Austria

Minimum Bayes Risk (MBR) decoding optimizes output selection by maximizing the expected utility value of an underlying human distribution. While prior work has shown the effectiveness of MBR decoding through empirical evaluation, few studies have analytically investigated why the method is effective. As a result of our analysis, we show that, given the size $n$ of the reference hypothesis set used in computation, MBR decoding approaches the optimal solution with high probability at a rate of $\mathcal{O}(n^{-\frac{1}{2}})$, under certain assumptions, even though the language space $\mathcal{Y}$ is significantly larger $|\mathcal{Y}| \gg n$.
This result helps to theoretically explain the strong performance observed in several prior empirical studies on MBR decoding. In addition, we provide the performance gap for maximum-a-posteriori (MAP) decoding and compare it to MBR decoding. The result of this paper indicates that MBR decoding tends to converge to the optimal solution faster than MAP decoding in several cases.

ACL 2025

Theoretical Guarantees for Minimum Bayes Risk Decoding

efficient inference for mt

mt theory

theory

### Welcome to The 63rd Annual Meeting of the Association for Computational Linguistics (ACL 2025)

Message from the General Chair: 
*It is my great pleasure and honor to welcome you to the 63rd Annual Meeting of the Association for Computational Linguistics (ACL 2025), held in beautiful Vienna, Austria, from July 27 to August 1, 2025. ACL2025continues our field’s tradition of excellence in scholarship, innovation, and inclusivity, and I am deeply grateful to the many volunteers who have worked tirelessly to bring this event to life.* 
[Read more](https://drive.google.com/file/d/1GI_hvOpjswAuYdUTromfeDiPpCcqidwg/view?usp=sharing)

To access this event page, you need to log in with the **email address you registered with**. Access credentials will be sent to your email from Underline - subject line "Welcome to ACL 2025". Please be sure to check your spam email folder if you do not see an email confirmation right away.

Please log in

To access this event page, you need to log in with the **email address you registered with**. 

Welcome to The 63rd Annual Meeting of the Association for Computational Linguistics (ACL 2025)

poster

During the preference optimization of large language models (LLMs), distribution shifts may arise between newly generated model samples and the data used to train the reward model (RM). This shift reduces the efficacy of the RM, which in turn negatively impacts the performance of the policy model (PM). To address this challenge, we propose $\textbf{Mutual-Taught}$, a self-training method that iteratively improves both the PM and RM without requiring additional human annotation. Our approach mirrors the expectation-maximization (EM) algorithm. In the E-step, the PM is updated using feedback from the current RM, guiding the PM toward a better approximation of the latent optimal preference distribution.
In the M-step, we update the RM by constructing training data from the outputs of the PM before and after the E-step update. This process ensures that the RM adapts to the evolving policy distribution. Experimental results demonstrate that this iterative approach leads to consistent improvements in both models. Specifically, our 8B policy model, LLaMA-3-8B-Instruct-MT, achieves a length-controlled win rate of 54.1\% on AlpacaEval-2, while our 8B reward model, FsfairX-LLaMA3-RM-MT, performs on par with GPT-4o-2024-08-06 on RewardBench.

Mutual-Taught for Co-adapting Policy and Reward Models

Natural language processing applications are increasingly prevalent, but social biases in their outputs remain a critical challenge. While various bias evaluation methods have been proposed, these assessments show unexpected instability when input texts undergo minor stylistic changes. This paper conducts a comprehensive analysis of how different style transformations impact bias evaluation results across multiple language models and bias types using causal inference techniques. Our findings reveal that formality transformations significantly affect bias scores, with informal style showing substantial bias reductions (up to 8.33% in LLaMA-2-13B). We identify appearance bias, sexual orientation bias, and religious bias as most susceptible to style changes, with variations exceeding 20%. Larger models demonstrate greater sensitivity to stylistic variations, with bias measurements fluctuating up to 3.1% more than in smaller models. These results highlight critical limitations in current bias evaluation methods and emphasize the need for reliable and fair assessments of language models.

Unmasking Style Sensitivity: A Causal Analysis of Bias Evaluation Instability in Large Language Models

In English education tutoring, teacher feedback is essential for guiding students. Recently, AI-based tutoring systems have emerged to assist teachers; however, these systems require high-quality and large-scale teacher feedback data, which is both time-consuming and costly to generate manually. In this study, we propose FEAT, a cost-effective framework for generating teacher feedback, and have constructed three complementary datasets: (1) DIRECT-Manual (DM), where both humans and large language models (LLMs) collaboratively generate high-quality teacher feedback, albeit at a higher cost; (2) DIRECT-Generated (DG), an LLM-only generated, cost-effective dataset with lower quality;, and (3) DIRECT-Augmented (DA), primarily based on DG with a small portion of DM added to enhance quality while maintaining cost-efficiency. Experimental results showed that incorporating a small portion of DM (5–10%) into DG leads to superior performance compared to using 100% DM alone.

FEAT: A Preference Feedback Dataset through a Cost-Effective Auto-Generation and Labeling Framework for English AI Tutoring

We introduce a novel framework for consolidating multi-turn adversarial “jailbreak” prompts into single-turn queries, significantly reducing the manual overhead required for adversarial testing of large language models (LLMs). While multi-turn human jailbreaks have been shown to yield high attack success rates (ASRs), they demand considerable human effort and time. Our proposed Multi-turn-to-Single-turn (M2S) methods—Hyphenize, Numberize, and Pythonize—systematically reformat multi-turn dialogues into structured single-turn prompts. Despite eliminating iterative back-and-forth interactions, these reformatted prompts preserve and often enhance adversarial potency: in extensive evaluations on the Multi-turn Human Jailbreak (MHJ) dataset, M2S methods yield ASRs ranging from 70.6 % to 95.9 % across various state-of-the-art LLMs. Remarkably, our single-turn prompts outperform the original multi-turn attacks by up to 17.5 % in absolute ASR, while reducing token usage by more than half on average. Further analyses reveal that embedding malicious requests in enumerated or code-like structures exploits “contextual blindness,” undermining both native guardrails and external input-output safeguards. By consolidating multi-turn conversations into efficient single-turn prompts, our M2S framework provides a powerful tool for large-scale red-teaming and exposes critical vulnerabilities in contemporary LLM defenses. All code, data, and conversion prompts are available for reproducibility and further investigations: https://github.com/Junuha/M2S_DATA

One-Shot is Enough: Consolidating Multi-Turn Attacks into Efficient Single-Turn Prompts for LLMs

Large language models (LLMs) have demonstrated impressive instruction following capabilities, while still struggling to accurately manage the length of the generated text, which is a fundamental requirement in many real-world applications. Existing length control methods involve fine-tuning the parameters of LLMs, which is inefficient and suboptimal for practical use. In this paper, we propose a novel iterative sampling framework for text length control, integrating the Metropolis-Hastings algorithm with an importance sampling acceleration strategy. This framework efficiently and reliably regulates LLMs to generate length-constrained text without modifying the underlying parameters, thereby preserving the original capabilities of LLMs. Experimental results demonstrate that our framework achieves almost 100% success rates of length control on Llama3.1 for tasks such as length-controlled abstractive summarization and length-constrained instruction following, with minimal additional computational overhead. This also highlights the significant potential of our method for precise length control across a broader range of applications, without compromising the versatility of LLMs.

Length Controlled Generation for Black-box LLMs

Retrieval-augmented generation (RAG) systems often struggle with narrative-rich documents and event-centric reasoning, particularly when synthesizing information across multiple sources. We present EventRAG, a novel framework that enhances text generation through structured event representations. We first construct an Event Knowledge Graph by extracting events and merging semantically equivalent nodes across documents, while expanding under-connected relationships. We then employ an iterative retrieval and inference strategy that explicitly captures temporal dependencies and logical relationships across events. Experiments on UltraDomain and MultiHopRAG benchmarks show EventRAG's superiority over baseline RAG systems, with substantial gains in generation effectiveness, logical consistency, and multi-hop reasoning accuracy. Our work advances RAG systems by integrating structured event semantics with iterative inference, particularly benefiting scenarios requiring temporal and logical reasoning across documents.

EventRAG: Enhancing LLM Generation with Event Knowledge Graphs

Chain-of-Though (CoT) represents a common strategy for reasoning in Large Language Models (LLMs) by decomposing complex tasks into intermediate inference steps. However, explanations generated via CoT are susceptible to content biases that negatively affect their robustness and faithfulness. To mitigate existing limitations, recent work has proposed using logical formalisms coupled with external symbolic solvers. However, fully symbolic approaches possess the bottleneck of requiring a complete translation from natural language to formal languages, a process that affects efficiency and flexibility. To achieve a trade-off, this paper investigates methods to disentangle content from logical reasoning without a complete formalisation. In particular, we present QuaSAR (for Quasi-Symbolic Abstract Reasoning), a variation of CoT that guides LLMs to operate at a higher level of abstraction via quasi-symbolic explanations. Our framework leverages the capability of LLMs to formalise only relevant variables and predicates, enabling the coexistence of symbolic elements with natural language. We show the impact of QuaSAR for in-context learning and for constructing demonstrations to improve the reasoning capabilities of smaller models. Our experiments show that quasi-symbolic abstractions can improve CoT-based methods by up to 8% accuracy, enhancing robustness and consistency on challenging adversarial variations on both natural language (i.e. MMLU-Redux) and symbolic reasoning tasks (i.e., GSM-Symbolic).

Improving Chain-of-Thought Reasoning via Quasi-Symbolic Abstractions

Demographics and cultural background of annotators influence the labels they assign in text annotation – for instance, an elderly woman might find it offensive to read a message addressed to a "bro", but a male teenager might find it appropriate. It is therefore important to acknowledge label variations to not under-represent members of a society. Two research directions developed out of this observation in the context of using large language models (LLM) for data annotations, namely (1) studying biases and inherent knowledge of LLMs and (2) injecting diversity in the output by manipulating the prompt with demographic information. We combine these two strands of research and ask the question to which demographics an LLM resorts to when no demographics is given. To answer this question, we evaluate which attributes of human annotators LLMs inherently mimic. Furthermore, we compare non-demographic conditioned prompts and placebo-conditioned prompts (e.g., "you are an annotator who lives in house number 5") to demographics-conditioned prompts ("You are a 45 year old man and an expert on politeness annotation. How do you rate {instance}"). We study these questions for politeness and offensiveness annotations on the POPQUORN data set, a corpus created in a controlled manner to investigate human label variations based on demographics which has not been used for LLM-based analyses so far. We observe notable influences related to gender, race, and age in demographic prompting, which contrasts with previous studies that found no such effects.

Which Demographics do LLMs Default to During Annotation?

This paper explores the problem of commonsense level vision-knowledge conflict in Multimodal Large Language Models (MLLMs), where visual information contradicts model's internal commonsense knowledge. To study this issue, we introduce an automated framework, augmented with human-in-the-loop quality control, to generate inputs designed to simulate and evaluate these conflicts in MLLMs. Using this framework, we have crafted a diagnostic benchmark consisting of 374 original images and 1,122 high-quality question-answer (QA) pairs. The benchmark covers two aspects of conflict and three question types, providing a thorough assessment tool. We apply this benchmark to assess the conflict-resolution capabilities of nine representative MLLMs from various model families. Our results indicate an evident over-reliance on parametric knowledge for approximately 20\% of all queries, especially among Yes-No and action-related problems. Based on these findings, we evaluate the effectiveness of existing approaches to mitigating the conflicts and compare them to our ``Focus-on-Vision'' prompting strategy. Despite some improvement, the vision-knowledge conflict remains unresolved and can be further scaled through our data construction framework. Our proposed framework, benchmark, and analysis contribute to the understanding and mitigation of vision-knowledge conflicts in MLLMs.

Insight Over Sight: Exploring the Vision-Knowledge Conflicts in Multimodal LLMs

While multilingual language models like XLM-R have advanced multilingualism in NLP, they still perform poorly in extremely low-resource languages. This situation is exacerbated by the fact that modern LLMs such as LLaMA and Qwen support far fewer languages than XLM-R, making text generation models non-existent for many languages in the world. To tackle this challenge, we propose a novel framework for adapting multilingual encoders to text generation in extremely low-resource languages. By reusing the weights between the encoder and the decoder, our framework allows the model to leverage the learned semantic space of the encoder, enabling efficient learning and effective generalization in low-resource languages. Applying this framework to four Chinese minority languages, we present XLM-SWCM, and demonstrate its superior performance on various downstream tasks even when compared with much larger models.

Downloads

Next from ACL 2025

Mutual-Taught for Co-adapting Policy and Reward Models

Similar lecture

Comprehensive Analysis of Minimum Bayes Risk Decoding through Bias and Diversity Decomposition