Papers with internal representations
Deep RNNs Encode Soft Hierarchical Syntax (P18-2)
Copied to clipboard
| Challenge: | Existing studies show that syntactic information is useful for a wide variety of NLP tasks. |
| Approach: | They propose to use word-level representations to learn internal representations that capture soft hierarchical notions of syntax from highly varied supervision. |
| Outcome: | The proposed model encodes significant amounts of syntax even without explicit supervision. |
BERT is Not an Interlingua and the Bias of Tokenization (D19-61)
Copied to clipboard
| Challenge: | Cananical Correlation Analysis (CCA) of the internal representations of a pre- trained, multilingual BERT model reveals that the model partitions representations for each language rather than using a common, shared, interlingual space. |
| Approach: | They propose to use a multilingual BERT model to partition representations for each language rather than using a common, shared, interlingual space. |
| Outcome: | The results show that the model partitions representations for each language rather than using a common, shared, interlingual space. |
On the Limitations of Language-targeted Pruning: Investigating the Calibration Language Impact in Multilingual LLM Pruning (2026.tacl-1)
Copied to clipboard
| Challenge: | Recent advances in large language model pruning have shown high predictive performance in post-training settings. |
| Approach: | They conduct an empirical study on the performance and internal representation changes associated with pruning multilingual models for monolingual applications. |
| Outcome: | The proposed pruning methods retain perplexity and yield high signal-to-noise ratios, but not consistently improve downstream tasks. |
Thesis proposal: COGNILENS: Analyzing Cognitive Decline in Language Models for Alzheimer’s Monitoring (2026.eacl-srw)
Copied to clipboard
| Challenge: | Existing methods to detect AD and Mild Cognitive Impairment (MCI) are not effective in early stages. |
| Approach: | They propose to develop digital twins of Alzheimer's Disease using language models to mimic functional deficits observed in AD patients. |
| Outcome: | The proposed models will mimic the functional deficits observed in AD patients and evaluate their effects on brain score against the state-of-the-art models. |
Aligned Probing: Relating Toxic Behavior and Model Internals (2026.tacl-1)
Copied to clipboard
| Challenge: | Language models (LMs) may produce toxic text that contains hate speech, insults, or vulgarity, even when prompted with innocuous text. |
| Approach: | They propose an interpretability framework that aligns the behavior of language models based on their outputs and internal representations. |
| Outcome: | The proposed framework bridges behavioral and internal perspectives for toxicity for the first time. |
Language Models Understand Us, Poorly (2022.findings-emnlp)
Copied to clipboard
| Challenge: | Recent large language models have achieved impressive results on benchmark tasks. |
| Approach: | They examine three views of human language understanding: as-mapping, as-reliability and as-representation. |
| Outcome: | The authors argue that language models are inadequate and that they can't understand us . they also argue that as-representation advances a science of understanding . |
Monotonic Representation of Numeric Attributes in Language Models (2024.acl-short)
Copied to clipboard
| Challenge: | Language models (LMs) can express factual knowledge involving numeric properties such as Karl Popper was born in 1902, but how this information is encoded in the model’s internal representations is not understood well. |
| Approach: | They propose a method for finding and editing representations of numeric properties such as Karl Popper’s birth year. |
| Outcome: | The proposed method can express an increasingly late birthyear by patching activations along a “birthyear” direction. |
Compositional Generalization in Grounded Language Learning via Induced Model Sparsity (2022.naacl-srw)
Copied to clipboard
| Challenge: | induced model sparsity can help achieve compositional generalization and sample efficiency in grounded language learning problems. |
| Approach: | They propose a model that encourages sparse correlations between words and attributes to find a goal in a language-conditioned navigation problem with disentangled observations. |
| Outcome: | The proposed agent maintains high performance even when learning from a handful of demonstrations. |
Auditing Language Model Unlearning via Information Decomposition (2026.eacl-long)
Copied to clipboard
| Challenge: | Existing approaches to unlearning in language models do not account for information about forgotten data . despite the apparent success of unlearning, information about the forgotten data remains linearly decodable from internal representations. |
| Approach: | They propose an interpretable framework for auditing unlearning using Partial Information Decomposition . they propose a representation-based risk score that can guide abstention on sensitive inputs . |
| Outcome: | The proposed framework can guide abstention on sensitive inputs at inference time. |
Probing the Geometry of Truth: Consistency and Generalization of Truth Directions in LLMs Across Logical Transformations and Question Answering Tasks (2025.findings-acl)
Copied to clipboard
| Challenge: | Large language models (LLMs) are trained on vast corpora that contain substantial knowledge but their outputs often contain confidently stated inaccuracies. |
| Approach: | They propose to encode truthfulness as a distinct linear feature, termed the "truth direction", which can classify truthfulness reliably. |
| Outcome: | The proposed model can generalize to logical transformations, question-answering tasks, in-context learning, and external knowledge sources. |
Pref-CTRL: Preference Driven LLM Alignment using Representation Editing (2026.acl-short)
Copied to clipboard
| Challenge: | Recent work suggests that test-time alignment methods are inefficient because they require a large number of computational resources. |
| Approach: | They propose a preference-based training framework that uses a multi-objective value function to better reflect the structure of preference data. |
| Outcome: | The proposed framework outperforms RE-Control and shows greater generalization on out-of-domain datasets. |
Streamlining LLMs: Adaptive Knowledge Distillation for Tailored Language Models (2025.naacl-srw)
Copied to clipboard
| Challenge: | Large language models (LLMs) have transformative potential across industries, e.g., enhancing customer service, revolutionizing medical diagnostics, or identifying crises in news articles. |
| Approach: | They propose to distill compact, parameter-efficient tailored language models from LLMs for domain-specific tasks with comparable performance. |
| Outcome: | The proposed framework outperforms knowledge distillation frameworks in the crisis domain, where labeled data is scarce. |
How do Multimodal Foundation Models Encode Text and Speech? An Analysis of Cross-Lingual and Cross-Modal Representations (2025.naacl-short)
Copied to clipboard
| Challenge: | Recent advances in foundation models have sparked growing interest in expanding their text processing capabilities to speech. |
| Approach: | They analyze the model activations from semantically equivalent sentences across languages in the text and speech modalities and examine how text and spoken are represented in recent multimodal foundation models. |
| Outcome: | The proposed models exhibit cross-lingual differences, but are not explicitly trained for modality-agnostic representations. |
Beyond Detection: Evaluating Fallacy Awareness of LLMs in Interactive Scenarios (2026.acl-long)
Copied to clipboard
| Challenge: | Large Language Models fail to recognize fallacious reasoning in real-world interactions despite strong performance on static fallacy detection tasks. |
| Approach: | They propose a Chinese benchmark to assess fallacy awareness without explicit cues . they propose 'fate' evaluation framework that assesses fallacy without explicit . |
| Outcome: | The proposed framework assesses fallacy awareness without explicit cues, combining natural dialogue responses and reasoning-based decisions. |
When Will the Tokens End? Graph-Based Forecasting for LLMs Output Length (2025.acl-srw)
Copied to clipboard
Grzegorz Piotrowski, Mateusz Bystroński, Mikołaj Hołysz, Jakub Binkowski, Grzegorz Chodak, Tomasz Jan Kajdanowicz
| Challenge: | Large Language Models (LLMs) are typically trained to predict the next token in a sequence. However, their internal representations encode signals that go beyond immediate next-token prediction. |
| Approach: | They propose an aggregation-based model that combines hidden states from multiple transformer layers l 8, dots, 15 using element-wise operations such as mean or sum. |
| Outcome: | The proposed model reduces NMAE by over 50% on the Alpaca dataset. |
A Shared Geometry of Difficulty in Multilingual Language Models (2026.acl-short)
Copied to clipboard
| Challenge: | Large language models encode problem difficulty as an internal signal that can be linearly decoded from their residuals. |
| Approach: | They train linear probes on the AMC subset of the Easy2Hard benchmark, translated into 21 languages, and found difficulty-related signals emerge at two distinct stages of the model internals. |
| Outcome: | The results show that difficulty-related signals emerge at two distinct stages of the model internals, corresponding to shallow (early-layers) and deep (later-layer) representations, that exhibit functionally different behaviors. |
Explaining Language Model Predictions with High-Impact Concepts (2024.findings-eacl)
Copied to clipboard
| Challenge: | Existing methods to explain large language models (LLMs) are mostly correlational and lack causal features due to compositional nature of languages. |
| Approach: | They propose a framework to provide impact-aware explanations for large language models that are robust to feature changes and influential to the model’s predictions. |
| Outcome: | The proposed explanations improve on real and synthetic tasks and are robust to feature changes and influential to the model’s predictions. |
Visual Grounding Helps Learn Word Meanings in Low-Data Regimes (2024.naacl-long)
Copied to clipboard
| Challenge: | Modern neural language models (LMs) require distinctly un-human-like ways to achieve these results. |
| Approach: | They train a diverse set of LM architectures with and without auxiliary visual supervision on datasets of varying scales. |
| Outcome: | The proposed models exhibit better learning of syntactic categories, lexical relations, semantic features, word similarity and alignment with human neural representations. |
Mitigating Hallucinations in Multi-modal Large Language Models via Image Token Attention-Guided Decoding (2025.naacl-long)
Copied to clipboard
| Challenge: | Multi-modal large language models (MLLMs) generate plausible but incorrect content, resulting in hallucinations . recent advances in MLLM technology have demonstrated their outstanding performance in a variety of visual tasks, such as object detection. |
| Approach: | They propose a plug-and-play method which leverages MLLMs’ internal representations to mitigate hallucinations by analyzing input and output tokens. |
| Outcome: | The proposed method exploits MLLMs’ internal representations to mitigate hallucinations. |
Tensor Product Generation Networks for Deep NLP Modeling (N18-1)
Copied to clipboard
| Challenge: | Using Tensor Product Representations (TPRs) we propose a new architecture for natural language processing based on the principle that hypothesis space for learning includes network hypotheses that are independently known to be suitable for performing the target task. |
| Approach: | They propose a Tensor Product Generation Network (TPGN) which is capable of carrying out TPR computation but uses unconstrained deep learning to design its internal representations. |
| Outcome: | The proposed architecture outperforms baselines on the COCO dataset and can interpret internal representations and operations. |
The Model Agreed, But Didn’t Learn: Diagnosing Surface Compliance in Large Language Models (2026.findings-acl)
Copied to clipboard
| Challenge: | Large Language Models internalize vast world knowledge as parametric memory, yet inherit the staleness and errors of their source corpora. |
| Approach: | They propose a framework that subjects models to discriminative self-assessment under diverse contextual pressures to scrutinize subtle behavioral nuances induced by memory modifications. |
| Outcome: | The proposed framework achieves high benchmarks without overwriting internal beliefs, while recursive modifications accumulate representational residues, triggering cognitive instability and permanently diminishing the reversibility of the model’s memory state. |
Can Activation Steering Generalize Across Languages? A Study on Syllogistic Reasoning in Language Models (2026.eacl-long)
Copied to clipboard
| Challenge: | Prior work has focused on activation steering for Large Language Models (LLMs) this technique can be used to improve reasoning accuracy and transferability across languages. |
| Approach: | They propose to use activation steering to steer models towards a cross-lingual reasoning space. |
| Outcome: | The proposed techniques generalise well to multilingual datasets while minimizing language modelling performance. |
Unveiling Decision-Making in LLMs for Text Classification : Extraction of influential and interpretable concepts with Sparse Autoencoders (2026.findings-eacl)
Copied to clipboard
| Challenge: | Concept-based explanations for large language models are not well understood in text classification. |
| Approach: | They propose a model with a specialized classifier head and activation rate sparsity loss for sentence classification . they compare it to existing models with HI-Concept and ConceptShap . |
| Outcome: | The proposed model improves both the causality and interpretability of the extracted features. |
Active Layer-Contrastive Decoding Reduces Hallucination in Large Language Model Generation (2025.emnlp-main)
Copied to clipboard
| Challenge: | Recent decoding methods improve the factuality of large language models (LLMs) by refining how the next token is selected during generation. |
| Approach: | They propose a decoding strategy that actively decides when to apply contrasting layers during generation by casting decoding as a sequential decision-making problem. |
| Outcome: | The proposed method surpasses state-of-the-art methods across five benchmarks and mitigates hallucinations in diverse generation scenarios. |
Multi-component Causal Tracing in Large Language Models (2026.acl-long)
Copied to clipboard
| Challenge: | Large language models (LLMs) are prone to various forms of safety risks, such as learning and propagating societal biases and even creating harmful or deceptive content through jailbreak attacks. |
| Approach: | They propose a framework for causally tracing multiple components simultaneously that systematically identifies the subsets of components most critical to a desired performance metric. |
| Outcome: | The proposed method outperforms existing methods in identifying components critical to a desired performance metric. |
How Gender Debiasing Affects Internal Model Representations, and Why It Matters (2022.naacl-main)
Copied to clipboard
| Challenge: | Existing studies of gender bias in NLP focus on extrinsic or intrinsic bias, but the relationship between extrindic and intrinsic bias is relatively unknown. |
| Approach: | They propose a framework to measure extrinsic and intrinsic bias together and propose metric to measure debiasing and intrinsic debiases. |
| Outcome: | The proposed framework provides a comprehensive perspective on bias in NLP models, which can be applied to deploy NLP systems in a more informed manner. |
Towards Integration of Discriminability and Robustness for Document-Level Relation Extraction (2023.eacl-main)
Copied to clipboard
| Challenge: | Document-level relation extraction (DocRE) predicts relations for entity pairs relying on context-dependent reasoning . a large number of annotation errors can make it difficult to distinguish large semantically close relations . |
| Approach: | They propose a loss function to improve discriminability and robustness for DocRE . they also propose supervised contrastive learning and negative label sampling strategy . |
| Outcome: | The proposed method achieves state-of-the-art results on the DocRED dataset and its recently cleaned version. |
Silencing the Guardrails: Inference-Time Jailbreaking via Dynamic Contextual Representation Ablation (2026.findings-acl)
Copied to clipboard
| Challenge: | Existing strategies to circumvent safety constraints face significant trade-offs between effectiveness and efficiency. |
| Approach: | They propose a framework that allows to infer model refusal behaviors without expensive parameter updates or training. |
| Outcome: | The proposed framework outperforms baselines in multiple safety-aligned open-source LLMs. |
FlashMem: Distilling Intrinsic Latent Memory via Computation Reuse (2026.findings-acl)
Copied to clipboard
| Challenge: | Large Language Models lack the mechanism to preserve dynamic context, forcing agents to redundantly reprocess history to maintain long-horizon autonomy. |
| Approach: | They propose a framework that distills intrinsic memory directly from transient reasoning states via computation reuse. |
| Outcome: | Experiments show that FlashMem matches heavy baselines while reducing inference latency by 5 times, effectively bridging the gap between efficiency and persistent cognition. |
Do Androids Know They’re Only Dreaming of Electric Sheep? (2024.findings-acl)
Copied to clipboard
| Challenge: | Detecting hallucinations in grounded generation tasks is commonly framed as a textual entailment problem. |
| Approach: | They develop probes that are narrowly trained to predict hallucination in a transformer language model. |
| Outcome: | The probes can detect hallucinations at many transformer layers outperforming baselines and human annotators on two out of three generation tasks. |
Neural Breadcrumbs: Membership Inference Attacks on LLMs Through Hidden State and Attention Pattern Analysis (2026.eacl-long)
Copied to clipboard
| Challenge: | Membership inference attacks (MIAs) reveal whether specific data was used to train machine learning models, serving as important tools for privacy auditing and compliance assessment. |
| Approach: | They propose to examine LLMs’ internal representations rather than just their outputs to gain additional insights into potential membership inference signals. |
| Outcome: | The proposed framework yields strong membership detection across several model families achieving average AUC scores of 0.85 on popular MIA benchmarks. |
Bridging the Language Gaps in Large Language Models with Inference-Time Cross-Lingual Intervention (2025.acl-long)
Copied to clipboard
| Challenge: | Existing approaches to address performance gaps in LLMs rely on pretraining or fine-tuning, which are resource-intensive. |
| Approach: | They propose a framework that aligns LLMs' internal representations with those of high-performing languages during inference. |
| Outcome: | The proposed framework improves performance on low-performing (source) languages by aligning their internal representations with those of high-performing languages during inference. |
Mary, the Cheeseburger-Eating Vegetarian: Do LLMs Recognize Incoherence in Narratives? (2026.eacl-long)
Copied to clipboard
Karin De Langis, Püren Öncel, Ryan Peters, Andrew Elfenbein, Laura Kristen Allen, Andreas Schramm, Dongyeop Kang
| Challenge: | Contemporary models of (human) reading comprehension characterize comprehension as a dynamic process in which the reader continually builds and updates representations to maintain coherence and integrate new information with prior knowledge. |
| Approach: | They use a paired narrative dataset to examine the extent to which large language models can reliably separate incoherent and coherent stories. |
| Outcome: | The proposed models do not eliminate the deficits in the model internal state and behavior. |
HALP: Detecting Hallucinations in Vision-Language Models without Generating a Single Token (2026.eacl-long)
Copied to clipboard
| Challenge: | Existing methods for detection of hallucinations operate after text generation, making intervention costly and untimely. |
| Approach: | They examine whether hallucination risk can instead be predicted before any token is generated by probing a model's internal representations in a single forward pass. |
| Outcome: | The proposed model can detect hallucinations before token generation, while query-token representations can be more accurate. |
Beyond Words: Integrating Theory of Mind into Conversational Agents for Human-Like Belief, Desire, and Intention Alignment (2025.findings-acl)
Copied to clipboard
| Challenge: | Empirical evaluations of LLaMA-3 models demonstrate that ToM-informed alignment improves response quality, achieving win rates of 63% and 67%, respectively. |
| Approach: | They investigate whether open-source LLaMA models can represent and retain ToM-related constructs and whether they can be used to generate more aligned responses. |
| Outcome: | The proposed models can represent and retain ToM-related constructs and improve response quality. |
Correlations between Multilingual Language Model Geometry and Crosslingual Transfer Performance (2024.lrec-main)
Copied to clipboard
| Challenge: | Pre-trained multilingual language models represent multiple languages in a single vector space, a feature hypothesized to enable impressive crosslingual transfer capabilities. |
| Approach: | They propose to use a multilingual representation space that sorts axes based on their language-separability to determine whether geometric distances between languages correlate with crosslingual transfer performance. |
| Outcome: | The proposed measures do not generalize well across models, layers, and tasks. |
Understanding the Inner-workings of Language Models Through Representation Dissimilarity (2023.emnlp-main)
Copied to clipboard
| Challenge: | Dissimilarity measures measure the extent to which two model’s internal representations differ . they can identify and locate generalization properties of models that are invisible via in-distribution test set performance. |
| Approach: | They propose to use representation dissimilarity measures to measure the extent to which two model’s internal representations differ. |
| Outcome: | The proposed dissimilarity measures can identify and locate generalization properties of models that are invisible via in-distribution test set performance and new evaluations of how language model features vary as width and depth are increased. |
Similarity Analysis of Contextual Word Representation Models (2020.acl-main)
Copied to clipboard
| Challenge: | Existing and novel similarity measures are used to analyze contextual word representations . different architectures have rather similar representations, but different individual neurons. |
| Approach: | They propose a method to analyze contextual word representation models using similarity analysis. |
| Outcome: | The proposed approach can be used to analyze model similarity without external annotations. |
How does BERT’s attention change when you fine-tune? An analysis methodology and a case study in negation scope (2020.acl-main)
Copied to clipboard
| Challenge: | Recent work probing pre-trained language models for downstream tasks is difficult to explain . a growing body of research is devoted to understanding what linguistic properties these language models have acquired. |
| Approach: | They propose a procedure and analysis method that takes a hypothesis of how a transformer-based model might encode a linguistic phenomenon and tests its validity. |
| Outcome: | The proposed method tests a hypothesis that some attention heads will consistently attend from a word in negation scope to the negation cue. |
Unlocking the Future: Exploring Look-Ahead Planning Mechanistic Interpretability in Large Language Models (2024.emnlp-main)
Copied to clipboard
| Challenge: | Recent studies have shown that large language models may possess preliminary planning capabilities. |
| Approach: | They examine the look-ahead planning mechanism in large language models from the perspectives of information flow and internal representations. |
| Outcome: | The proposed model can decode the decision from the output of MHSA in the middle layers at the last token. |
How Language Models Conflate Logical Validity with Plausibility: A Representational Analysis of Content Effects (2026.findings-acl)
Copied to clipboard
| Challenge: | a number of theories have been proposed to account for content effects in large language models, including the dual-process theory of reasoning, but the mechanisms behind content effects remain unclear. |
| Approach: | They propose to encode validity and plausibility concepts in LLMs by aligning them in representational geometry. |
| Outcome: | The proposed model conflates validity and plausibility, and vice versa. |
Model Selection for Cross-lingual Transfer (2021.emnlp-main)
Copied to clipboard
| Challenge: | Existing work has relied on English dev data to select among models that are fine-tuned with different learning rates, number of steps and other hyperparameters, often resulting in suboptimal choices. |
| Approach: | They propose a machine learning approach that uses the fine-tuned model’s internal representations to predict its cross-lingual capabilities. |
| Outcome: | The proposed model selects better than English validation data across twenty five languages, including eight low-resource languages, and often achieves comparable results to model selection using target language development data. |
SGT: Securing Open-Source LLMs Against Malicious Fine-tuning via Safety Guidance Trigger (2026.acl-long)
Copied to clipboard
| Challenge: | open-weight large language models increase vulnerability to malicious fine-tuning . despite these advantages, open-source LLMs increase the potential for misuse . |
| Approach: | They propose a safety guide for open-weight large language models that guides fine-tuning toward the safety manifold to preserve alignment. |
| Outcome: | The proposed safety guidance trigger significantly improves robustness against malicious fine-tuning. |
Finding Universal Grammatical Relations in Multilingual BERT (2020.acl-main)
Copied to clipboard
| Challenge: | Recent work has found that multilingual masked language models learn a surprising amount of linguistic structure, despite a lack of direct linguistic supervision. |
| Approach: | They propose an unsupervised method to find syntactic tree distances in languages other than English and that these subspaces are approximately shared across languages. |
| Outcome: | The proposed method shows that mBERT learns representations of syntactic dependency labels, in the form of clusters, which largely agree with the Universal Dependencies taxonomy. |
Steering LVLMs via Sparse Autoencoder for Hallucination Mitigation (2025.findings-emnlp)
Copied to clipboard
| Challenge: | Existing approaches to address hallucinations in large vision-language models require substantial computational cost and time. |
| Approach: | They propose to leverage sparse autoencoders to identify semantic directions closely associated with faithfulness or hallucination, extracting more precise and disentangled hallucinian-related representations. |
| Outcome: | The proposed method outperforms existing decoding approaches while maintaining transferability across different model architectures with negligible additional time overhead. |
NER Retriever: Zero-Shot Named Entity Retrieval with Type-Aware Embeddings (2025.findings-emnlp)
Copied to clipboard
| Challenge: | NER Retriever uses a user-defined type description to retrieve documents mentioning entities of that type. |
| Approach: | They propose a zero-shot retrieval framework for ad-hoc Named Entity Recognition . a user-defined type description is used to retrieve documents mentioning entities of that type . |
| Outcome: | The proposed framework outperforms lexical and dense retrieval baselines on three benchmarks. |
Implicit Representations of Grammaticality in Language Models (2026.acl-long)
Copied to clipboard
| Challenge: | Pretrained language models generate grammatically well-formed text and discriminate well between grammatical and ungrammatically sentences in tightly controlled minimal pairs. |
| Approach: | They propose a method to train pretrained LMs for representations of grammaticality by applying perturbations to a naturalistic text corpus. |
| Outcome: | The proposed model outperforms probability-based models on human-curated grammaticality judgment benchmarks and performs worse than string probabilities on plausibility benchmarks. |
How a Bilingual LM Becomes Bilingual: Tracing Internal Representations with Sparse Autoencoders (2025.findings-emnlp)
Copied to clipboard
Tatsuro Inaba, Go Kamoda, Kentaro Inui, Masaru Isonuma, Yusuke Miyao, Yohei Oseki, Yu Takagi, Benjamin Heinzerling
| Challenge: | Using sparse autoencoders, we explore how bilingual language models develop complex internal representations. |
| Approach: | They employ sparse autoencoders to analyze bilingual language models' internal representations. |
| Outcome: | The proposed method integrates decomposed representations from a fully trained model into a mid-training model. |
An information theoretic view on selecting linguistic probes (2020.emnlp-main)
Copied to clipboard
| Challenge: | Recent advances in NLP tasks require a question of how much linguistic knowledge is encoded in neural networks. |
| Approach: | They propose to use diagnostic classifiers to perform supervised classification from internal representations. |
| Outcome: | Empirically, the two proposed criteria lead to results that agree with each other. |
Emergence of Abstract State Representations in Embodied Sequence Modeling (2023.emnlp-main)
Copied to clipboard
| Challenge: | Sequence modeling is a simple yet versatile task that can be applied to more complex decision-making domains. |
| Approach: | They build a sequence modeling Transformer which takes a language instruction, actions, and environmental observations as inputs and then trains a model to reconstruct environmental layouts. |
| Outcome: | The proposed model can reconstruct environmental layouts from the inputs of the model and language instructions play a role in the reconstruction accuracy. |
Strong hallucinations from negation and how to fix them (2024.findings-acl)
Copied to clipboard
| Challenge: | Despite great performance on many tasks, language models still struggle with reasoning, sometimes providing responses that cannot possibly be true because they stem from logical incoherence. |
| Approach: | They propose a way to treat negation as an operation over latent representations that constrains how they may evolve. |
| Outcome: | The proposed approach improves model performance in cloze prompting and natural language inference tasks without training on sparse negative data. |
How Retrieved Context Shapes Internal Representations in RAG (2026.findings-acl)
Copied to clipboard
| Challenge: | Retrieval-augmented generation (RAG) is a widely adopted approach for enhancing large language models with external knowledge. |
| Approach: | They analyze how different types of retrieved documents affect the hidden states of large language models and how these internal representation shifts relate to downstream generation behavior. |
| Outcome: | The results show that context relevancy and layer-wise processing influence internal representations, providing explanations of LLMs’ output behaviors and insights for RAG system design. |
Improving Cross-lingual Transfer with Contrastive Negative Learning and Self-training (2024.lrec-main)
Copied to clipboard
| Challenge: | Recent studies improve cross-lingual transfer learning by better aligning the internal representations within the multilingual model or exploring the information of the target language using self-training. |
| Approach: | They propose to use negative pairs to align the multilingual model and self-train the model to converge on the obtained clean pseudo-labels. |
| Outcome: | The proposed method improves upon the baseline models and can serve as a beneficial complement to the alignment-based methods. |
Multi-Source Probing for Open-Domain Conversational Understanding (2023.emnlp-main)
Copied to clipboard
| Challenge: | Existing models of open-domain dialogue comprehension have limited conversational understanding and response generation. |
| Approach: | They propose a multi-source probing method to probe dialogue comprehension abilities of open-domain dialogue models. |
| Outcome: | The proposed method aggregates features from multiple sources to accomplish diverse task goals and conducts downstream tasks in a generative manner consistent with dialogue model pre-training to leverage model capabilities. |
MULTIGUARD: An Efficient Approach for AI Safety Moderation Across Languages and Modalities (2025.emnlp-main)
Copied to clipboard
Sahil Verma, Keegan Hines, Jeff Bilmes, Charlotte Siska, Luke Zettlemoyer, Hila Gonen, Chandan Singh
| Challenge: | Existing approaches to detect harmful queries to large language models are fallible and vulnerable to attacks that exploit mismatched generalization of model capabilities. |
| Approach: | They propose an approach to detect harmful queries to large language models (LLMs) OMNIGUARD identifies internal representations of an LLM/MLLM that are aligned across languages or modalities and builds a language-agnostic or modality-adic classifier for detecting harmful prompts. |
| Outcome: | OMNIGUARD improves harmful prompt classification accuracy by 11.57% over the strongest baseline in a multilingual setting, by 20.44% for image-based prompts, and sets a new SOTA for audio-based ones. |
Adapting Where It Matters: Depth-Aware Adaptation for Efficient Multilingual Speech Recognition in Low-Resource Languages (2026.findings-acl)
Copied to clipboard
| Challenge: | Recent speech foundation models excel at multilingual automatic speech recognition (ASR) for high-resource languages, but their performance drops substantially on low-resourced languages due to the limited data availability. |
| Approach: | They propose a Depth-Aware Model Adaptation framework that allocates adaptation capacity according to each layer’s role. |
| Outcome: | The proposed framework matches or surpasses state-of-the-art accuracy with 80% fewer trainable parameters and achieves 29% error reduction under extreme data scarcity. |
RECALL: REpresentation-aligned Catastrophic-forgetting ALLeviation via Hierarchical Model Merging (2025.emnlp-main)
Copied to clipboard
Bowen Wang, Haiyuan Wan, Liwen Shi, Chen Yang, Peng He, Yue Ma, Haochen Han, Wenhao Li, Tiao Tan, Yongjian Li, Fangming Liu, Gong Yifan, Sheng Zhang
| Challenge: | Existing models that require task labels or performance trade-offs are susceptible to catastrophic forgetting. |
| Approach: | They propose a representation-aware model merging framework for continual learning without access to historical data. |
| Outcome: | The proposed framework outperforms baselines in knowledge retention and generalization across five NLP tasks and multiple continual learning scenarios. |
AGD: Adversarial Game Defense Against Jailbreak Attacks in Large Language Models (2025.acl-long)
Copied to clipboard
Shilong Pan, Zhiliang Tian, Zhen Huang, Wanlong Yu, Zhihua Wen, Xinwang Liu, Kai Lu, Minlie Huang, Dongsheng Li
| Challenge: | Existing defenses, including post-training alignment and prompt engineering, struggle with adaptability to out-of-distribution (OOD) attacks. |
| Approach: | They propose an adversarial game-based defense method that dynamically adjusts LLMs’ internal representations to achieve a balanced trade-off between helpfulness and harmlessness. |
| Outcome: | The proposed method improves LLMs’ safety over all baselines. |
FineSteer: A Unified Framework for Fine-Grained Inference-Time Steering in Large Language Models (2026.acl-long)
Copied to clipboard
| Challenge: | Existing methods for inference-time steering fail to be effective, utility-preserving and training-efficient due to rigid, one-size-fits-all designs and limited adaptability. |
| Approach: | They propose a steering framework that decomposes inference-time steering into two stages . they propose 'conditional steering' mechanism that preserves model utility by avoiding unnecessary steering . a 'mixture-of-Steering-Experts' mechanism captures multimodal nature of desired steering behaviors . |
| Outcome: | The proposed framework outperforms the state-of-the-art methods on safety and truthfulness benchmarks. |
DP-BART for Privatized Text Rewriting under Local Differential Privacy (2023.findings-acl)
Copied to clipboard
| Challenge: | Existing systems that guarantee privacy protection to individuals face several problems, such as formal mathematical flaws, unrealistic privacy guarantees, and privatization of only individual words. |
| Approach: | They propose a system that enables sharing of sensitive textual documents while formally guaranteeing privacy protection to individuals. |
| Outcome: | The proposed system outperforms existing LDP systems in terms of noise and privacy guarantees. |
ReFL: Reflective Feedback Learning for Hallucination Detection of Large Language Models (2026.acl-long)
Copied to clipboard
| Challenge: | Existing methods for detecting hallucinations depend on external knowledge sources, incurring high computational costs and limiting real-time applicability, or extract the model’s internal states, leading to poor generalization. |
| Approach: | They propose a hallucination detection framework that leverages corrective in-context learning to guide LLMs to recognize their own prediction errors and adjust internal representations, critically without updating model weights. |
| Outcome: | The proposed framework outperforms existing methods on two benchmark datasets and achieves state-of-the-art performance. |
Segment-Based Attention Masking for GPTs (2025.acl-long)
Copied to clipboard
| Challenge: | Causal masking is a fundamental component in Generative Pre-Trained Transformers (GPT) models, playing a crucial role during training. |
| Approach: | They propose to apply causal masking to all input tokens step-by-step, mimicking the generation process. |
| Outcome: | The proposed model can process the entire user prompt at once, but it is applied to all input tokens step-by-step, mimicking the generation process. |
Linear Steerability in Language Models: When It Emerges and How It Evolves (2025.findings-emnlp)
Copied to clipboard
| Challenge: | a new framework for steering language models reveals how concepts become linearly separable as training progresses . |
| Approach: | They propose a framework to analyze steerability in language models by using hidden state and representation analysis. |
| Outcome: | The proposed framework reveals how steerability evolves over training . concepts become linearly separable as training progresses, the framework shows . |
Rethinking Jailbreak Detection of Large Vision Language Models with Representational Contrastive Scoring (2026.acl-long)
Copied to clipboard
| Challenge: | Large Vision-Language Models (LVLMs) are vulnerable to a growing array of multimodal jailbreak attacks, necessitating a generalizable defense that is efficient for practical deployment. |
| Approach: | They propose a framework that uses a lightweight projection to separate benign and malicious inputs in safety-critical layers. |
| Outcome: | The proposed framework enables a simple yet powerful contrastive score that differentiates true malicious intent from mere distribution shift. |
On the Universal Truthfulness Hyperplane Inside LLMs (2024.emnlp-main)
Copied to clipboard
| Challenge: | Recent studies have explored hallucinations through the lens of internal representations, proposing mechanisms to decipher LLMs’ adherence to facts. |
| Approach: | They propose to train a universal truthfulness hyperplane that distinguishes the model’s factually correct and incorrect outputs on a diverse collection of over 40 datasets and examine its cross-task, cross-domain, and in-domain generalization. |
| Outcome: | The proposed model is able to distinguish factual outputs from incorrect outputs on a diverse collection of over 40 datasets. |
Reading Between the Prompts: How Stereotypes Shape LLM’s Implicit Personalization (2025.emnlp-main)
Copied to clipboard
| Challenge: | Prior work has shown that such inferences can lead to lower quality responses for users assumed to be from minority groups. |
| Approach: | They analyze LLMs' latent user representations through both model internals and generated answers to targeted user questions. |
| Outcome: | The proposed models infer demographic attributes based on stereotypical signals, which persists even when the user explicitly identifies with a different demographic group. |
Memorization ≠ Understanding: Do Large Language Models Have the Ability of Scenario Cognition? (2025.emnlp-main)
Copied to clipboard
| Challenge: | Large language models (LLMs) have demonstrated impressive performance across NLP tasks. |
| Approach: | They propose a framework to assess LLMs’ scenario cognition . they examine the ability to link semantic scenario elements with their arguments in context . |
| Outcome: | The proposed framework assesses large language models’ scenario cognition . it shows that current models rely on superficial memorization, failing to achieve robust semantic scenario cognition even in simple cases. |
Length Representations in Large Language Models (2025.findings-emnlp)
Copied to clipboard
| Challenge: | Large language models (LLMs) have shown remarkable capabilities across various tasks, that are learned from massive amounts of text-based data. |
| Approach: | They propose to scale hidden units within the model to control output sequence length without losing the informativeness of the generated text. |
| Outcome: | The output sequence length is controlled by multiple head attention mechanisms, which can be adjusted in a disentangled manner. |
Keys to Robust Edits: From Theoretical Insights to Practical Advances (2025.acl-long)
Copied to clipboard
| Challenge: | Existing methods for modifying parametric memory are prone to inaccuracies due to conflicting or outdated information. |
| Approach: | They propose a plug-and-play module that disentangles editing keys from native model representations and dynamically adjusts keys via contrastive learning to achieve robustness-specificity balance. |
| Outcome: | The proposed method improves over robustness tests by up to 66.4% while maintaining the success rate unaffected. |
Looking Beyond the One: Operationalizing and Eliciting Visual Ambiguity in VLLMs (2026.acl-long)
Copied to clipboard
| Challenge: | Visual question answering systems typically collapse ambiguity, committing to a single interpretation during decoding and evaluation. |
| Approach: | They operationalize ambiguity as the existence of multiple answer-supporting regions in an image . they show that ambiguities are already encoded in their internal representations . |
| Outcome: | The proposed approach makes ambiguity observable without exhaustive annotations . ambiguities are already encoded in models, but not reliably expressed in outputs despite hidden states . |
Unsupervised Hallucination Detection by Inspecting Reasoning Processes (2025.emnlp-main)
Copied to clipboard
| Challenge: | Unsupervised hallucination detection aims to identify hallucines generated by large language models without relying on labeled data. |
| Approach: | They propose an unsupervised method to detect hallucinated content by large language models . they use internal representations intrinsic to factual correctness to prompt the model to verify the truthfulness of a given statement . |
| Outcome: | The proposed framework outperforms existing unsupervised methods and is fully unsupervised and low cost. |
Beyond Prompt Engineering: Robust Behavior Control in LLMs via Steering Target Atoms (2025.acl-long)
Copied to clipboard
| Challenge: | Recent research has explored the use of sparse autoencoders (SAE) to disentangle knowledge in high-dimensional spaces for steering. |
| Approach: | They propose a method that isolates and manipulates disentangled knowledge components to enhance safety by using sparse autoencoders to disentangle knowledge in high-dimensional spaces for steering. |
| Outcome: | The proposed method is able to isolate and manipulate disentangled knowledge components to enhance safety in large reasoning models. |
Two Pathways to Truthfulness: On the Intrinsic Encoding of LLM Hallucinations (2026.acl-long)
Copied to clipboard
Wen Luo, Guangyue Peng, Wei Li, Shaohang Wei, Feifan Song, Liang Wang, Nan Yang, Xingxing Zhang, Jing Jin, Furu Wei, Houfeng Wang
| Challenge: | Previous work shows that large language models generate hallucinations, yet the origins and mechanisms of these signals remain unclear. |
| Approach: | They propose to validate and disentangle two different pathways for truthfulness cues . they also propose to use the same mechanism to derive self-contained evidence from the generated answer . |
| Outcome: | The proposed applications improve hallucination detection performance by integrating two different inputs. |
Analyzing LLMs’ Knowledge Boundary Cognition Across Languages Through the Lens of Internal Representations (2025.acl-long)
Copied to clipboard
| Challenge: | Understanding the knowledge boundaries of Large Language Models (LLMs) is crucial to prevent hallucination, but research on the knowledge boundary perceptions of LLMs has predominantly focused on English. |
| Approach: | They propose a training-free alignment method that effectively transfers knowledge boundary perception ability across languages, thereby helping reduce hallucination risk in low-resource languages. |
| Outcome: | The proposed method reduces hallucination risk in low-resource languages by fine-tuning on bilingual question pair translation. |
Compiling Activation Steering into Weights via Null-Space Constraints for Stealthy Backdoors (2026.acl-long)
Copied to clipboard
Rui Yin, Tianxu Han, Naen Xu, Changjiang Li, Ping He, Chunyi Zhou, Jun Wang, Zhihui Fu, Tianyu Du, Jinbao Li, Shouling Ji
| Challenge: | Existing methods to inject safety-aligned large language models rely on token-level mappings, which do not guarantee sustained harmful output. |
| Approach: | They propose a method that directly modifies model weights to map a trigger to an attacker-specified response. |
| Outcome: | The proposed method achieves high triggered attack success while maintaining non-triggered safety and general utility. |
Knowledge Graph Enhanced Large Language Model Editing (2024.emnlp-main)
Copied to clipboard
| Challenge: | Existing methods for editing large language models struggle to track and incorporate changes in knowledge associated with edits, which limits the generalization ability of post-edit LLMs in processing edited knowledge. |
| Approach: | They propose a model editing method that leverages knowledge graphs to enhance LLM editing by capturing changes in associated knowledge by constructing an external graph. |
| Outcome: | The proposed method improves the generalization ability of LLMs in processing edited knowledge. |
SafeKey: Amplifying Aha-Moment Insights for Safety Reasoning (2025.emnlp-main)
Copied to clipboard
| Challenge: | Large Reasoning Models (LRMs) introduce a new paradigm of explicitly reasoning before answering, but they pose great safety risks against harmful queries and adversarial attacks. |
| Approach: | They propose a safety aha moment that activates safety reasoning and leads to a safe response. |
| Outcome: | The proposed model can generalize to unseen jailbreak prompts while maintaining general abilities. |
False Sense of Security: Why Probing-based Malicious Input Detection Fails to Generalize (2026.findings-acl)
Copied to clipboard
| Challenge: | Recent work has leveraged probing-based approaches to study the separability of malicious and benign inputs in Large Language Models’ internal representations. |
| Approach: | They propose to use probing-based methods to study separability of malicious and benign inputs in LLMs' internal representations to detect harmful and benign content. |
| Outcome: | The proposed methods show that they learn superficial patterns rather than semantic harmfulness. |
Understanding and Mitigating Political Stance Cross-topic Generalization in Large Language Models (2026.acl-long)
Copied to clipboard
| Challenge: | Recent studies have focused on the internal representations of large language models and the mechanisms that lead to unintended cross-topic generalization. |
| Approach: | They propose a method that uses inhibition to localize political neurons and a technique that uses topic-specific blocking to mitigate the cross-topic generalization. |
| Outcome: | The proposed method reduces cross-topic generalization by 20% while preserving topic-specific performance. |
Fairness Evaluation and Inference Level Mitigation in LLMs (2026.findings-acl)
Copied to clipboard
| Challenge: | Large language models display undesirable behaviors embedded in their internal representations, undermining fairness, inconsistency drift, and the propagation of unwanted patterns during extended dialogues. |
| Approach: | They propose a pruning-based framework that detects context-aware neuron activations and applies adaptive masking to modulate their influence during generation. |
| Outcome: | The proposed framework detects context-aware neuron activations and applies adaptive masking to modulate their influence during generation. |
It’s Not a Walk in the Park! Challenges of Idiom Translation in Speech-to-text Systems (2025.acl-long)
Copied to clipboard
| Challenge: | idioms are defined as words with a figurative meaning not deducible from their individual components. |
| Approach: | They compare idiom translation as compared to conventional news translation in two languages . they compare MT and SLT systems with MT, Large Language Models and cascaded alternatives . |
| Outcome: | The proposed systems show better handling of idioms than standard news translation systems. |
Exploiting Contextual Knowledge in LLMs through 𝒱-usable Information based Layer Enhancement (2025.acl-long)
Copied to clipboard
| Challenge: | Existing approaches to enhance the context-faithfulness of Large Language Models (LLMs) ignore the fundamental mechanism of how contextual information is processed within LLMs’ internal states. |
| Approach: | They propose a method that enhances the utilization of contextual knowledge within LLMs’ internal representations by employing V-usable information analysis. |
| Outcome: | The proposed method improves context-faithfulness generation in Question-Answering tasks, particularly in scenarios involving unknown or conflicting contextual knowledge. |
Do LLMs Adhere to Label Definitions? Examining Their Receptivity to External Label Definitions (2025.emnlp-main)
Copied to clipboard
Seyedali Mohammadi, Bhaskara Hanuma Vedula, Hemank Lamba, Edward Raff, Ponnurangam Kumaraguru, Francis Ferraro, Manas Gaur
| Challenge: | Exact label definitions are considered as clues to disambiguate unclear labels, helping models perform their tasks more effectively. |
| Approach: | They conducted controlled experiments on multiple explanation benchmark datasets and label definition conditions using expert-curated, LLM-generated, perturbed, and swapped definitions. |
| Outcome: | The results suggest that models often default to internal representations, particularly in general tasks, while domain-specific tasks benefit more from explicit definitions. |
Activation Steering for Chain-of-Thought Compression (2026.findings-acl)
Copied to clipboard
| Challenge: | Large language models produce intermediate explanations, commonly referred to as chains of thought (CoTs), but the generated rationales are typically verbose, consuming many additional tokens, and thus degrading throughput and increasing inference energy consumption. |
| Approach: | They propose to generate concise reasoning traces by directly adjusting internal representations via activation steering. |
| Outcome: | The proposed method reduces generated token length by 69.4% across five reasoning benchmarks while maintaining accuracy. |
Learning Uncertainty from Sequential Internal Dispersion in Large Language Models (2026.acl-long)
Copied to clipboard
| Challenge: | Recent approaches to detect hallucinations depend on model internal states to estimate uncertainty, but they focus on last or mean tokens. |
| Approach: | They propose a supervised hallucination detection framework that leverages token-wise, layer-wise features derived from hidden states. |
| Outcome: | The proposed framework outperforms baseline models and avoids large training sets. |
SiLP: Enhancing Non-Dominant Language Capabilities with a Selective Bidirectional Language Projection Framework (2026.acl-long)
Copied to clipboard
| Challenge: | Existing methods to improve performance of large language models rely on additional training objectives or language-specific parameters. |
| Approach: | They propose a bidirectional language projection framework that enables efficient multilingual alignment and language shift using the intrinsic parameters. |
| Outcome: | The proposed framework improves performance of non-dominant languages and improves internal representations. |