ACL 2025 Papers

Fine-grained classification via LLMs is susceptible to more complex label biases compared to traditional classification tasks. Existing bias mitigation strategies, such as retraining, post-hoc adjustment, and parameter-efficient fine-tuning (PEFT) are primarily effective for simple classification biases, such as stereotypes, but fail to adequately address prediction propensity and discriminative ability biases. In this paper, we analyze these two bias phenomena and observe their progressive accumulation from intermediate to deeper layers within LLMs. To mitigate this issue, we propose a bias-aware optimization framework that incorporates two distinct label balance constraints with a PEFT strategy targeting an intermediate layer. Our approach adjusts less than 1% of the model’s parameters while effectively curbing bias amplification in deeper layers. Extensive experiments conducted across 12 datasets and 5 LLMs demonstrate that our method consistently outperforms or matches the performance of full-parameter fine-tuning and LoRA, achieving superior results with lower perplexity.

Recent preference learning frameworks for large language models (LLMs) simplify human preferences with binary pairwise comparisons and scalar rewards. This simplification could make LLMs’ responses biased to mostly preferred features, and would be exacerbated during the iterations of online preference learning steps. To address these challenges, we propose a novel framework coined PFP (Preference Feature Preservation). The key idea of PFP is maintaining the distribution of human preference features and utilizing such rich signals throughout the online preference learning process. Specifically, PFP first extract preference features from offline pairwise human preference data and trains a feature classifier. Then, using trained classifier and the distribution preserving optimization, PFP maps appropriate preference features for a new input instruction during online learning. Lastly, PFP trains LLM using the existing preference learning method, by incorporating the preference feature into system prompts and enabling LLM to explicitly handle various human preferences. Our experiments demonstrate that PFP successfully mitigates the bias in preference features during online learning, and hence achieves superior performance compared to previous preference learning methods on standard benchmarks to evaluate LLM alignment.

With the impressive reasoning and text generation capabilities of large language models (LLMs), methods leveraging multiple LLMs to debate each other have garnered increasing attention. However, existing debate-based approaches remain limited in effectiveness in structured and detailed domains represented by code generation due to several reasons: 1) Reliance on different instances of the same LLM for debate, neglecting the potential benefits of integrating diverse models with varied internal knowledge for more comprehensive code generation, 2) under-utilization of test cases, and 3) reliance on third-party LLM moderators for result consolidation and decision-making, probably introducing hallucinations and judgment errors. To address these challenges, we propose DebateCoder to collect intelligence of LLMs via test case-driven debate for code generation. In DebateCoder, test cases serve as a medium for models to analyze code and identify bugs, while opposing models generate test cases to challenge each other’s code during the debate process. These test cases, along with their execution results, are elaborately leveraged to refine and enhance the code through a novel contrastive analysis process. Furthermore, DebateCoder leverages test case outcomes to assess code quality and determine convergence criteria. Unlike previous approaches, DebateCoder emphasizes the collaborative improvement of both models through competitive debate and interactive analysis. Abundant experimental results on two datasets demonstrate the effectiveness of DebateCoder.

Large language models (LLMs) have demonstrated impressive performance in reasoning. However, existing data annotation methods usually suffer from high annotation cost and the lack of effective automatic validation. To address these issues, we propose a Fine-grained Multi-Agent Debate framework (FMAD) and MMATH-Data, a dataset created by FMAD, which consists of 46K reasoning steps. By prompting multiple agents to debate, FMAD assesses the contribution of each reasoning step to the final solution, with labels based on the judge’s confidence score and the winner’s position. To facilitate reasoning in math and examine FMAD and MMATH-Data, we further propose two key components: a Multi-Agent Debate Reward Model (MRM) trained on MMATH-Data, which serves as a reward model to provide robust feedback during the optimization process, and MMATH-LLM, a model designed specifically for mathematical reasoning. MMATH-LLM is fine-tuned using reinforcement learning with supervised feedback from MRM, aiming at improving its mathematical reasoning capabilities. Extensive experiments demonstrate that our model achieves 83.4% accuracy on the GSM8K dataset and 45.1% on the MATH dataset, outperforming the state-of-the-art methods by 1.2% and 3.5%, respectively. All data and code will be available soon at GitHub.

Large Language Models (LLMs) continue to set new standards in knowledge-intensive and complex reasoning tasks, yet their high computational demands limit widespread adoption. While distilling large models into smaller ones offers a sustainable solution, current techniques—such as static knowledge distillation, resource-intensive reinforcement learning from human feedback, or limited self-reflection—struggle to yield substantial and lasting performance gains. In this paper, we present a novel Debate and Reflect (D&R) framework that orchestrates multi-turn debates between smaller models and stronger teacher models, eliciting actionable feedback (e.g., error analysis, corrective strategies) to guide student models. Further, we introduce Tree-structured Direct Preference Optimization (T-DPO) to efficiently leverage these debate logs, organizing interactions into a hierarchical format for effective training. Empirical evaluations across diverse NLP benchmarks demonstrate that our approach significantly improves smaller-model accuracy, robustness, and generalization, outperforming conventional baselines by a large margin.

With the widespread adoption of Large Language Models (LLMs), jailbreak attacks have become an increasingly pressing safety concern. While safety-aligned LLMs can effectively defend against normal harmful queries, they remain vulnerable to such attacks. Existing defense methods primarily rely on fine-tuning or input modification, which often suffer from limited generalization and reduced utility. To address this, we introduce DeTAM, a finetuning-free defense approach that improves the defensive capabilities against jailbreak attacks of LLMs via targeted attention modification. Specifically, we analyze the differences in attention scores between successful and unsuccessful defenses to identify the attention heads sensitive to jailbreak attacks. During inference, we reallocate attention to emphasize users’ core intentions, minimizing interference from attack tokens. Our experimental results demonstrate that DeTAM outperforms various baselines in jailbreak defense and exhibits robust generalization across different attacks and models, maintaining its effectiveness even on in-the-wild jailbreak data. Furthermore, we compare DeTAM with the baselines on over-defense datasets, further validating its superior balance between helpfulness and harmlessness.

We present DeRAGEC, a method for improving Named Entity (NE) correction in Automatic Speech Recognition (ASR) systems. By extending the Retrieval-Augmented Generative Error Correction (RAGEC) framework, DeRAGEC employs synthetic denoising rationales to filter out noisy NE candidates before correction. By leveraging phonetic similarity and augmented definitions, it refines noisy retrieved NEs using in-context learning, requiring no additional training. Experimental results on CommonVoice and STOP datasets show significant improvements in Word Error Rate (WER) and NE hit ratio, outperforming baseline ASR and RAGEC methods. Specifically, we achieved a 28% relative reduction in WER compared to ASR without postprocessing.

LLMs are ideal for decision-making thanks to their ability to reason over long contexts. However, challenges arise when processing speech transcripts that describe complex scenarios, as they are verbose and include repetition, hedging, and vagueness. E.g., during a company’s earnings call, an executive might project a positive revenue outlook to reassure investors, despite uncertainty regarding future earnings. It is crucial for LLMs to incorporate this uncertainty systematically when making decisions. In this paper, we introduce DeFine, a modular framework that constructs probabilistic factor profiles from complex scenarios. It then integrates these profiles with analogical reasoning, leveraging insights from similar past experiences to guide LLMs in making critical decisions in new situations. Our framework separates the tasks of quantifying uncertainty and incorporating it into LLM decision-making. This approach is particularly useful in areas such as consulting and financial deliberation, where making decisions under uncertainty is vital.

Large Language Models (LLMs) are nowadays expected to generate content aligned with human preferences. Current work focuses on alignment at model training time, through techniques such as Reinforcement Learning with Human Feedback (RLHF). However, it is unclear if such methods are an effective choice to teach alignment objectives to the model. First, the inability to incorporate multiple, custom rewards and reliance on a model developer’s view of universal and static principles are key limitations. Second, the reliability of such approaches is also questionable (e.g. susceptibility to jailbreaking even after safety training). To address these issues, we propose DeAL, a framework that allows the user to customize reward functions and enables Decoding-time Alignment of LLMs (DeAL). At its core, we view decoding as a heuristic-guided search process and facilitate the use of a wide variety of alignment objectives. Our experiments with programmatic constraints such as keyword and length constraints, and abstract objectives such as harmlessness and helpfulness, show that we can DeAL with fine-grained trade-offs and improve adherence to alignment objectives. Lastly, we demonstrate that DeAL is largely complementary to existing alignment strategies, and can be effectively paired with RLHF and prompting techniques to achieve better alignment.

We introduce DavIR, a model-based data selection method for post-training Large Language Models. DavIR generalizes Reducible Holdout Loss to core-set selection problem of causal language modeling, and quantifies the learnability of a given datum with respect to a pre-trained LLM based on relative reduction in loss during fine-tuning, a metric we show to be closely related to the implicit reward model described in Direct Preference Optimization (DPO). We show that 6% of Alpaca dataset selected with DavIR can steer both the LLaMA and Gemma model family to produce superior performance compared to the same models trained on the full 52K dataset. We also show that Alpaca dataset compressed with DavIR can be combined with GSM8K dataset to effectively balance open-domain freeform QA and mathematical reasoning capabilities. Finally, we apply the DavIR objective to DPO and develop a normalized DavIR-DPO objective which improves alignment performance of Zephyr-7B-SFT model by 8% (relative) on AlpacaEval, compared against training on vanilla DPO objective.

Many sensitive domains — such as the clinical domain — lack widely available datasets due to privacy risks. The increasing generative capabilities of large language models (LLMs) have made synthetic datasets a viable path forward. In this study, we domain-adapt LLMs to the clinical domain and generate synthetic clinical texts that are machine-annotated with tags for personally identifiable information using capable encoder-based NER models. The synthetic corpora are then used to train synthetic NER models. The results show that training NER models using synthetic corpora incurs only a small drop in predictive performance. The limits of this process are investigated in a systematic ablation study — using both Swedish and Spanish data. Our analysis shows that smaller datasets can be sufficient for domain-adapting LLMs for data synthesis. Instead, the effectiveness of this process is almost entirely contingent on the performance of the machine-annotating NER models trained using the original data.

Conversational assistants are increasingly popular across diverse real-world applications, highlighting the need for advanced multimodal speech modeling. Speech, as a natural mode of communication, encodes rich user-specific characteristics such as speaking rate and pitch, making it critical for effective interaction. Our work introduces a data-centric customization approach for efficiently enhancing multimodal understanding in conversational speech modeling. Central to our contributions is a novel multi-task learning paradigm that involves designing auxiliary tasks to utilize a small amount of speech data. Our approach achieves state-of-the-art performance on the Spoken-SQuAD benchmark, using only 10% of the training data with open-weight models, establishing a robust and efficient framework for audio-centric conversational modeling. We also introduce ASK-QA, the first dataset for multi-turn spoken dialogue with ambiguous user requests and dynamic evaluation inputs.

Fine-tuning large language models (LLMs) on task-specific data is essential for their effective deployment. As dataset sizes grow, efficiently selecting optimal subsets for training becomes crucial to balancing performance and computational costs. Traditional data selection methods often require fine-tuning a scoring model on the target dataset, which is time-consuming and resource-intensive, or rely on heuristics that fail to fully leverage the model’s predictive capabilities. To address these challenges, we propose Data Whisperer, an efficient, training-free, attention-based method that leverages few-shot in-context learning with the model to be fine-tuned. Comprehensive evaluations were conducted on both raw and synthetic datasets across diverse tasks and models. Notably, Data Whisperer achieves superior performance compared to the full GSM8K dataset on the Llama-3-8B-Instruct model, using just 10% of the data, and outperforms existing methods with a 3.1-point improvement and a 7.4× speedup.

Our quality audit for three widely used public multilingual speech datasets Mozilla Common Voice 17.0, FLEURS, and VoxPopuli shows that in some languages, these datasets suffer from significant quality issues. We believe addressing these issues will make these datasets more useful as evaluation sets, and improve downstream models. We divide these quality issues into two categories: micro-level and macro-level. We find that macro-level issues are more prevalent in less institutionalized, often under-resourced languages. We provide a case analysis of Taiwanese Southern Min (nan_tw) that highlights the need for proactive language planning (e.g. orthography prescriptions, dialect boundary definition) and enhanced data quality control in the process of Automatic Speech Recognition (ASR) dataset creation. We conclude by proposing guidelines and recommendations to mitigate these issues in future dataset development, emphasizing the importance of sociolinguistic awareness in creating robust and reliable speech data resources.

In this paper, we show that knowledge distillation can be subverted to manipulate language model benchmark scores, revealing a critical vulnerability in current evaluation practices. We introduce “Data Laundering,” a process that enables the covert transfer of benchmark-specific knowledge through seemingly legitimate intermediate training steps. Through extensive experiments with a 2-layer BERT student model, we show how this approach can achieve substantial improvements in benchmark accuracy (up to 75% on GPQA) without developing genuine reasoning capabilities. Notably, this method can be exploited intentionally or even unintentionally, as researchers may inadvertently adopt this method and inflate scores without realising the implications. While our findings demonstrate the effectiveness of this technique, we present them as a cautionary tale highlighting the urgent need for more robust evaluation methods in AI. This work aims to contribute to the ongoing discussion about evaluation integrity in AI development and the need for benchmarks that more accurately reflect true model capabilities. The code is available at https://github.com/mbzuai-nlp/data_laundering.

Large Language Model (LLM)-based agents have excelled in various domains but face significant challenges when applied to data science workflows due to their complex, multi-stage nature. Current LLM-based agents struggle with non-linear relationships, recursive dependencies, implicit data- and logic-dependent reasoning, and managing extensive context. In this paper, we introduce Data Interpreter, an LLM-based agent that addresses these challenges through hierarchical graph-based modeling to represent the complexity and a progressive strategy for step-by-step verification, refinement, and consistent context management. Extensive experiments confirm the effectiveness of Data Interpreter. On InfiAgent-DABench, it boosts performance by 25% (from 75.9% to 94.9%), and on machine learning and open-ended tasks, it lifts accuracy from 88% to 95% and from 60% to 97%, respectively. Moreover, our method surpasses state-of-the-art baselines by 26% on the MATH dataset. We will release the code upon publication.

With a combination of quantitative experiments, human judgments, and qualitative analyses, we evaluate the quantity and quality of African American Language (AAL) representation in 12 predominantly English, open-source pretraining corpora. We specifically focus on the sources, variation, and naturalness of included AAL texts representing the AAL speaking community. We find that AAL is underrepresented in all evaluated pretraining corpora compared to US demographics, constituting as few as 0.007% and at most 0.18% of documents. We also find that more than 25% of AAL texts in C4 may be perceived as inappropriate for LLMs to generate and to reinforce harmful stereotypes. Finally, we find that most automated filters are more likely to conserve White Mainstream English (WME) texts over AAL in pretraining corpora.

Multimodal Aspect-Based Sentiment Analysis (MABSA) aims to extract aspect-sentiment pairs from text and image data. While significant progress has been made in image-aspect alignment, due to the subtlety and complexity of language expressions, there are not always explicit aspect words in the language to align with images. Existing methods typically assume a direct alignment between images and aspects, matching the entire image with a corresponding aspect. This rough alignment of images and aspects introduces noise. To address the above issues, this paper proposes a Dual-Aware Enhanced Alignment Network (DaNet) designed for fine-grained multimodal aspect-image alignment and denoising. Specifically, we first introduce a Multimodal Denoising Encoder (MDE) that jointly image and text to guide the compression and denoising of visual sequences. And then, aspect-aware and sentiment-aware networks are constructed to jointly enhance fine-grained alignment and denoising of text-image information. To better align implicit aspects, an Implicit Aspect Opinion Generation (IAOG) pretraining is designed under the guidance of large language model. Extensive experiments across three MABSA subtasks demonstrate that DaNet outperforms existing methods. Code will be available at https://github.com/***/DaNet.

LLMs face privacy risks when handling sensitive data. To ensure privacy, researchers use differential privacy (DP) to provide protection by adding noise during LLM training. However, users may be hesitant to share complete data with LLMs. Researchers follow local DP to sanitize the text on the user side and feed non-sensitive text to LLMs. The sanitization usually uses a fixed non-sensitive token list or a fixed noise distribution, which induces the risk of being attacked or semantic distortion. We argue that the token’s protection level should be adaptively adjusted according to its semantic-based information to balance the privacy-utility trade-off. In this paper, we propose DYNTEXT, an LDP-based Dynamic Text sanitization for privacy-preserving LLM inference, which dynamically constructs semantic-aware adjacency lists of sensitive tokens to sample non-sensitive tokens for perturbation. Specifically, DYNTEXT first develops a semantic-based density modeling under DP to extract each token’s density information. We propose token-level smoothing sensitivity by combining the idea of global sensitivity (GS) and local sensitivity (LS), which dynamically adjusts the noise scale to avoid excessive noise in GS and privacy leakage in LS. Then, we dynamically construct an adjacency list for each sensitive token based on its semantic density information. Finally, we apply the replacement mechanism to sample non-sensitive, semantically similar tokens from the adjacency list to replace sensitive tokens. Experiments show that DYNTEXT excels strong baselines on three datasets.