NAACL 2025 Papers

Large language models (LLMs) have shown remarkable performance on code generation tasks. A recent use case is iterative code repair, where an LLM fixes an incorrect program by rationalizing about errors and generating new code. Recent works augment the code repair process by integrating modern techniques such as chain-of-thought reasoning or distillation, but only study their benefits on high-resource languages like Python, and ignore low-resource languages like Perl. To address this gap of knowledge, we investigate the benefits of distilling code repair for both high and low resource languages to determine if the techniques that are effective in a high resource setting are also applicable in a low resource setting. Our evaluation shows that distilling the ability to repair code has language dependent benefits. To explain this behavior, we perform a further analysis and find that contrary to preexisting beliefs, the correlation between reasoning ability and code correction ability is weak. We hypothesize this weak correlation is magnified in low-resource settings where base models lack deep knowledge of a programming language, leading to wavering benefits of code repair.

Reasoning abilities of LLMs have been a key focus in recent years. One challenging reasoning domain with interesting nuances is legal reasoning, which requires careful application of rules, and precedents while balancing deductive and analogical reasoning, and conflicts between rules. Although there have been a few works on using LLMs for legal reasoning, their focus has been on overall accuracy. In this paper, we dig deeper to do a step-by-step analysis and figure out where they commit errors. We use the college-level Multiple Choice Question-Answering (MCQA) task from the Civil Procedure dataset and propose a new error taxonomy derived from initial manual analysis of reasoning chains with respect to several LLMs, including two objective measures: soundness and correctness scores. We then develop an LLM-based automated evaluation framework to identify reasoning errors and evaluate the performance of LLMs. The computation of soundness and correctness on the dataset using the auto-evaluator framework reveals several interesting insights. Furthermore, we show that incorporating the error taxonomy as feedback in popular prompting techniques marginally increases LLM performance. Our work will also serve as an evaluation framework that can be used in detailed error analysis of reasoning chains for logic-intensive complex tasks.

Humans have strong capabilities of decomposition and composition in natural-to-formal language conversion (N2F) when faced with an unfamiliar formal language, and can easily cope with compositional gaps and counter-intuitive symbolic names. To investigate whether large language models (LLMs) have this set of basic capabilities in N2F, we propose the STD framework. This framework semi-automatically performs sample and task construction, allowing decoupled evaluation of the set of decomposition and composition capabilities of LLMs in N2F. Based on this framework, we evaluate and analyze the most advanced LLMs, and the main findings include that: (1) the LLMs are deficient in both decomposition and composition; (2) the LLMs show a wide coverage of error types that can be attributed to deficiencies in natural language understanding and the learning and use of symbolic systems; (3) compositional gaps and counter-intuitive symbolic names both affect the decomposition and composition of the LLMs. Our work provides a new perspective for investigating the basic capabilities of decomposition and composition of LLMs in N2F. The detailed analysis of deficiencies and attributions can help subsequent improvements of LLMs.

Studying human values is instrumental for cross-cultural research, enabling a better understanding of preferences and behaviour of society at large and communities therein. To study the dynamics of communities online, we propose a method to computationally analyse values present on Reddit. Our method allows analysis at scale, complementing survey based approaches. We train a value relevance and a value polarity classifier, which we thoroughly evaluate using in-domain and out-of-domain human annotations. Using these, we automatically annotate over nine million posts across 12k subreddits with Schwartz values. Our analysis unveils both previously recorded and novel insights into the values prevalent within various online communities. For instance, we discover a very negative stance towards conformity in the Vegan and AbolishTheMonarchy subreddits. Additionally, our study of geographically specific subreddits highlights the correlation between traditional values and conservative U.S. states. Through our work, we demonstrate how our dataset and method can be used as a complementary tool for qualitative study of online communication.

Simultaneous Machine Translation (SiMT) generates target translation before receiving the whole source sentence and faces a serious hallucination problem. In contrast, traditional offline machine translation (OMT) models exhibit significantly fewer hallucinations. Motivated by this disparity, we propose Knowledge Distillation for SiMT (KD-SiMT), a simple yet effective method that utilizes the OMT model to mitigate hallucinations in SiMT. Experiments on Zh→En and De→En tasks demonstrate that KD-SiMT effectively reduces hallucinations and enhances the SiMT performance. Furthermore, we systematically investigate the deficiencies in SiMT models related to serious hallucinations and the effect of KD-SiMT. Specifically, we design targeted tasks and metrics to quantitatively evaluate the components in SiMT models from the perspectives of model structure and knowledge acquisition. Our analyses reveal that inaccurate source representations and imbalanced cross-attention are more likely to occur in SiMT models when generating hallucinations, while KD-SiMT alleviates these issues. Besides, we find that KD-SiMT equips SiMT models with sufficient faithfulness knowledge in training, thus reducing hallucinations.

Large language models (LLMs) have significantly influenced various industries but suffer from a critical flaw, the potential sensitivity of generating harmful content, which poses severe societal risks. We developed and tested novel attack strategies on popular LLMs to expose their vulnerabilities in generating inappropriate content. These strategies, inspired by psychological phenomena such as the “Priming Effect”, “Safe Attention Shift”, and “Cognitive Dissonance”, effectively attack the models’ guarding mechanisms. Our experiments achieved an attack success rate (ASR) of 100% on various open-source models, including Meta’s Llama-3.2, Google’s Gemma-2, Mistral’s Mistral-NeMo, Falcon’s Falcon-mamba, Apple’s DCLM, Microsoft’s Phi3, and Qwen’s Qwen2.5, among others. Similarly, for closed-source models such as OpenAI’s GPT-4o, Google’s Gemini-1.5, and Claude-3.5, we observed an ASR of at least 95% on the AdvBench dataset, which represents the current state-of-the-art. This study underscores the urgent need to reassess the use of generative models in critical applications to mitigate potential adverse societal impacts.

While recent work has found that vision-language models trained under the Contrastive Language Image Pre-training (CLIP) framework contain intrinsic social biases, the extent to which different upstream pre-training features of the framework relate to these biases, and hence how intrinsic bias and downstream performance are connected has been unclear. In this work, we present the largest comprehensive analysis to-date of how the upstream pre-training factors and downstream performance of CLIP models relate to their intrinsic biases. Studying 131 unique CLIP models, trained on 26 datasets, using 55 architectures, and in a variety of sizes, we evaluate bias in each model using 26 well-established unimodal and cross-modal principled Embedding Association Tests. We find that the choice of pre-training dataset is the most significant upstream predictor of bias, whereas architectural variations have minimal impact. Additionally, datasets curated using sophisticated filtering techniques aimed at enhancing downstream model performance tend to be associated with higher levels of intrinsic bias. Finally, we observe that intrinsic bias is often significantly correlated with downstream performance (0.3 ≤ r ≤ 0.8), suggesting that models optimized for performance inadvertently learn to amplify representational biases. Comparisons between unimodal and cross-modal association tests reveal that social group bias depends heavily on the modality. Our findings imply that more sophisticated strategies are needed to address intrinsic model bias for vision-language models across the entire model development pipeline.

Dense embeddings deliver strong retrieval performance but often lack interpretability and controllability. This paper introduces a novel approach using sparse autoencoders (SAE) to interpret and control dense embeddings via the learned latent sparse features. Our key contribution is the development of a retrieval-oriented contrastive loss, which ensures the sparse latent features remain effective for retrieval tasks and thus meaningful to interpret. Experimental results demonstrate that both the learned latent sparse features and their reconstructed embeddings retain nearly the same retrieval accuracy as the original dense vectors, affirming their faithfulness. Our further examination of the sparse latent space reveals interesting features underlying the dense embeddings and we can control the retrieval behaviors via manipulating the latent sparse features, for example, prioritizing documents from specific perspectives in the retrieval results.

Image Editing has made incredible progress in recent years. Earliest work only supported caption-guided editing. Recently, free-form text instructions and reference images are incorporated to allow more flexibility. However, existing methods still struggle with complicated editing instructions involving multiple objects or reference images. We present InstructAny2Pix, a novel image editing model that leverages a multi-modal LLM to execute complicated edit instructions. Compared with previous, works, InstructAny2Pix extends the flexibility of edit instructions in three ways: First, it can perform complex instructions involving multiple object edits; Second, it supports interleaving text instructions with multiple reference images; Third, it supports audio and music inputs as part of edit prompts, unlocking many creative applications, such as album cover generation and music-inspired merchandise design. To evaluate the effectiveness of InstructAny2Pix, we propose two new benchmark datasets MM-Inst and Dream-booth++ consisting of human written, multi-modal prompts. InstructAny2Pix outperforms baselines in these two proposed multi-modal benchmarks, as well as conventional image editing benchmarks such as InstructPix2Pix.

Self-reflection for Large LanguageModels (LLMs) has gained significant attention. Existing approaches involve models iterating and improving their previous responses based on LLMs’ internal reflection ability or external feedback. However, recent research has raised doubts about whether intrinsic self-correction without external feedback may even degrade performance. Based on our empirical evidence, we find that current static reflection methods may lead to redundant, drift, and stubborn issues. To mitigate this, we introduce **I**nstruct-**o**f-**R**eflec**t**ion (**IoRT**), a novel and general reflection framework that leverages dynamic-meta instruction to enhance the iterative reflection capability of LLMs. Specifically, we propose the instructor driven by the meta-thoughts and self-consistency classifier, generates various instructions, including refresh, stop, and select, to guide the next reflection iteration. Our experiments demonstrate that IoRT achieves an average improvement of 10.1% over established baselines in mathematical and commonsense reasoning tasks, highlighting its efficacy and applicability. Our code is available at https://github.com/llp635/IoRT.

Human Preference Alignment (HPA) can assist large language models (LLMs) to generate safe content. Due to the heavy cost of fine-tuning, tuning-free methods have emerged, typically modifying LLM decoding via post-processing. In this paper, we propose a novel and effective approach for HPA in a tuning-free way, named In-Context Direct Preference Optimization (ICDPO). We first rethink the derivation procedures of DPO, based on which we conversely build an instant scorer using the states of the LLM before and after ICL. It enables LLMs to both generate and select the well-aligned response, which is precisely estimated by the aforementioned instant scorer, thereby enhancing the final performance. ICDPO can be further enhanced with a two-stage retriever and an upgraded scorer. Extensive experiments show its effectiveness, particularly in outperforming multiple tuning-free baselines, even competitiveness with SFT and DPO. We also conduct detailed analyses to offer comprehensive insights into ICDPO.

Large Language Models (LLM) have become a popular approach for implementing Retrieval Augmented Generation (RAG) systems, and a significant amount of effort has been spent on building good models and metrics. In spite of increased recognition of the need for rigorous evaluation of RAG systems, few tools exist that go beyond the creation of model output and automatic calculation. We present InspectorRAGet, an introspection platform for performing a comprehensive analysis of the quality of RAG system output. InspectorRAGet allows the user to analyze aggregate and instance-level performance of RAG systems, using both human and algorithmicmetrics as well as annotator quality. InspectorRAGet is suitable for multiple use cases and is available publicly to the community.A live instance of the platform is available at https://ibm.biz/InspectorRAGet

High-quality training data has proven crucial for developing performant large language models (LLMs). However, commercial LLM providers disclose few, if any, details about the data used for training. This lack of transparency creates multiple challenges: it limits external oversight and inspection of LLMs for issues such as copyright infringement, it undermines the agency of data authors, and it hinders scientific research on critical issues such as data contamination and data selection. How can we recover what training data is known to LLMs? In this work we demonstrate a new method to identify training data known to proprietary LLMs like GPT-4 without requiring any access to model weights or token probabilities, by using information-guided probes. Our work builds on a key observation: text passages with high surprisal are good search material for memorization probes. By evaluating a model’s ability to successfully reconstruct high-surprisal tokens in text, we can identify a surprising number of texts memorized by LLMs.

Despite seemingly performant web agents on the task-completion benchmarks, most existing methods evaluate the agents based on a presupposition: the web navigation task consists of a linear sequence of actions with an end state that marks task completion. In contrast, our work focuses on web navigation for information aggregation, wherein the agent must explore different websites to gather information for a complex query. We consider web information aggregation from two different perspectives: i) Direct API-driven Access relies on a text-only view of the Web, leveraging external tools such as Google Search API to navigate the Web and a scraper to extract website contents. (ii) Interactive Visual Access uses screenshots of the webpages and requires interaction with the browser to navigate and access information. Motivated by these diverse information access settings, we introduce Infogent, a novel modular framework for web information aggregation involving three distinct components: Navigator, Extractor, and Aggregator. Experiments on different information access settings demonstrate that Infogent beats an existing SOTA multi-agent search framework by 7% under Direct API-Driven Access on FRAMES and improves over an existing information-seeking web agent by 4.3% under Interactive Visual Access on AssistantBench.

We study the post-training of large language models (LLMs) with human preference data. Recently, direct preference optimization and its variants have shown considerable promise in aligning language models, eliminating the need for reward models and online sampling. Despite these benefits, these methods rely on explicit assumptions about the Bradley-Terry (BT) model, which makes them prone to overfitting and results in suboptimal performance, particularly on reasoning-heavy tasks. To address these challenges, we propose a principled preference fine-tuning algorithm called InfoPO, which effectively and efficiently aligns large language models using preference data. InfoPO eliminates the reliance on the BT model and prevents the likelihood of the chosen response from decreasing. Extensive experiments confirm that InfoPO consistently outperforms established baselines on widely used open benchmarks, particularly in reasoning tasks.

We propose selective debiasing – an inference-time safety mechanism designed to enhance the overall model quality in terms of prediction performance and fairness, especially in scenarios where retraining the model is impractical. The method draws inspiration from selective classification, where at inference time, predictions with low quality, as indicated by their uncertainty scores, are discarded. In our approach, we identify the potentially biased model predictions and, instead of discarding them, we remove bias from these predictions using LEACE – a post-processing debiasing method. To select problematic predictions, we propose a bias quantification approach based on KL divergence, which achieves better results than standard uncertainty quantification methods. Experiments on text classification datasets with encoder-based classification models demonstrate that selective debiasing helps to reduce the performance gap between post-processing methods and debiasing techniques from the at-training and pre-processing categories.

Retrieval-augmented generation (RAG) has emerged as a popular approach to steering the output of a large language model (LLM) by incorporating retrieved contexts as inputs. However, existing work observed the generator bias, such that improving the retrieval results may negatively affect the outcome. In this work, we show such bias can be mitigated, from inference scaling, aggregating inference calls from the permuted order of retrieved contexts. The proposed Mixture-of-Intervention (MoI) explicitly models the debiased utility of each passage with multiple forward passes to construct a new ranking. We also show that MoI can leverage the retriever’s prior knowledge to reduce the computational cost by minimizing the number of permutations considered and lowering the cost per LLM call. We showcase the effectiveness of MoI on diverse RAG tasks, improving ROUGE-L on MS MARCO and EM on HotpotQA benchmarks by ~7 points.

Large language models (LLMs) have shown a remarkable ability to learn and perform complex tasks through in-context learning (ICL). However, a comprehensive understanding of its internal mechanisms is still lacking. This paper explores the role of induction heads in a few-shot ICL setting. We analyse two state-of-the-art models, Llama-3-8B and InternLM2-20B on abstract pattern recognition and NLP tasks. Our results show that even a minimal ablation of induction heads leads to ICL performance decreases of up to ~32% for abstract pattern recognition tasks, bringing the performance close to random. For NLP tasks, this ablation substantially decreases the model’s ability to benefit from examples, bringing few-shot ICL performance close to that of zero-shot prompts. We further use attention knockout to disable specific induction patterns, and present fine-grained evidence for the role that the induction mechanism plays in ICL.

Autoregressive transformer language models (LMs) possess strong syntactic abilities, often successfully handling phenomena from agreement to NPI licensing. However, the features they use to incrementally process their linguistic input are not well understood. In this paper, we fill this gap by studying the mechanisms underlying garden path sentence processing in LMs. Specifically, we ask: (1) Do LMs use syntactic features or shallow heuristics to perform incremental sentence processing? (2) Do LMs represent only one potential interpretation, or multiple? and (3) Do LMs reanalyze or repair their initial incorrect representations? To address these questions, we use sparse autoencoders to identify interpretable features that determine which continuation—and thus which reading—of a garden path sentence the LM prefers. We find that while many important features relate to syntactic structure, some reflect syntactically irrelevant heuristics. Moreover, though most active features correspond to one reading of the sentence, some features correspond to the other, suggesting that LMs assign weight to both possibilities. Finally, LMs fail to re-use features to answer follow-up questions.