ACL 2025 Papers

Fairness in NLP must extend beyond performance parity to encompass equitable reliability across groups. This study exposes a criticalblind spot: models often make less reliable or overconfident predictions for marginalized groups, even when overall performance appearsfair. Using the FairLex benchmark as a case study in legal NLP, we systematically evaluate both performance and reliability dispari-ties across demographic, regional, and legal attributes spanning four jurisdictions. We show that domain-specific pre-training consistentlyimproves both performance and reliability, especially for underrepresented groups. However, common bias mitigation methods frequentlyworsen reliability disparities, revealing a trade-off not captured by performance metrics alone. Our results call for a rethinking of fairnessin high-stakes NLP: To ensure equitable treatment, models must not only be accurate, but also reliably self-aware across all groups.

Large language models (LLMs) are prone to capturing biases from training corpus, leading to potential negative social impacts. Existing prompt-based debiasing methods exhibit instability due to their sensitivity to prompt changes, while fine-tuning-based techniques incur substantial computational overhead and catastrophic forgetting. In this paper, we propose FairSteer, a novel inference-time debiasing framework without requiring customized prompt design or model retraining. Motivated by the linear representation hypothesis, our preliminary investigation demonstrates that fairness-related features can be encoded into separable directions in the hidden activation space. FairSteer operates in three steps: biased activation detection, debiasing steering vector (DSV) computation, and dynamic activation steering. Specifically, it first trains a lightweight linear classifier to detect bias signatures in activations, and then computes DSVs as intervention directions derived from small contrastive prompt pairs. Subsequently, it performs debiasing by adjusting activations with DSVs in the inference stage. Comprehensive evaluation with six LLMs demonstrates the superiority of FairSteer across question-answering, counterfactual input evaluation and open-ended text generation tasks. Code will be released.

Existing studies on fairness are largely Western-focused, making them inadequate for culturally diverse countries such as India. To address this gap, we introduce INDIC-BIAS, a comprehensive India-centric benchmark designed to evaluate fairness of LLMs across 85 identity groups encompassing diverse castes, religions, regions, and tribes. We first consult domain experts to curate over 1,800 socio-cultural topics spanning behaviors and situations, where biases and stereotypes are likely to emerge. Grounded in these topics, we generate and manually validate 20,000 real-world scenario templates to probe LLMs for fairness. We structure these templates into three evaluation tasks: plausibility, judgment, and generation. Our evaluation of 14 popular LLMs on these tasks reveals strong negative biases against marginalized identities, with models frequently reinforcing common stereotypes. Additionally, we find that models struggle to mitigate bias even when explicitly asked to rationalize their decision. Our evaluation provides evidence of both allocative and representational harms that current LLMs could cause towards Indian identities, calling for a more cautious usage in practical applications. We release INDIC-BIAS as an open-source benchmark to advance research on benchmarking and mitigating biases and stereotypes in the Indian context.

Generative Error Correction (GEC) has emerged as a powerful post-processing method to boost the performance of Automatic Speech Recognition (ASR) systems. In this paper, we first show that GEC models struggle to generalize beyond the specific types of errors encountered during training, limiting their ability to correct new, unseen errors at test time, particularly in out-of-domain (OOD) scenarios. This phenomenon amplifies with named entities (NEs), where, in addition to insufficient contextual information or knowledge about the NEs, novel NEs keep emerging. To address these issues, we propose DARAG (Data- and Retrieval-Augmented Generative Error Correction), a novel approach designed to improve GEC for ASR in in-domain (ID) and OOD scenarios. First, we augment the GEC training dataset with synthetic data generated using foundational generative models, thereby simulating additional errors from which the model can learn from. For out-of-domain scenarios, we simulate test-time errors from new domains similarly and in an unsupervised fashion. Additionally, to better handle NEs, we introduce retrieval-augmented correction wherein we augment the model input with entities retrieved from a datastore of NEs. Our approach is simple, scalable, and both domain- and language-agnostic. We experiment on multiple datasets and settings, showing that DARAG outperforms all our baselines, achieving 8%–30% relative WER improvements in ID and 10%–33% improvements in OOD settings.

Large Language Models (LLMs) have shown impressive capability in language generation and understanding, but their tendency to hallucinate and produce factually incorrect information remains a key limitation. To verify LLM-generated contents and claims from other sources, traditional verification approaches often rely on holistic models that assign a single factuality label to complex claims, potentially obscuring nuanced errors. In this paper, we advocate for a shift towards fine-grained verification, where complex claims are broken down into smaller sub-claims for individual verification, allowing for more precise identification of inaccuracies, improved transparency, and reduced ambiguity in evidence retrieval. However, generating sub-claims poses challenges, such as maintaining context and ensuring semantic equivalence with respect to the original claim. We introduce **FactLens**, a benchmark for evaluating fine-grained fact verification, with metrics and automated evaluators of sub-claim quality. The benchmark data is manually curated to ensure high-quality ground truth. Our results show alignment between automated FactLens evaluators and human judgments, and we discuss the impact of sub-claim characteristics on the overall verification performance.

The rapid adoption of language models (LMs) across diverse applications has raised concerns about their factuality, i.e., their consistency with real-world facts. We introduce VERIFY, an evidence-based evaluation pipeline that measures LMs’ factuality in real-world user interactions. VERIFY considers the verifiability of LM-generated content and categorizes content units as Supported, Unsupported, or Undecidable based on Web-retrieved evidence. Importantly, factuality judgment by VERIFY more strongly correlates with human evaluations than existing methods. Using VERIFY, we identify “hallucination prompts,” i.e., those that frequently elicit factual errors in LM responses. These prompts form FactBench, a dataset of 1K prompts spanning 150 topics and tiered into Easy, Moderate, and Hard prompts. We benchmark widely-used openweight and proprietary LMs from six families, yielding three key findings: (i) LMs’ factual precision declines from Easy to Hard prompts, (ii) factuality does not necessarily improve with scale; Llama3.1-405B-Instruct performs comparably to or worse than its 70B variant, and (iii) Gemini1.5-Pro shows a notably higher refusal rate, with over-refusal in 25% of cases.

Language models (LMs) can make a correct prediction based on many possible signals in a prompt, not all corresponding to recall of factual associations. However, current interpretations of LMs fail to take this into account. For example, given the query “Astrid Lindgren was born in” with the corresponding completion “Sweden”, no difference is made between whether the prediction was based on knowing where the author was born or assuming that a person with a Swedish-sounding name was born in Sweden. In this paper, we present a model-specific recipe - PrISM - for constructing datasets with examples of four different prediction scenarios: generic language modeling, guesswork, heuristics recall and exact fact recall. We apply two popular interpretability methods to the scenarios: causal tracing (CT) and information flow analysis. We find that both yield distinct results for each scenario. Results for exact fact recall and generic language modeling scenarios confirm previous conclusions about the importance of mid-range MLP sublayers for fact recall, while results for guesswork and heuristics indicate a critical role of late last token position MLP sublayers. In summary, we contribute resources for a more extensive and granular study of fact completion in LMs, together with analyses that provide a more nuanced understanding of how LMs process fact-related queries.

Targeting long-form question-answering, chain-of-query (CoQ) has been studied, integrating chain-of-thought (CoT) with retrieval-augmented generation. CoQ answers the complex question step-by-step, through simpler subquestions (SQs) from which relevant knowledge is retrieved. By doing so, CoQ aims to improve the answer comprehensiveness and verifiability, at the expense of latency. Our first contribution is showing that the chaining often incurs harmful effects on both objectives, and SQs left unverified often fail to answer the given question. Second, we propose a better alternative to CoQ, union-of-query which adopts a factored approach to break the harmful chain. Finally, we propose to verify SQs before answers, by fine-tuning the SQ generator using verified SQs and introducing a selector verifying SQs in test time. Employing vicuna-13b, our approach, denoted by FaVe (short for Factored and Verified search), even outperforms ChatGPT baselines while maintaining efficiency.

In recent years, Vision-Language Models (VLMs) have shown remarkable performance improvements in Vision-Language tasks. However, their large size poses challenges for real-world applications where inference latency is a concern. To tackle this issue, we propose employing Early Exit (EE) strategies in VLMs. However, training exit classifiers in VLMs is challenging, particularly with limited labeled training data. To address this, we introduce FREE, an adversarial training approach within a GAN-based framework. Here, each exit consists of a transformer layer and a classifier. The transformer layer is adversarially trained to produce feature representations similar to the final layer, while a feature classifier serves as the discriminator. Our method focuses on performing input-adaptive inference that increases inference speed with minimal drop in performance. Experimental results demonstrate the effectiveness of our approach in enhancing accuracy and model robustness by mitigating overthinking and the phenomenon of mid-crisis that we highlight. We experimentally validate that our method speeds up the inference process by more than 1.51× while retaining comparable performance. The anonymized source code is available at https://github.com/Div290/BLIPEE.

The automation of scientific research through large language models (LLMs) presents significant opportunities but faces critical challenges in knowledge synthesis and quality assurance. We introduce Feedback-Refined Agent Methodology (FRAME), a novel framework that enhances medical paper generation through iterative refinement and structured feedback. Our approach comprises three key innovations: (1) A structured dataset construction method that decomposes 4,287 medical papers into essential research components through iterative refinement; (2) A tripartite architecture integrating Generator, Evaluator, and Reflector agents that progressively improve content quality through metric-driven feedback; and (3) A comprehensive evaluation framework that combines statistical metrics with human-grounded benchmarks. Experimental results demonstrate FRAME’s effectiveness, achieving significant improvements over conventional approaches across multiple models (9.91% average gain with DeepSeek V3, comparable improvements with GPT-4o Mini) and evaluation dimensions. Human evaluation confirms that FRAME-generated papers achieve quality comparable to human-authored works, with particular strength in synthesizing future research directions. The results demonstrated our work could efficiently assist medical research by building a robust foundation for automated medical research paper generation while maintaining rigorous academic standards.

Large language models (LLMs) have significantly advanced human language understanding and generation, with pretraining data quality and organization being crucial to their performance. Multi-stage pretraining is a promising approach, but existing methods often lack quantitative criteria for data partitioning and instead rely on intuitive heuristics. In this paper, we propose the novel Four-quadRAnt Multi-stage prEtraining strategy (FRAME), guided by the established principle of organizing the pretraining process into four stages to achieve significant loss reductions four times. This principle is grounded in two key findings: first, training on high Perplexity (PPL) data followed by low PPL data, and second, training on low PPL difference (PD) data followed by high PD data, both causing the loss to drop significantly twice and performance enhancements. By partitioning data into four quadrants and strategically organizing them, FRAME achieves a remarkable 16.8% average improvement over random across MMLU and CMMLU for the 3B model, effectively boosting LLM performance.

To mitigate the hallucination and knowledge deficiency in large language models (LLMs), Knowledge Graph (KG)-based Retrieval-Augmented Generation (RAG) has shown promising potential by utilizing KGs as an external resource to enhance LLM reasoning.However, existing KG-RAG approaches struggle with a trade-off between flexibility and retrieval quality. Modular methods prioritize flexibility by avoiding the use of KG-fine-tuned models during retrieval, leading to fixed retrieval strategies and suboptimal retrieval quality. Conversely, coupled methods embed KG information within models to improve retrieval quality but at the expense of flexibility.In this paper, we propose a novel flexible modular KG-RAG framework, termed FRAG, which synergizes the advantages of both approaches. FRAG estimates the hop range of reasoning paths based solely on the query and classifies it as either simple or complex.To match the complexity of the query, tailored pipelines are applied to ensure efficient and accurate reasoning path retrieval, thus fostering the final reasoning process. By using the query text instead of the KG to infer the structural information of reasoning paths and employing adaptable retrieval strategies, FRAG improves retrieval quality while maintaining flexibility. Moreover, FRAG does not require extra LLM fine-tuning or calls, significantly boosting efficiency and conserving resources. Extensive experiments show that FRAG achieves state-of-the-art performance with high efficiency and low resource consumption. The code for our method is publicly available at https://github.com/gzy02/FRAG.

Fine-Tuning of LLMs using RLHF / RLAIF has been shown as a critical step to improve the performance of LLMs in complex generation tasks. These methods typically use response-level human or model feedback for alignment. Recent works indicate that finer sentence or span-level labels provide more accurate and interpretable feedback for LLM optimization. In this work, we propose FRACTAL, a suite of models to disaggregate response-level labels into sentence-level (pseudo-)labels through Multiple Instance Learning (MIL) and Learning from Label Proportions (LLP) formulations, novel usage of prior information, and maximum likelihood calibration. We perform close to 2000 experiments across 6 datasets and 4 tasks that show that FRACTAL can reach up to 93% of the performance of the fully supervised baseline while requiring only around 10% of the gold labels. Furthermore, in a downstream eval, employing step-level pseudo scores in RLHF for a math reasoning task leads to 5% absolute improvement in performance. Our work is the first to develop response-level feedback to sentence-level scoring techniques leveraging sentence-level prior information, along with comprehensive evaluations on multiple tasks as well as end-to-end finetuning evaluations.

Speculative sampling has emerged as an important technique for accelerating the auto-regressive generation process of large language models (LLMs) by utilizing a draft-then-verify mechanism to produce multiple tokens per forward pass. While state-of-the-art speculative sampling methods use only a single layer and a language modeling (LM) head as the draft model to achieve impressive layer compression, their efficiency gains are substantially reduced for large-vocabulary LLMs, such as Llama-3-8B with a vocabulary of 128k tokens. To address this, we present FR-Spec, a frequency-ranked speculative sampling framework that optimizes draft candidate selection through vocabulary space compression. By constraining the draft search to a frequency-prioritized token subset, our method reduces LM Head computation overhead by 75% while ensuring the equivalence of the final output distribution. Experiments across multiple datasets demonstrate an average of 1.12× speedup over the state-of-the-art speculative sampling method EAGLE-2. Code is availableat https://github.com/thunlp/FR-Spec.

Auto-regressive decoding is a memory-bound job, meaning decoding inference performance is limited by the bandwidth rather than the computational capabilities of the GPU. Weight-only quantization is a promising method to address the memory-bound limitations. Previous studies have followed one of two approaches. Some have exclusively studied integer quantization while ignoring the Gaussian distribution nature of LLMs’ weights. Others have proposed non-uniform quantization but incurred additional I/O overhead due to lookup tables, e.g. NF4. In this work, we extend the IEEE 754 float-point standard to the ExMy quantization schema, which allocates x bit for the exponent and y bit for the mantissa to represent a number. In terms of runtime efficiency, we demonstrate that the conversion from ExMy to FP16 can be realized through register-level operations, which can get almost the same performance as INT5. In terms of quantization loss, we analyze that of different ExMy settings, where the E2M2 schema achieves an optimal balance, offering the highest efficiency with lossless accuracy. We further propose the FPE2M2 framework that supports lossless weight-only quantization inference and validate the FPE2M2 framework on Qwen and LLaMA Models across various modalities, such as text, image, and audio tasks, which achieves a faster inference speed while maintaining nearly lossless accuracy.

Humans possess a remarkable ability to interpret underspecified ambiguous statements by inferring their meanings from contexts such as visual inputs. This ability, however, may not be as developed in recent pre-trained vision-language models (VLMs). In this paper, we introduce a novel probing dataset called FOCUS to evaluate whether state-of-the-art VLMs have this ability. FOCUS consists of underspecified sentences paired with image contexts and carefully designed probing questions. Our experiments reveal that VLMs still fall short in handling underspecification even when visual inputs that can help resolve the ambiguities are available. To further support research in underspecification, FOCUS will be released for public use. We hope this dataset will inspire further research on the reasoning and contextual understanding capabilities of VLMs.

Large language models (LLMs) fine-tuned on multimodal financial data have demonstrated impressive reasoning capabilities in various financial tasks. However, they often struggle with multi-step, goal-oriented scenarios in interactive financial markets, such as trading, where complex agentic approaches are required to improve decision-making. To address this, we propose FLAG-Trader, a unified architecture integrating linguistic processing (via LLMs) with gradient-driven reinforcement learning (RL) policy optimization, in which a partially fine-tuned LLM acts as the policy network, leveraging pre-trained knowledge while adapting to the financial domain through parameter-efficient fine-tuning. Through policy gradient optimization driven by trading rewards, our framework not only enhances LLM performance in trading but also improves results on other financial-domain tasks. We present extensive empirical evidence to validate these enhancements.

The rapid development of Large Vision-Language Models (LVLMs) often comes with widespread hallucination issues, making cost-effective and comprehensive assessments increasingly vital. Current approaches mainly rely on costly annotations and are not comprehensive – in terms of evaluating all aspects, such as relations, attributes, and dependencies between aspects. Therefore, we introduce the FIHA (automated Fine-graIned Hallucination evAluation in LVLMs), which could access LVLMs hallucination in an LLM-free and annotation-free way and model the dependency between different types of hallucinations. FIHA can generate Q&A pairs on any image dataset at minimal cost, enabling hallucination assessment from both image and caption. Based on this approach, we introduce a benchmark called FIHA-v1, which consists of diverse questions on various images from three datasets. Furthermore, we use the Davidson Scene Graph (DSG) to organize the structure among Q&A pairs, in which we can increase the reliability of the evaluation. We evaluate representative models using FIHA-v1, highlighting their limitations and challenges. We released our code and data at https://github.com/confidentzzzs/FIHA.

Detecting rumors on social media has become a crucial issue.Propagation structure-based methods have recently attracted increasing attention.When the propagation structure is represented by the dynamic graph, temporal information is considered.However, existing rumor detection models using dynamic graph typically focus only on coarse-grained temporal information and ignore the fine-grained temporal dynamics within individual snapshots and across snapshots.In this paper, we propose a novel Fine-Grained Dynamic Graph Neural Network (FGDGNN) model, which can incorporate the fine-grained temporal information of dynamic propagation graph in the intra-snapshot and dynamic embedding update mechanism in the inter-snapshots into a unified framework for rumor detection.Specifically, we first construct the edge-weighted propagation graph and the edge-aware graph isomorphism network is proposed.To obtain fine-grained temporal representations across snapshots, we propose an embedding transformation layer to update node embeddings.Finally, we integrate the temporal information in the inter-snapshots at the graph level to enhance the effectiveness of the proposed model.Extensive experiments conducted on three public real-world datasets demonstrate that our FGDGNN model achieves significant improvements compared with the state-of-the-art baselines.