论文检索

Abhirama Subramanyam Penamakuri, Navlika Singh, Piyush Arora, Anand Mishra

Large Vision-Language Models (L-VLMs) have demonstrated remarkable performance in various vision and language tasks, including Visual Question Answering (VQA). However, their high computational cost makes them impractical for resource-constrained settings and inference-heavy applications. In contrast, Small Vision-Language Models (S-VLMs) offer efficiency but suffer from a significant performance gap compared to their larger counterparts. In this work, we introduce the Model Parity Aligner (MPA), a novel framework designed to systematically improve S-VLMs by leveraging unlabeled images and effective knowledge transfer from L-VLMs. Instead of traditional knowledge distillation methods that rely on labeled training data, MPA employs a strategic parity-based approach that precisely identifies the knowledge disparities between S-VLMs and L-VLMs, and optimizes training by targeting only these disparities. We conduct extensive experiments on four diverse VQA benchmarks, namely TextVQA, ST-VQA, ChartQA, and OKVQA, each of which required specialized reasoning capabilities such as text recognition, chart interpretation, and commonsense and factual understanding. Our results demonstrate that MPA consistently enhances the performance of S-VLM on all benchmarks, reducing the performance gap while maintaining computational efficiency. We shall make our code and MPA-aligned models publicly available upon acceptance of this work.

Mengying Yuan, WenHao Wang, Zixuan Wang, Yujie Huang, Kangli Wei, Fei Li, Chong Teng, Donghong Ji

Natural Language Inference (NLI) is a fundamental task in natural language processing. While NLI has developed many subdirections such as sentence-level NLI, document-level NLI and cross-lingual NLI, Cross-Document Cross-Lingual NLI (CDCL-NLI) remains largely unexplored. In this paper, we propose a novel paradigm: CDCL-NLI, which extends traditional NLI capabilities to multi-document, multilingual scenarios. To support this task, we construct a high-quality CDCL-NLI dataset including 25,410 instances and spanning 26 languages.To address the limitations of previous methods on CDCL-NLI task, we further propose an innovative method that integrates RST-enhanced graph fusion with interpretability-aware prediction.Our approach leverages RST (Rhetorical Structure Theory) within heterogeneous graph neural networks for cross-document context modeling, and employs a structure-aware semantic alignment based on lexical chains for cross-lingual understanding. For NLI interpretability, we develop an EDU (Elementary Discourse Unit)-level attribution framework that produces extractive explanations.Extensive experiments demonstrate our approach’s superior performance, achieving significant improvements over both conventional NLI models as well as large language models.Our work sheds light on the study of NLI and will bring research interest on cross-document cross-lingual context understanding, hallucination elimination and interpretability inference.Our code and dataset are available at CDCL-NLI-link.

Hongyu Chen, Neele Falk, Michael Roth, Agnieszka Falenska

In NLP, language–gender associations are commonly grounded in the author’s gender identity, inferred from their language use. However, this identity-based framing risks reinforcing stereotypes and marginalizing individuals who do not conform to normative language–gender associations. To address this, we operationalize the language–gender association as a perceived gender expression of language, focusing on how such expression is externally interpreted by humans, independent of the author’s gender identity. We present the first dataset of itskind: 5,100 human annotations of perceived gendered style—human-written texts rated on a five-point scale from very feminine to verymasculine. While perception is inherently subjective, our analysis identifies textual features associated with higher agreement among annotators: formal expressions and lower emotional intensity. Moreover, annotator demographics influence their perception: women annotators are more likely to label texts as feminine, and men and non-binary annotators as masculine. Finally, feature analysis reveals that the text’s perceived gendered style is shaped by both affective and function words, partially overlapping with known patterns of language variation across gender identities. Our findings lay the groundwork for operationalizing gendered style through human annotation, while also highlighting annotators’ subjective judgments as meaningful signals to understand perception-based concepts.

Sabri Boughorbel, Fahim Dalvi, Nadir Durrani, Majd Hawasly

As fine-tuning becomes the dominant paradigm for improving large language models (LLMs), understanding what changes during this process is increasingly important. Traditional benchmarking often fails to explain _why_ one model outperforms another. In this work, we use model diffing, a mechanistic interpretability approach, to analyze the specific capability differences between Gemma-2-9b-it and a SimPO-enhanced variant. Using crosscoders, we identify and categorize latent representations that differentiate the two models. We find that SimPO acquired latent concepts predominantly enhance safety mechanisms (+32.8%), multilingual capabilities (+43.8%), and instruction-following (+151.7%), while its additional training also reduces emphasis on model self-reference (-44.1%) and hallucination management (-68.5%). Our analysis shows that model diffing can yield fine-grained insights beyond leaderboard metrics, attributing performance gaps to concrete mechanistic capabilities. This approach offers a transparent and targeted framework for comparing LLMs.

Prasanna Reddy Pulakurthi, Jiamian Wang, Majid Rabbani, Sohail Dianat, Raghuveer Rao, Zhiqiang Tao

Prevalent text-to-video retrieval systems mainly adopt embedding models for feature extraction and compute cosine similarities for ranking. However, this design presents two limitations. Low-quality text-video data pairs could compromise the retrieval, yet are hard to identify and examine. Cosine similarity alone provides no explanation for the ranking results, limiting the interpretability. We ask that can we interpret the ranking results, so as to assess the retrieval models and examine the text-video data? This work proposes X-CoT, an explainable retrieval framework upon LLM CoT reasoning in place of the embedding model-based similarity ranking. We first expand the existing benchmarks with additional video annotations to support semantic understanding and reduce data bias. We also devise a retrieval CoT consisting of pairwise comparison steps, yielding detailed reasoning and complete ranking. X-CoT empirically improves the retrieval performance and produces detailed rationales. It also facilitates the model behavior and data quality analysis. Code and data are available at: https://github.com/PrasannaPulakurthi/X-CoT.

Fengze Liu, Haoyu Wang, Joonhyuk Cho, Dan Roth, Andrew Lo

Clinical trials are critical for advancing medical treatments but remain prohibitively expensive and time-consuming. Accurate prediction of clinical trial outcomes can significantly reduce research and development costs and accelerate drug discovery. While recent deep learning models have shown promise by leveraging unstructured data, their black-box nature, lack of interpretability, and vulnerability to label leakage limit their practical use in high-stakes biomedical contexts. In this work, we propose AutoCT, a novel framework that combines the reasoning capabilities of large language models with the explainability of classical machine learning. AutoCT autonomously generates, evaluates, and refines tabular features based on public information without human input. Our method uses Monte Carlo Tree Search to iteratively optimize predictive performance. Experimental results show that AutoCT performs on par with or better than SOTA methods on clinical trial prediction tasks within only a limited number of self-refinement iterations, establishing a new paradigm for scalable, interpretable, and cost-efficient clinical trial prediction.

Keenan Samway, Max Kleiman-Weiner, David Guzman Piedrahita, Rada Mihalcea, Bernhard Schölkopf, Zhijing Jin

As AI systems increasingly navigate applications in healthcare, law, and governance, understanding how they handle ethically complex scenarios becomes critical. Previous work has mainly examined the moral judgments in large language models (LLMs), rather than their underlying moral reasoning process. In contrast, we focus on a large-scale analysis of the moral reasoning traces provided by LLMs. Furthermore, unlike prior work that attempted to draw inferences from only a handful of moral dilemmas, our study leverages over 600 distinct trolley problems as probes for revealing the reasoning patterns that emerge within different LLMs. We introduce and test a taxonomy of moral rationales to systematically classify reasoning traces according to two main normative ethical theories: consequentialism and deontology. Our analysis reveals that LLM chains-of-thought favor deontological principles based on moral obligations, while post-hoc explanations shift notably toward consequentialist rationales that emphasize utility. Our framework provides a foundation for understanding how LLMs process and articulate ethical considerations, an important step toward safe and interpretable deployment of LLMs in high-stakes decision-making environments.

Varun Dhanraj, Chris Eliasmith

Large language models (LLMs) continue to face challenges in reliably solving reasoning tasks, particularly tasks that involve precise rule following, as often found in mathematical reasoning tasks. This paper introduces a novel neurosymbolic method that improves LLM reasoning by encoding hidden states into neurosymbolic vectors, enabling problem-solving within a neurosymbolic vector space. The results are decoded and merged with the original hidden state, significantly boosting the model’s performance on numerical reasoning tasks. By offloading computation through neurosymbolic representations, this method enhances efficiency, reliability, and interpretability. Our experimental results demonstrate an average of 88.6% lower cross-entropy loss and 15.4 times more problems correctly solved on a suite of mathematical reasoning tasks compared to chain-of-thought prompting and supervised fine-tuning (LoRA), while not hindering the LLM’s performance on other tasks. We make our code available at https://github.com/vdhanraj/Neurosymbolic-LLM.

Oghenevovwe Ikumariegbe, Eduardo Blanco, Ellen Riloff

Distinguishing between rhetorical questions and informational questions is a challenging task, as many rhetorical questions have similar surface forms to informational questions. Existing datasets, however, do not contain many questions that can be rhetorical or informational in different contexts. We introduce Studying Rhetorically Ambiguous Questions (SRAQ), a new dataset explicitly constructed to support the study of such rhetorical ambiguity. The questions in SRAQ can be interpreted as either rhetorical or informational depending on the context. We evaluate the performance of state-of-the-art language models on this dataset and find that they struggle to recognize many rhetorical questions.

Massimiliano Pronesti, Michela Lorandi, Paul Flanagan, Oisín Redmond, Anya Belz, Yufang Hou

Systematic reviews in medicine play a critical role in evidence-based decision-making by aggregating findings from multiple studies. A central bottleneck in automating this process is extracting numeric evidence and determining study-level conclusions for specific outcomes and comparisons. Prior work has framed this problem as a textual inference task by retrieving relevant content fragments and inferring conclusions from them. However, such approaches often rely on shallow textual cues and fail to capture the underlying numeric reasoning behind expert assessments.In this work, we conceptualise the problem as one of quantitative reasoning. Rather than inferring conclusions from surface text, we extract structured numerical evidence (e.g., event counts or standard deviations) and apply domain knowledge informed logic to derive outcome-specific conclusions. We develop a numeric reasoning system composed of a numeric data extraction model and an effect estimate component, enabling more accurate and interpretable inference aligned with the domain expert principles. We train the numeric data extraction model using different strategies, including supervised fine-tuning (SFT) and reinforcement learning (RL) with a new value reward model.When evaluated on the CochraneForest benchmark, our best-performing approach – using RL to train a small-scalenumber extraction model – yields up to a 21% absolute improvement in F1 score over retrieval-based systems and outperforms general-purpose LLMs of over 400B parameters by up to 9%.Our results demonstrate the promise of reasoning-driven approaches for automating systematic evidence synthesis.

Mohna Chakraborty, Lu Wang, David Jurgens

Large language models (LLMs) are increasingly deployed in domains requiring moral understanding, yet their reasoning often remains shallow, and misaligned with human reasoning. Unlike humans, whose moral reasoning integrates contextual trade-offs, value systems, and ethical theories, LLMs often rely on surface patterns, leading to biased decisions in morally and ethically complex scenarios. To address this gap, we present a value-grounded framework for evaluating and distilling structured moral reasoning in LLMs. We benchmark 12 open-source models across four moral datasets using a taxonomy of prompts grounded in value systems, ethical theories, and cognitive reasoning strategies. Our evaluation is guided by four questions: (1) Does reasoning improve LLM decision-making over direct prompting? (2) Which types of value/ethical frameworks most effectively guide LLM reasoning? (3) Which cognitive reasoning strategies lead to better moral performance? (4) Can small-sized LLMs acquire moral competence through distillation? We find that prompting with explicit moral structure consistently improves accuracy and coherence, with first-principles reasoning and Schwartz’s + care-ethics scaffolds yielding the strongest gains. Furthermore, our supervised distillation approach transfers moral competence from large to small models without additional inference cost. Together, our results offer a scalable path toward interpretable and value-grounded models.

Seoyoon Park, Hyeji Choi, Minseon Kim, Subin An, Xiaonan Wang, Gyuri Choi, Hansaem Kim

Figurative language conveys stance, emotion, and social nuance, making its appropriate use essential in dialogue. While large language models (LLMs) often succeed in recognizing figurative expressions at the sentence level, their ability to use them coherently in conversation remains uncertain. We introduce FLUID QA, the first multilingual benchmark that evaluates figurative usage in dialogue across English, Korean, and Chinese. Each item embeds figurative choices into multi-turn contexts. To support interpretation, we include FLUTE-bi, a sentence-level diagnostic task. Results reveal a persistent gap: models that perform well on FLUTE-bi frequently fail on FLUID QA, especially in sarcasm and metaphor. These errors reflect systematic rhetorical confusion and limited discourse reasoning. FLUID QA provides a scalable framework for assessing usage-level figurative competence across languages.

Mostafa Saeed, Nizar Habash

Lemmatization is crucial for NLP tasks in morphologically rich languages with ambiguous orthography like Arabic, but existing tools face challenges due to inconsistent standards and limited genre coverage. This paper introduces two novel approaches that frame lemmatization as classification into a Lemma-POS-Gloss (LPG) tagset, leveraging machine translation and semantic clustering. We also present a new Arabic lemmatization test set covering diverse genres, standardized alongside existing datasets. We evaluate character-level sequence-to-sequence models, which perform competitively and offer complementary value, but are limited to lemma prediction (not LPG) and prone to hallucinating implausible forms. Our results show that classification and clustering yield more robust, interpretable outputs, setting new benchmarks for Arabic lemmatization.

Jack Gallifant, Shan Chen, Kuleen Sasse, Hugo Aerts, Thomas Hartvigsen, Danielle Bitterman

Sparse Autoencoders (SAEs) provide potential for uncovering structured, human-interpretable representations in Large Language Models (LLMs), making them a crucial tool for transparent and controllable AI systems. We systematically analyze SAE for interpretable feature extraction from LLMs in safety-critical classification tasks. Our framework evaluates (1) model-layer selection and scaling properties, (2) SAE architectural configurations, including width and pooling strategies, and (3) the effect of binarizing continuous SAE activations. SAE-derived features achieve macro F1 > 0.8, outperforming hidden-state and BoW baselines while demonstrating cross-model transfer from Gemma 2 2B to 9B-IT models. These features generalize in a zero-shot manner to cross-lingual toxicity detection and visual classification tasks. Our analysis highlights the significant impact of pooling strategies and binarization thresholds, showing that binarization offers an efficient alternative to traditional feature selection while maintaining or improving performance. These findings establish new best practices for SAE-based interpretability and enable scalable, transparent deployment of LLMs in real-world applications.

Thibaud Ardoin, Yi Cai, Gerhard Wunder

Hallucination remains a critical failure mode of large language models (LLMs), undermining their trustworthiness in real-world applications. In this work, we focus on confabulation, a foundational aspect of hallucination where the model fabricates facts about unknown entities. We introduce a targeted dataset designed to isolate and analyze this behavior across diverse prompt types. Using this dataset, and building on recent progress in interpreting LLM internals, we extract latent directions associated with confabulation using sparse projections. A simple vector-based steering method demonstrates that these directions can modulate model behavior with minimal disruption, shedding light on the inner representations that drive factual and non-factual output. Our findings contribute to a deeper mechanistic understanding of LLMs and pave the way toward more trustworthy and controllable generation. We release the code and dataset at https://github.com/Thibaud-Ardoin/where-confabulation-lives.

Qi Lin, Weikai Xu, Lisi Chen, Bin Dai

With the continued proliferation of Large Language Model (LLM) based chatbots, there is a growing demand for generating responses that are not only linguistically fluent but also consistently aligned with persona-specific traits in conversations. However, existing role-play and persona-based chat approaches rely heavily on static role descriptions, coarse-grained signal space, and low-quality synthetic data, which fail to capture dynamic fine-grained details in human-like chat. Human-like chat requires modeling subtle latent traits, such as emotional tone, situational awareness, and evolving personality, which are difficult to predefine and cannot be easily learned from synthetic or distillation-based data. To address these limitations, we propose a Verbal Variational Auto-Encoding (V-VAE) framework, containing a variational auto-encoding module and fine-grained control space which dynamically adapts dialogue behaviour based on fine-grained, interpretable latent variables across talking style, interaction patterns, and personal attributes. We also construct a high-quality dataset, HumanChatData, and benchmark HumanChatBench to address the scarcity of high-quality data in the human-like domain. Experiments show that LLMs based on V-VAE consistently outperform standard baselines on HumanChatBench and DialogBench, which further demonstrates the effectiveness of V-VAE and HumanChatData.

Kerem Zaman, Shashank Srivastava

Large Language Models (LLMs) offer natural language explanations as an alternative to feature attribution methods for model interpretability. However, despite their plausibility, they may not reflect the model’s truereasoning faithfully. While several faithfulness metrics have been proposed, they are often evaluated in isolation, making principled comparisons between them difficult. We present Causal Diagnosticity, a testbed framework for evaluating faithfulness metrics for natural language explanations. We use the concept of diagnosticity, and employ model-editing methods to generate faithful-unfaithful explanation pairs. Our benchmark includes four tasks: fact-checking, analogy, object counting, and multi-hop reasoning. We evaluate prominent faithfulness metrics, including post-hoc explanation and chain-of-thought methods. Diagnostic performance varies across tasks and models, with Filler Tokens performing best overall. Additionally, continuous metrics are generally more diagnostic than binary ones but can be sensitive to noise and model choice. Our results highlight the need for more robust faithfulness metrics.

Abir Harrasse, Philip Quirke, Clement Neo, Dhruv Nathawani, Luke Marks, Amir Abdullah

Mechanistic interpretability research faces a gap between analyzing simple circuits in toy tasks and discovering features in large models. To bridge this gap, we propose text-to-SQL generation as an ideal task to study, as it combines the formal structure of toy tasks with real-world complexity. We introduce TinySQL, a synthetic dataset, progressing from basic to advanced SQL operations, and train models ranging from 33M to 1B parameters to establish a comprehensive testbed for interpretability. We apply multiple complementary interpretability techniques, including Edge Attribution Patching and Sparse Autoencoders, to identify minimal circuits and components supporting SQL generation. We compare circuits for different SQL subskills, evaluating their minimality, reliability, and identifiability. Finally, we conduct a layerwise logit lens analysis to reveal how models compose SQL queries across layers: from intent recognition to schema resolution to structured generation. Our work provides a robust framework for probing and comparing interpretability methods in a structured, progressively complex setting.

Pranav Bhandari, Nicolas Fay, Michael J Wise, Amitava Datta, Stephanie Meek, Usman Naseem, Mehwish Nasim

Large Language Models (LLMs) are widely used as conversational agents exploiting their capabilities in various sectors such as education, law, medicine, and more. However, LLMs are often subjected to context-shifting behaviour, resulting in a lack of consistent and interpretable personality-aligned interactions. Adherence to psychological traits lacks comprehensive analysis, especially in the case of dyadic (pairwise) conversations. We examine this challenge from two viewpoints, initially using two conversation agents to generate a discourse on a certain topic with an assigned personality from the OCEAN framework (Openness, Conscientiousness, Extraversion, Agreeableness, and Neuroticism) as High/Low for each trait. This is followed by using multiple judge agents to infer the original traits assigned to explore prediction consistency, inter-model agreement, and alignment with the assigned personality. Our findings indicate that while LLMs can be guided toward personality-driven dialogue, their ability to maintain personality traits varies significantly depending on the combination of models and discourse settings. These inconsistencies emphasise the challenges in achieving stable and interpretable personality-aligned interactions in LLMs.

Ikhyun Cho, Julia Hockenmaier

Sparse autoencoders (SAEs) have emerged as a powerful analytical tool in mechanistic interpretability for large language models (LLMs), with growing success in applications beyond interpretability. Building on this momentum, we present a novel approach that leverages SAEs to enhance the general in-context learning (ICL) performance of LLMs.Specifically, we introduce Feature Detection through Prompt Variation (FDPV), which leverages the SAE’s remarkable ability to capture subtle differences between prompts, enabling efficient feature selection for downstream steering. In addition, we propose a novel steering method tailored to ICL—Selective In-Context Steering (SISTER)—grounded in recent insights from ICL research that LLMs utilize label words as key anchors. Our method yields a 3.5% average performance improvement across diverse text classification tasks and exhibits greater robustness to hyperparameter variations compared to standard steering approaches. Our code is available at https://github.com/ihcho2/SAE-ICL.