factual knowledge in PLMs without re-training from scratch. |
| Outcome: | The proposed method can be used to calibrate factual knowledge in PLMs without re-training from scratch. |
TADPOLE: Task ADapted Pre-Training via AnOmaLy DEtection (2021.emnlp-main)
Copied to clipboard
| Challenge: | Existing approaches to solve domain shifts in NLP tasks require additional pre-training . current approaches focus on the downstream corpus when it is small, but are not effective . |
| Approach: | They propose a task-adapted pre-training framework that can be used when the downstream corpus is too small for additional pre-tuning. |
| Outcome: | The proposed framework outperforms baseline methods on biomedical, computer science, news, and movie reviews tasks. |
Domain-Agnostic Adapter Architecture for Deception Detection: Extensive Evaluations with the DIFrauD Benchmark (2024.lrec-main)
Copied to clipboard
| Challenge: | Existing research focuses predominantly on specific fields, which results in the need for clarity on linguistic markers associated with deception. |
| Approach: | They propose a domain-independent fraud detection benchmark with 100,000 honest and misleading statements in seven domains and a parameter-efficient finetuning adapter to improve tuning methods. |
| Outcome: | The proposed adapter outperforms all competition on the DIFrauD benchmark and is able to predict the performance of the proposed model. |
On Robustness of Finetuned Transformer-based NLP Models (2023.findings-emnlp)
Copied to clipboard
| Challenge: | Pretrained Transformer-based language models have been finetuned for a large number of tasks. |
| Approach: | They characterize changes between pretrained and finetuned models with CKA and STIR metrics. |
| Outcome: | The proposed models are more robust to perturbations than BERT and T5 on classification tasks and generation tasks. |
NaijaHate: Evaluating Hate Speech Detection on Nigerian Twitter Using Representative Data (2024.acl-long)
Copied to clipboard
Manuel Tonneau, Pedro Quinta De Castro, Karim Lasri, Ibrahim Farouq, Lakshmi Subramanian, Victor Orozco-Olvera, Samuel Fraiberger
| Challenge: | a recent study shows that hate speech detection systems are often evaluated on non-representative samples, raising concerns about overestimating performance in real-world settings. |
| Approach: | They propose a pretrained hate speech detection model that is annotated on a representative sample of Nigerian tweets and propose heuristics for domain-adaptive pretraining and finetuning. |
| Outcome: | The proposed model overestimates real-world performance by at least twofold compared to a dataset from the United States and Nigeria . the proposed model requires ten thousand Nigerian tweets flagged as hateful daily to moderate 60% of hateful content . |
Distilling Step-by-Step! Outperforming Larger Language Models with Less Training Data and Smaller Model Sizes (2023.findings-acl)
Copied to clipboard
Cheng-Yu Hsieh, Chun-Liang Li, Chih-kuan Yeh, Hootan Nakhost, Yasuhisa Fujii, Alex Ratner, Ranjay Krishna, Chen-Yu Lee, Tomas Pfister
| Challenge: | Deploying large language models (LLMs) is difficult because they are memory inefficient and compute-intensive for practical applications. |
| Approach: | They propose a mechanism that fine tunes or distills small models that outperform LLMs . they use human labels to fine tune models or LLM-generated labels to train models . |
| Outcome: | The proposed method outperforms LLMs by using fewer training examples compared to few-shot prompted models using substantially smaller model sizes. |
Parameter-Efficient Tuning Makes a Good Classification Head (2022.emnlp-main)
Copied to clipboard
| Challenge: | In recent years, pretrained models revolutionized the paradigm of natural language understanding . but the final-layer output of the backbone, i.e. the input of the classification head, will change greatly during finetuning . |
| Approach: | They propose to append a randomly initialized classification head after the pretrained backbone and finetune the whole model. |
| Outcome: | The proposed classification head can be replaced with the randomly initialized heads for a stable performance gain. |
Data-scarce Behavior Editing of Language Models (2025.findings-emnlp)
Copied to clipboard
| Challenge: | Prior studies show that noisy neural circuitries coexist with generalizable abilities within LLMs. |
| Approach: | a new method is proposed to improve the generalizability of large-scale web-based text models . a TaRot method is based on learnable rotation matrices optimized for Bayesian optimization . |
| Outcome: | a new method for task adaptation improves on multiple classification and generation tasks . it improves upon zero- and few-shot performance, with average improvements of 14% and 15% . |
Impact of Co-occurrence on Factual Knowledge of Large Language Models (2023.findings-emnlp)
Copied to clipboard
| Challenge: | Large language models (LLMs) often make factually incorrect responses despite their success in various applications. |
| Approach: | They propose to fine tune large language models to mitigate the co-occurrence bias by filtering out biased samples with high subject-object co-occurring counts. |
| Outcome: | The proposed model scales up to debiased datasets to mitigate the co-occurrence bias, but is not effective in recalling rare facts unseen during finetuning. |
Polyjuice: Generating Counterfactuals for Explaining, Evaluating, and Improving Models (2021.acl-long)
Copied to clipboard
| Challenge: | Existing counterfactual generation methods rely on manual labor to create very few counterf actuals or only instantiate limited types of perturbations such as paraphrases or word substitutions. |
| Approach: | They propose a general-purpose counterfactual generator that allows for control over perturbation types and locations. |
| Outcome: | The proposed generator produces diverse sets of realistic counterfactuals that are useful in various applications. |
Cross-Lingual Retrieval Augmented Prompt for Low-Resource Languages (2023.findings-acl)
Copied to clipboard
| Challenge: | Multilingual pretrained language models (MPLMs) perform strongly in cross-lingual transfer. |
| Approach: | They propose to augment context with similar sentences retrieved from a high-resource language (HRL) they find a significant correlation between cross-lingual transfer performance and similarity between high- and low-resourced languages . |
| Outcome: | The proposed model outperforms finetuning by 3.7% on three downstream tasks with multilingual parallel test sets across 10 LRLs covering 6 language families in unlabeled and labeled settings. |
Unsupervised Summarization Re-ranking (2023.findings-acl)
Copied to clipboard
| Challenge: | Abstractive summarization models have been gaining popularity, but performance of unsupervised models still lags behind supervised models. |
| Approach: | They propose to re-rank summary candidates in an unsupervised manner to close the performance gap between unsupervised and supervised models. |
| Outcome: | The proposed model improves unsupervised models by up to 7.27% and ChatGPT by up 6.86% relative mean ROUGE across four widely-adopted summarization benchmarks. |
Generating Datasets with Pretrained Language Models (2021.emnlp-main)
Copied to clipboard
| Challenge: | Recent approaches to obtain high-quality sentence embeddings from pretrained language models require labeled data or finetuned on large set of labeles. |
| Approach: | They propose to use generative abilities of large and high-performing PLMs to generate entire datasets of labeled text pairs from scratch and fine tune much smaller and more efficient models. |
| Outcome: | The proposed approach outperforms baselines on several semantic textual similarity datasets. |
NeuroAda: Activating Each Neuron’s Potential for Parameter-Efficient Fine-Tuning (2025.emnlp-main)
Copied to clipboard
| Challenge: | Existing methods for parameter-efficient fine-tuning are limited and require computational and memory resources. |
| Approach: | They propose a parameter-efficient fine-tuning method that enables fine-grained model finetunation while maintaining high memory efficiency. |
| Outcome: | The proposed method reduces CUDA memory usage by up to 60% while maintaining high performance. |
NLKI: A Lightweight Natural Language Knowledge Integration Framework for Improving Small VLMs in Commonsense VQA Tasks (2025.findings-emnlp)
Copied to clipboard
| Challenge: | Small vision-language models lag behind their larger generative counterparts due to lack of knowledge. |
| Approach: | They propose a framework that integrates commonsense knowledge into small vision-language models . the framework retrieves natural language facts and prompts an LLM to craft natural language explanations . |
| Outcome: | The proposed framework retrieves natural language facts and prompts an LLM to craft natural language explanations. |
Data-Efficient Finetuning Using Cross-Task Nearest Neighbors (2023.findings-acl)
Copied to clipboard
| Challenge: | Prior work shows training models on multitask data augmented with task descriptions transfers knowledge to new tasks. |
| Approach: | They propose to use unlabeled target-task data to train models on task descriptions . they use only 2% of the data from the P3 pool without labeled target task data . |
| Outcome: | The proposed model outperforms baseline models on 12 out of 14 datasets . it also provides better initialization than single model on target-task data . |
In-Context Learning with Long-Context Models: An In-Depth Exploration (2025.naacl-long)
Copied to clipboard
| Challenge: | In-context learning is limited by context length, but it can be used for many tasks. |
| Approach: | They study the behavior of in-context learning at an extreme context length . example retrieval shows excellent performance at low context lengths but has diminished gains . |
| Outcome: | The proposed model can perform many tasks with reasonable accuracy when a few examples are provided in-context. |
Simple and Effective Input Reformulations for Translation (2023.emnlp-main)
Copied to clipboard
| Challenge: | Foundation language models learn from their finetuning input context in different ways. |
| Approach: | They propose three different data efficient techniques to improve translation performance . they reformulate inputs during finetuning for challenging translation tasks . |
| Outcome: | The proposed techniques show significant improvements on the Flores200 translation benchmark. |
LANDeRMT: Dectecting and Routing Language-Aware Neurons for Selectively Finetuning LLMs to Machine Translation (2024.acl-long)
Copied to clipboard
| Challenge: | Existing studies have shown promising results in multilingual translation with limited bilingual supervision. |
| Approach: | They propose a Language-Aware Neuron Detecting and Routing framework that fine tunes LLMs to Machine Translation with diverse translation training data. |
| Outcome: | The proposed framework selectively finetunes LLMs to MT tasks with diverse translation training data. |
Discrete and Soft Prompting for Multilingual Models (2021.emnlp-main)
Copied to clipboard
| Challenge: | In few-shot learning, discrete and soft prompting perform better than finetuning in multilingual cases. |
| Approach: | They show that discrete and soft prompting perform better than finetuning in crosslingual transfer and in-language training of multilingual natural language inference. |
| Outcome: | The proposed prompting model outperforms finetuning in crosslingual transfer and in-language training of multilingual natural language inference. |
GALLa: Graph Aligned Large Language Models for Improved Source Code Understanding (2025.acl-long)
Copied to clipboard
| Challenge: | Programming languages have rich semantics that are represented by graphs and not available from the surface form of source code. |
| Approach: | They propose to use graph neural networks and cross-modal alignment technologies to inject structural information of code into LLMs as an auxiliary task during finetuning. |
| Outcome: | The proposed framework improves on five code tasks with six different baseline LLMs, while incurring no cost at inference time. |
Downstream Datasets Make Surprisingly Good Pretraining Corpora (2023.acl-long)
Copied to clipboard
| Challenge: | a dominant practice is to fine tune large pretrained transformer models using smaller downstream datasets . performance gains are not always attributable to the use of external data in massive amounts . |
| Approach: | They propose to use the same (downstream) training data for pretraining and finetuning to compare models. |
| Outcome: | The proposed model outperforms standard pretraining on the BookWiki corpus on 7 and 5 datasets. |
GPT-SW3: An Autoregressive Language Model for the Scandinavian Languages (2024.lrec-main)
Copied to clipboard
Ariel Ekgren, Amaru Cuba Gyllensten, Felix Stollenwerk, Joey Öhman, Tim Isbister, Evangelia Gogoulou, Fredrik Carlsson, Judit Casademont, Magnus Sahlgren
| Challenge: | a growing interest in building and applying large language models for languages other than English is fueling interest in developing LLMs for smaller languages. |
| Approach: | They describe the development process for the first native large generative language model for the North Germanic languages, GPT-SW3. |
| Outcome: | The proposed model is based on the generative language model for the North Germanic languages . it is a first-generation model with a high-quality data set and a low cost of implementation . |
Avoiding Inference Heuristics in Few-shot Prompt-based Finetuning (2021.emnlp-main)
Copied to clipboard
| Challenge: | Recent prompt-based approaches allow pretrained language models to achieve strong performances on few-shot finetuning by reformulating downstream task instances as a language modeling problem. |
| Approach: | They propose to reformulate downstream tasks as a language modeling problem and add a regularization that preserves pretraining weights to the model to mitigate the destructive tendency of few-shot finetuning. |
| Outcome: | The proposed model performs better on low data regimes than the standard model on few-shot finetuning. |
Exploiting Structured Knowledge in Text via Graph-Guided Representation Learning (2020.emnlp-main)
Copied to clipboard
| Challenge: | Existing methods for integrating knowledge graphs into pre-trained language models have been poorly implemented. |
| Approach: | They propose a self-supervised entity masking scheme that exploits relational knowledge underlying the text. |
| Outcome: | The proposed model achieves improved performance on five benchmarks, including question answering and knowledge base completion. |
Continual Contrastive Finetuning Improves Low-Resource Relation Extraction (2023.acl-long)
Copied to clipboard
| Challenge: | Relation extraction (RE) has been challenging in low-resource domains and with limited resources. |
| Approach: | They propose to pretrain and finetune the RE model using consistent objectives of contrastive learning. |
| Outcome: | The proposed method outperforms PLM-based RE classifier on two document-level RE datasets. |
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. |
Steering Large Language Models for Machine Translation with Finetuning and In-Context Learning (2023.findings-emnlp)
Copied to clipboard
Duarte Alves, Nuno Guerreiro, João Alves, José Pombal, Ricardo Rei, José de Souza, Pierre Colombo, Andre Martins
| Challenge: | Large language models (LLMs) are a promising avenue for machine translation (MT) however, their effectiveness depends on the choice of few-shot examples and they often require extra post-processing due to overgeneration. |
| Approach: | They propose a method that incorporates few-shot examples during finetuning to improve performance on MT tasks. |
| Outcome: | The proposed method outperforms few-shot prompting while eliminating the need for in-context examples. |
EvoBench: Towards Real-world LLM-Generated Text Detection Benchmarking for Evolving Large Language Models (2025.findings-acl)
Copied to clipboard
| Challenge: | Existing methods to detect LLM-generated texts rely on static benchmarks that neglect the evolving nature of LLMs. |
| Approach: | They propose a benchmark to evaluate the generalization of LLM-generated text detection methods. |
| Outcome: | The proposed benchmark measures generalization of 14 detection methods across LLMs. |
SMART: Submodular Data Mixture Strategy for Instruction Tuning (2024.findings-acl)
Copied to clipboard
| Challenge: | Existing methods for fine tuning language models are manual or rely on intuition. |
| Approach: | They propose a method which uses a submodular function to assign importance scores to tasks and then use them to determine mixture weights. |
| Outcome: | The proposed method outperforms traditional methods such as examples proportional mixing and equal mixing. |
Turning English-centric LLMs Into Polyglots: How Much Multilinguality Is Needed? (2024.findings-emnlp)
Copied to clipboard
| Challenge: | Existing models that target a single language are not seen during finetuning, but are able to respond in multiple languages once deployed in downstream applications. |
| Approach: | They investigate the minimal amount of multilinguality required during finetuning to elicit effective cross-lingual generalisation in English-centric LLMs. |
| Outcome: | The proposed model can respond in as few as two to three languages to a user's query in English, but the degree to which a target language is seen during pretraining is limiting. |
Jointly Reparametrized Multi-Layer Adaptation for Efficient and Private Tuning (2023.findings-acl)
Copied to clipboard
| Challenge: | Efficient finetuning of pretrained language transformers requires a large number of tunable parameters. |
| Approach: | They propose a language transformer finetuning strategy that introduces task-specific parameters in multiple transformer layers. |
| Outcome: | The proposed method outperforms other methods with 4,100 parameters on GLUE tasks with 5% of full finetuning performance. |
Geographical Erasure in Language Generation (2023.findings-emnlp)
Copied to clipboard
| Challenge: | Large language models encode vast amounts of world knowledge but are at risk of inordinately capturing information about dominant groups. |
| Approach: | They propose to operationalise a form of geographical erasure wherein language models underpredict certain countries. |
| Outcome: | The proposed model underpredicts certain countries by a factor 3 . the model is based on large datasets and is able to mitigate the effects . |
JDocQA: Japanese Document Question Answering Dataset for Generative Language Models (2024.lrec-main)
Copied to clipboard
| Challenge: | Document question answering is a task of question answering on given documents such as reports, slides, pamphlets, and websites. |
| Approach: | They propose a large-scale document-based QA dataset that requires both visual and textual information to answer questions. |
| Outcome: | The proposed dataset incorporates multiple categories of questions and unanswerable questions from the document for realistic question-answering applications. |
GRASS: Compute Efficient Low-Memory LLM Training with Structured Sparse Gradients (2024.emnlp-main)
Copied to clipboard
| Challenge: | Existing projection-based methods that project gradients into a lower-dimensional subspace can introduce computational and memory overheads. |
| Approach: | They propose a novel approach that leverages sparse projections to transform gradients into structured sparser updates. |
| Outcome: | The proposed approach significantly reduces memory usage for optimizer states and minimizes memory footprint, computation, and communication costs, leading to substantial throughput improvements. |
FLEXITOKENS: Flexible Tokenization for Evolving Language Models (2026.findings-acl)
Copied to clipboard
| Challenge: | Widely used subword tokenizers overfragment sequences in unseen domains, languages, and scripts . inefficient tokenizer models can cause overfragments in out-of-distribution domains if not trained properly . |
| Approach: | They propose a byte-level LM with learnable tokenizers to make tokenization adaptive . they propose 'flexitoken' which enables significantly greater flexibility during adaptation . |
| Outcome: | The proposed method significantly reduces token overfragmentation and improves on multilingual benchmarks and domains. |
AppBench: Planning of Multiple APIs from Various APPs for Complex User Instruction (2024.emnlp-main)
Copied to clipboard
| Challenge: | Existing state-of-the-art Large Language Models (LLMs) still cannot perform well in this situation even with the help of in-context learning and finetuning. |
| Approach: | They propose a benchmark to evaluate LLMs’ ability to plan and execute multiple APIs from various sources in order to complete the user’s task. |
| Outcome: | The proposed benchmarks show that the existing state-of-the-art LLMs still cannot perform well in this situation even with in-context learning and finetuning. |
MAPLE: Multilingual Evaluation of Parameter Efficient Finetuning of Large Language Models (2024.findings-acl)
Copied to clipboard
| Challenge: | Prior work on multilingual evaluation has shown that there is a large gap between the performance of Large Language Models on English and other languages. |
| Approach: | They propose to finetune Llama-2 and Mistral models on two datasets to determine their effect on model performance on six downstream tasks covering forty one languages. |
| Outcome: | The proposed model can improve on six multilingual tasks while degrading on high-resource languages. |
Modeling Gender and Dialect Bias in Automatic Speech Recognition (2024.findings-emnlp)
Copied to clipboard
| Challenge: | Dialect and gender-based biases have become an area of concern in language-dependent AI systems. |
| Approach: | They construct a podcast audio dataset and evaluate its performance . they then refine the models to better understand how finetuning can impact performance. |
| Outcome: | The proposed model improves on 13 hours of podcast audio transcribed by speakers of four US-based English dialects. |
Crosslingual Generalization through Multitask Finetuning (2023.acl-long)
Copied to clipboard
Niklas Muennighoff, Thomas Wang, Lintang Sutawika, Adam Roberts, Stella Biderman, Teven Le Scao, M Saiful Bari, Sheng Shen, Zheng Xin Yong, Hailey Schoelkopf, Xiangru Tang, Dragomir Radev, Alham Fikri Aji, Khalid Almubarak, Samuel Albanie, Zaid Alyafeai, Albert Webson, Edward Raff, Colin Raffel
| Challenge: | Multitask prompted finetuning (MTF) has been shown to help large language models generalize to new tasks in a zero-shot setting, but so far explorations of MTF have focused on English data and models. |
| Approach: | They apply multitask prompted finetuning to pretrained multilingual models and generate variants called BLOOMZ and mT0. |
| Outcome: | The proposed models can generalize to non-English languages that have never been seen before. |
Set the Clock: Temporal Alignment of Pretrained Language Models (2024.findings-acl)
Copied to clipboard
| Challenge: | Language models (LMs) are trained on web text originating from many points in time and, in general, without any explicit temporal grounding. |
| Approach: | They construct a time-sensitive question dataset and use it to examine temporal alignment methods to align their internal knowledge to a target time. |
| Outcome: | The proposed methods improve LLaMa2's performance by 62% if they are fine tuned to the year 2022 . |
LLoCO: Learning Long Contexts Offline (2024.emnlp-main)
Copied to clipboard
Sijun Tan, Xiuyu Li, Shishir G Patil, Ziyang Wu, Tianjun Zhang, Kurt Keutzer, Joseph Gonzalez, Raluca Popa
| Challenge: | Large language models are still unable to handle long contexts due to the quadratic computational and memory overhead of the self-attention mechanism and the substantial KV cache sizes during generation. |
| Approach: | They propose a method to learn contexts offline through context compression and in-domain parameter-efficient finetuning with LoRA. |
| Outcome: | The proposed model outperforms in-context learning while using 30 fewer tokens during inference and significantly reduces the cost of long document question answering. |
Mitigating Catastrophic Forgetting in Language Transfer via Model Merging (2024.findings-emnlp)
Copied to clipboard
| Challenge: | Large language models have shown remarkable capabilities, particularly in English, but for less prevalent languages, performance can be significantly lower, making additional adaptation paramount. |
| Approach: | They propose a new adaptation method based on iteratively merging multiple models fine-tuned on a subset of available training data that reduces forgetting while maintaining learning on the target domain. |
| Outcome: | The proposed method outperforms LLAMA-3-8B-based models in German and German while maintaining learning on the target domain. |
Multimodal Cross-Document Event Coreference Resolution Using Linear Semantic Transfer and Mixed-Modality Ensembles (2024.lrec-main)
Copied to clipboard
Abhijnan Nath, Huma Jamil, Shafiuddin Rehan Ahmed, George Arthur Baker, Rahul Ghosh, James H. Martin, Nathaniel Blanchard, Nikhil Krishnaswamy
| Challenge: | Existing methods for cross-document coreference resolution do not provide images for all mentions of events. |
| Approach: | They propose a multimodal cross-document event coreference resolution method that integrates visual and textual cues with a simple linear map between vision and language models. |
| Outcome: | The proposed method improves on a popular ECB+ and AIDA datasets. |
ChildEval:WHEN LARGE LANGUAGE MODELS MEET CHILDREN’S PERSONALITIES (2026.findings-acl)
Copied to clipboard
| Challenge: | Large Language Models (LLMs) have achieved remarkable success in effectively understanding and generating human language, leading to a revolutionary era in LLMs. |
| Approach: | They propose a benchmark to evaluate LLMs' ability to infer and follow child-centered preferences in long-context conversations. |
| Outcome: | The proposed benchmark spans five top-level and fourteen sub-level categories covering children’s daily lives and development. |
War of Thoughts: Competition Stimulates Stronger Reasoning in Large Language Models (2025.findings-acl)
Copied to clipboard
| Challenge: | Recent advances in Large Language Models (LLMs) have reshaped the landscape of reasoning tasks. |
| Approach: | They propose a method that enhances LLM reasoning without finetuning by using test-time scaling. |
| Outcome: | The proposed method outperforms baseline models in both budget and model size. |
ApiQ: Finetuning of 2-Bit Quantized Large Language Model (2024.emnlp-main)
Copied to clipboard
| Challenge: | Memory-efficient finetuning of large language models (LLMs) has attracted huge attention with the increasing size of LLMs due to the constraints posed by GPU memory limitations and the effectiveness of these methods compared to full finetune. |
| Approach: | They propose a memory-efficient finetuning framework called ApiQ to restore lost information from quantization by initializing LoRA components and quantizing weights of LLMs. |
| Outcome: | The proposed framework maintains the original LLM’s activation precision while mitigating error propagation from shallower into deeper layers. |
Qsnail: A Questionnaire Dataset for Sequential Question Generation (2024.lrec-main)
Copied to clipboard
| Challenge: | Questionnaires are a professional research methodology used for qualitative and quantitative analysis of human opinions, preferences, and behaviors. |
| Approach: | They propose a questionnaire-based dataset that consists of 13,168 human-written questionnaires. |
| Outcome: | The proposed dataset contains 13,168 human-written questionnaires gathered from online platforms. |
Preserving Language Capabilities in Vision-Language Models via Representation Regulation (2026.findings-acl)
Copied to clipboard
| Challenge: | Vision-Language Models (VLMs) provide a unified framework to process both text-only and vision-language tasks. |
| Approach: | They propose a method to reduce the distance between visual and textual representations by introducing a Representation Distribution Difference (RDD) loss. |
| Outcome: | Empirical evidence shows that finetuning VLMs on vision-language data has degraded language capabilities. |
Nearest Neighbor Normalization Improves Multimodal Retrieval (2024.emnlp-main)
Copied to clipboard
| Challenge: | Recent training-free methods suggest that accuracy can be improved without fine-tuning. |
| Approach: | They propose a method for correcting errors in trained contrastive image-text retrieval models with no additional training, called Nearest Neighbor Normalization. |
| Outcome: | The proposed method improves retrieval metrics for all contrastive models and datasets and does not require training on the reference database. |
ClusComp: A Simple Paradigm for Model Compression and Efficient Finetuning (2025.findings-acl)
Copied to clipboard
| Challenge: | Weight-only quantization reduces model size but suffers from performance degradation at lower bit widths. |
| Approach: | They propose a weight-only quantization paradigm that clusters weight matrices into codebooks and finetunes them block-by-block. |
| Outcome: | The proposed paradigm outperforms quantization methods and fine tunes LLMs to 1-bit compression and fine tuning. |
Do Generalisation Results Generalise? (2026.findings-acl)
Copied to clipboard
| Challenge: | Existing studies evaluating generalisation performance on large language models focuses on a single out-of-distribution dataset . |
| Approach: | They examine whether OOD generalisation results generalise across multiple OOD testsets throughout a finetuning run and then evaluate the partial correlation of results . |
| Outcome: | The proposed model achieves high scores on multiple OOD testsets, regressing out in-domain performance. |
Finetuning LLMs for Human Behavior Prediction in Social Science Experiments (2025.emnlp-main)
Copied to clipboard
| Challenge: | Large language models can be used to simulate social science experiments . finetuning LLMs directly on individual-level responses from past experiments improves accuracy . |
| Approach: | They propose to fine tune large language models directly on individual responses from past experiments to achieve multiple levels of generalization. |
| Outcome: | The proposed model outperforms GPT-4o in completely unseen studies by 36% . the proposed model reduces demographic parity difference by 10.6% compared to GPT-4) |
Efficient Many-Shot In-Context Learning with Dynamic Block-Sparse Attention (2025.acl-long)
Copied to clipboard
| Challenge: | Many-shot in-context learning shifts computational burden from training-time to inference-time, making deployment of many-shot ICL challenging to justify in-practice. |
| Approach: | They propose a method for retrieval-based many-shot in-context learning that uses blocks-sparse attention and retrieval of cached demonstrations to achieve comparable per-example latency to finetuning. |
| Outcome: | The proposed method achieves comparable per-example latency to finetuning while maintaining on average >95% of the best method’s accuracy across strong ICL and finetuned baselines. |
Measuring scalar constructs in social science with LLMs (2025.emnlp-main)
Copied to clipboard
Hauke Licht, Rupak Sarkar, Patrick Y. Wu, Pranav Goel, Niklas Stoehr, Elliott Ash, Alexander Miserlis Hoyle
| Challenge: | Valid scalar measurement of skalar constructs is a fundamental task in text analysis. |
| Approach: | They evaluate four approaches to measuring scalar constructs using large language models . pairwise comparisons produced better measurements than prompting LLMs, they say . validation of skalar measurement enables wide range of substantive applications in social science research . |
| Outcome: | The proposed methods improve on pairwise comparisons and finetuning . the proposed methods can be used in social science research . |
Activation Reward Models for Few-Shot Model Alignment (2026.findings-acl)
Copied to clipboard
Tianning Chai, Chancharik Mitra, Brandon Huang, Gautam Rajendrakumar Gare, Zhiqiu Lin, Assaf Arbelle, Leonid Karlinsky, Rogerio Feris, Trevor Darrell, Deva Ramanan, Roei Herzig
| Challenge: | A common approach is to use reward models that enable reinforcement-learning post-training. |
| Approach: | They propose a method that steers LLM activations to align with few-shot preference data without finetuning. |
| Outcome: | The proposed method surpasses zero-shot, few-shot and voting-based benchmarks on reward hacking and noise signals. |
Reading Between the Lines: The One-Sided Conversation Problem (2026.findings-acl)
Copied to clipboard
| Challenge: | In many real-world scenarios, only one side of a conversation is available for processing. |
| Approach: | They propose a one-sided conversation problem to reconstruct the missing speaker's turns and generate faithful summaries from one-side transcripts. |
| Outcome: | The proposed model improves reconstructions with prompting, but smaller models require fine tuning. |
EduVidQA: Generating and Evaluating Long-form Answers to Student Questions based on Lecture Videos (2025.emnlp-main)
Copied to clipboard
| Challenge: | This paper explores using Multimodal Large Language Models (MLLMs) to respond to student questions from online lectures . MLLM is a novel question answering task of real world significance . |
| Approach: | They propose to use Multimodal Large Language Models to automatically respond to student questions from online lectures by using a dataset of 5252 question-answer pairs from 296 computer science videos. |
| Outcome: | The proposed model can fine tune and fine tune questions from 296 computer science videos and show that students' preferences are important to the task. |
PROBE: PROcess-Based BEnchmark for Hallucination Detection (2026.findings-acl)
Copied to clipboard
Yu Zhang, Peter Belcak, Shizhe Diao, Yonggan Fu, Shaona Ghosh, Morteza Mardani, Eileen Margaret Peters Long, Bei Yu, Pavlo Molchanov
| Challenge: | Existing agentic applications rely on LLMs to self-assess the factuality of outputs . but current LLM systems fail to detect hallucinations . |
| Approach: | They propose a benchmark that breaks down hallucination detection into four critical steps . they show that when halluciation detection is treated as a multi-step process, all models achieve considerably better performance. |
| Outcome: | The proposed benchmark breaks down hallucination detection into four critical steps . it shows that when halluciation detection is treated as a multi-step process, all models achieve considerably better performance. |