China

As large language models (LLMs) are increasingly applied across various domains, enhancing safety while maintaining the helpfulness of LLMs has become a critical challenge. Recent studies solve this problem through safety-constrained online preference optimization or safety-constrained offline preference optimization. However, the safety-constrained online methods often suffer from excessive safety, which might reduce helpfulness, while the safety-constrained offline methods perform poorly in adaptively balancing safety and helpfulness. To address these limitations, we propose MidPO, a Mixture of Experts (MoE) framework for safety-helpfulness dual Preference Optimization. Firstly, MidPO devises single-preference enhanced direct preference optimization approach to transform the base model into two independent experts, termed safety and helpfulness experts, and fine-tunes the two independent experts for optimal safety or helpfulness performance. Secondly, to achieve an effective balance between safety and helpfulness, MidPO incorporates the two experts into the MoE framework and designs a dynamic routing mechanism to allocate contributions from each expert adaptively. We conduct quantitative and qualitative experiments on three popular datasets to demonstrate the proposed MidPO significantly outperforms state-of-the-art approaches in both safety and helpfulness. Code is available at https://shorturl.at/BmNa8.

EMNLP 2025

MidPO: Dual Preference Optimization for Safety and Helpfulness in Large Language Models via a Mixture of Experts Framework

llm safety alignment

ai security

large language model

mixture of experts

### Welcome!
The 62nd Annual Meeting of the Association for Computational Linguistics (ACL 2024) will take place in Bangkok, Thailand from August 11th to 16th, 2024. Our Virtual Poster Sessions will take place online Thursday, August 22, 2024.

You are required to register for this event. **Please register [here](https://2024.aclweb.org/registration). **

If you have already registered, please check your inbox for an email from Underline granting you access to ACL 2024 content.

Please register!

ACL 2024

The 62nd Annual Meeting of the Association for Computational Linguistics (ACL 2024) will take place in Bangkok, Thailand from August 11th to 16th, 2024. More information will be announced soon.

poster

## Welcome!
"I am excited to welcome you to this year’s edition of the Conference on Empirical Methods in Natural Language Processing! Importantly, it marks the 30th edition of EMNLP. With over 8,000 submissions, more than 3,000 accepted papers, and thousands of attendees, we have come a long way from that first
workshop, which had 14 accepted papers. As the field looks ahead, Suzhou is the fitting location for celebrating this milestone: rooted in a long literary tradition, yet modern and forward-looking, and home to a large share of the NLP community." 

*Message from the General Chair, Dirk Hovy*

Celebrate 30 Years of EMNLP! 
EMNLP 2025 will be held in Suzhou, China from November 5th to November 9th, 2025.

We introduce FreqRank, a mutation-based defense to localize malicious components in LLM outputs and their corresponding backdoor triggers. FreqRank assumes that the malicious sub-string(s) consistently appear in outputs for triggered inputs and uses a frequency-based ranking system to identify them. Our ranking system then leverages this knowledge to localize the backdoor triggers present in the inputs. We train six malicious models for three downstream tasks, namely, code completion (CC), code generation (CG), and code summarization (CS), and show that they have an average attack success rate (ASR) of 80.9%. Furthermore, FreqRank’s ranking system highlights the malicious outputs as one of the top five suggestions in over 99.0% of cases. We also demonstrate that FreqRank is capable of localizing the backdoor trigger effectively even with a limited number of triggered samples. Finally, we show that our approach is 40-50% more effective than other defense methods.

Localizing Malicious Outputs from CodeLLM

Modern ASR systems are increasingly deployed in high-stakes settings, including clinical interviews, public services, and educational tools, where equitable performance across speaker groups is essential. While pre-trained models like Whisper achieve strong overall accuracy, they often exhibit inconsistent group-level performance that varies across domains. These disparities are not fixed properties of the model, but emerge from the interaction between model, data, and task—posing challenges for fairness interventions designed in-domain. We frame fairness in ASR as a generalisation problem. We fine-tune Whisper on a diverse corpus using four strategies: standard fine-tuning, demographic rebalancing, gender-swapped data augmentation, and a novel contrastive learning objective that encourages gender-invariant representations. We evaluate performance across multiple aspects of fairness and utility, both in-domain and on three out-of-domain test sets: LibriSpeech, EdAcc, and CognoSpeak. Our findings show that the method with the best in-domain fairness performed worst out-of-domain, illustrating that fairness gains do not always generalise. Demographic balancing generalises more consistently, while our contrastive method offers a practical alternative: it achieves stable, cross-domain fairness improvements without requiring changes to the training data distribution, and with minimal accuracy trade-offs.

Fairness in Automatic Speech Recognition Isn’t a One-Size-Fits-All

Large Language Models (LLMs) often exhibit tendencies that diverge from human preferences, such as favoring certain writing styles or producing verbose outputs. While crucial for improvement, identifying the factors driving these misalignments remains challenging due to existing evaluation methods' reliance on coarse-grained comparisons and lack of explainability. To address this, we introduce PROFILE, an automated framework to uncover and measure the alignment of factor-level preferences of humans and LLMs. Using PROFILE, we analyze preference alignment across summarization, instruction-following, and document-based question-answering tasks. We find a significant discrepancy: while LLMs show poor factor-level alignment with human preferences when generating texts, they demonstrate strong alignment in evaluation tasks. We demonstrate how leveraging the identified generation-evaluation gap can be used to improve LLM alignment through multiple approaches, including fine-tuning with self-guidance.

Uncovering Factor-Level Preference to Improve Human-Model Alignment

Offline preference optimization methods are efficient for large language models (LLMs) alignment. Direct Preference optimization (DPO)-like learning, one of the most popular approaches, stands out for its efficiency in reward modeling. However, these methods typically follow the convention to use Bradley-Terry (BT) reward modeling that faces several critical assumptions, including the requirement for pairwise training data, model distribution shifting, human rationality assumption, etc. To address these limitations, we propose a general framework for offline preference optimization methods, Adaptive Preference Optimization with Utility Anchor (UAPO), which introduces an anchoring function to estimate the uncertainties brought from preference data annotation. Our method enables training even in scenarios where the data is unpaired, significantly enhancing data utilization efficiency. Moreover, the anchor design makes UAPO more robust in the training process. Experimental results demonstrate that UAPO achieves competitive outcomes without the strict dependency on data pairing, paving the way for more flexible and effective preference optimization methods.

Adaptive Preference Optimization with Uncertainty-aware Utility Anchor

Retrieval-Augmented Generation (RAG) enhances the accuracy of Large Language Model (LLM) responses by leveraging relevant external documents during generation. Although previous studies noted that retrieving many documents can degrade performance, they did not isolate how the quantity of documents affects performance while controlling for context length. We evaluate various language models on custom datasets derived from a multi-hop QA task. We keep the context length and position of relevant information constant while varying the number of documents, and find that increasing the document count in RAG settings poses significant challenges for most LLMs, reducing performance by up to 20%. However, Qwen2 maintained consistent results across increasing document counts, indicating better multi-document handling capability. Finally, our results indicate that processing multiple documents is a separate challenge from handling long contexts. We will publicly release the datasets and code upon publication to facilitate further research in multi-document retrieval.

Stop Playing the Guessing Game! Evaluating Conversational Recommender Systems via Target-free User Simulation

Large language models (LLMs) have shown remarkable capabilities across various tasks, that are learned from massive amounts of text-based data. Although LLMs can control output sequence length, particularly in instruction-based settings, the internal mechanisms behind this control have been unexplored yet. In this study, we provide empirical evidence on how output sequence length information is encoded within the internal representations in LLMs. In particular, our findings show that multi-head attention mechanisms are critical in determining output sequence length, which can be adjusted in a disentangled manner. By scaling specific hidden units within the model, we can control the output sequence length without losing the informativeness of the generated text, thereby indicating that length information is partially disentangled from semantic information. Moreover, some hidden units become increasingly active as prompts become more length-specific, thus reflecting the model's internal awareness of this attribute. Our findings suggest that LLMs have learned robust and adaptable internal mechanisms for controlling output length without any external control.

Length Representations in Large Language Models

Persona-prompting is a growing strategy to steer LLMs toward simulating particular perspectives or linguistic styles through the lens of a specified identity. While this method is often used to personalize outputs, its impact on how LLMs represent social groups remains underexplored. In this paper, we investigate whether persona-prompting leads to different levels of linguistic abstraction—an established marker of stereotyping—when generating short texts linking socio-demographic categories with stereotypical or non-stereotypical attributes. Drawing on the Linguistic Expectancy Bias framework, we analyze outputs from six open-weight LLMs under three prompting conditions, comparing 11 persona-driven responses to those of a generic AI assistant. To support this analysis, we introduce Self-Stereo, a new dataset of self-reported stereotypes from Reddit. We measure abstraction through three metrics: concreteness, specificity, and negation. Our results highlight the limits of persona-prompting in modulating abstraction in language, confirming criticisms about the ecology of personas as representative of socio-demographic groups and raising concerns about the risk of propagating stereotypes even when seemingly evoking the voice of a marginalized groups.

Simulating Identity, Propagating Bias: Abstraction and Stereotypes in LLM-Generated Text

Recent advancements in Large Language Models (LLMs) have transformed natural language understanding and generation, leading to extensive benchmarking across diverse tasks. However, cryptanalysis—a critical area for data security and encryption—has not yet been thoroughly explored in LLM evaluations. To address this gap, we present the first study evaluating the cryptanalytic potential of state-of-the-art LLMs on encrypted texts generated using a range of cryptographic algorithms. We introduce a novel benchmark dataset comprising diverse plain texts—spanning various domains, lengths, writing styles, and topics—paired with their encrypted versions. Using a zero-shot approach, we assess multiple LLMs for both decryption accuracy and semantic comprehension across different encryption schemes. Our findings reveal key insights into the strengths and limitations of LLMs in side-channel communication, while raising concerns about their susceptibility to jailbreaking attacks. This research highlights the dual-use nature of LLMs in security contexts and contributes to the ongoing discussion on AI safety and security.

Benchmarking Large Language Models for Cryptanalysis and Side-Channel Vulnerabilities

Large language models (LLMs) have demonstrated remarkable performance in zero-shot dialogue state tracking (DST), reducing the need for task-specific training. However, conventional DST benchmarks primarily focus on structured user-agent conversations, failing to capture the complexities of real-world multi-user interactions. In this study, we assess the robustness of LLMs in multi-user DST while minimizing dataset construction costs. Inspired by recent advances in LLM-based data annotation, we extend an existing DST dataset by generating utterances of a second user based on speech act theory. Our methodology systematically incorporates a second user’s utterances into conversations, enabling a controlled evaluation of LLMs in multi-user settings. Experimental results reveal a significant performance drop compared to single-user DST, highlighting the limitations of current LLMs in extracting and tracking dialogue states amidst multiple speakers. Our findings emphasize the need for future research to enhance LLMs for multi-user DST scenarios, paving the way for more realistic and robust DST models.

Downloads

Next from EMNLP 2025

Localizing Malicious Outputs from CodeLLM

Similar lecture

From Representational Harms to Quality-of-Service Harms: A Case Study on Llama 2 Safety Safeguards