China

As large language models (LLMs) are increasingly deployed for complex reasoning tasks, Long Chain-of-Thought (Long-CoT) prompting has emerged as a key paradigm for structured inference. Despite early-stage safeguards enabled by alignment techniques such as RLHF, we identify a previously underexplored vulnerability: reasoning trajectories in Long-CoT models can drift from aligned paths, resulting in content that violates safety constraints. We term this phenomenon Path Drift. Through empirical analysis, we uncover three behavioral triggers of Path Drift: (1) first-person commitments that induce goal-driven reasoning that delays refusal signals; (2) ethical evaporation, where surface-level disclaimers bypass alignment checkpoints; (3) condition chain escalation, where layered cues progressively steer models toward unsafe completions. Building on these insights, we introduce a three-stage Path Drift Induction Framework comprising cognitive load amplification, self-role priming, and condition chain hijacking. Each stage independently reduces refusal rates, while their combination further compounds the effect. To mitigate these risks, we propose a path-level defense strategy incorporating role attribution correction and metacognitive reflection (reflective safety cues). Our findings highlight the need for trajectory-level alignment oversight in long-form reasoning beyond token-level alignment.

EMNLP 2025

Path Drift in Large Reasoning Models: How First-Person Commitments Override Safety

red teaming; security and privacy; prompting; robustness;

## Welcome to EMNLP 2024! 
We are excited to welcome you to one of the most prominent conferences in the field of Natural Language Processing. This year, EMNLP 2024 is being held in a hybrid format,
offering both virtual and in-person participation in beautiful Miami. Due to a record-breaking number of submissions, we've expanded the total number of accepted papers to accommodate more cutting-edge research from around the globe.
### [Conference Handbook](https://drive.google.com/file/d/1WPROgxjLAC96AJL7Ugy0tEnYm7dkrbHt/view?usp=sharing)

You are required to register for this event. **Please register [here](https://2024.emnlp.org/registration/).** The EMNLP 2024 event page on Underline will be open to public one week prior to the event.

Please register!

EMNLP 2024

EMNLP 2024 will take place in Miami, Florida from Nov 12th to Nov 16th, 2024, at the Hyatt Regency Miami Hote and on Underline for remote participants.

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.

Multimodal large language models (MLLMs) have demonstrated extraordinary capabilities in conducting conversations based on image inputs. However, we observe that MLLMs exhibit a pronounced form of visual sycophantic behavior. While similar behavior has also been noted in text-based large language models (LLMs), it becomes significantly more prominent when MLLMs process image inputs. We refer to this phenomenon as the "sycophantic modality gap." To better understand this issue, we further analyze the factors that contribute to the exacerbation of this gap. To mitigate the visual sycophantic behavior, we first experiment with naive supervised fine-tuning to help the MLLM resist misleading instructions from the user. However, we find that this approach also makes the MLLM overly resistant to corrective instructions (i.e., stubborn even if it is wrong). To alleviate this trade-off, we propose Sycophantic Reflective Tuning (SRT), which enables the MLLM to engage in reflective reasoning, allowing it to determine whether a user's instruction is misleading or corrective before drawing a conclusion. After applying SRT, we observe a significant reduction in sycophantic behavior toward misleading instructions, without resulting in excessive stubbornness when receiving corrective instructions.

Pointing to a Llama and Call it a Camel: On the Sycophancy of Multimodal Large Language Models

Subjectivity in NLP tasks, _e.g._, toxicity classification, has emerged as a critical challenge precipitated by the increased deployment of NLP systems in content-sensitive domains. Conventional approaches aggregate annotator judgements (labels), ignoring minority perspectives, and overlooking the influence of the sociocultural context behind such annotations. We propose a framework where subjectivity in binary labels is modeled as an empirical distribution accounting for the variation in annotators through human values extracted from sociocultural descriptors using a language model. The framework also allows for downstream tasks such as population and sociocultural group-level majority label prediction. Experiments on three toxicity datasets covering human-chatbot conversations and social media posts annotated with diverse annotator pools demonstrate that our approach yields well-calibrated toxicity distribution predictions across binary toxicity labels, which are further used for majority label prediction across cultural subgroups, improving over existing methods.

Learning Subjective Label Distributions via Sociocultural Descriptors

Generative Large Language Models (LLMs) infer user's demographic information from subtle cues in the conversation --- a phenomenon called implicit personalization. Prior work has shown that such inferences can lead to lower quality responses for users assumed to be from minority groups, even when no demographic information is explicitly provided. In this work, we systematically explore how LLMs respond to stereotypical cues using controlled synthetic conversations, by analyzing the models' latent user representations through both model internals and generated answers to targeted user questions. Our findings reveal that LLMs do infer demographic attributes based on these stereotypical signals, which for a number of groups even persists when the user explicitly identifies with a different demographic group. Finally, we show that this form of stereotype-driven implicit personalization can be effectively mitigated by intervening on the model's internal representations using a trained linear probe to steer them toward the explicitly stated identity. Our results highlight the need for greater transparency and control in how LLMs represent user identity.

Reading Between the Prompts: How Stereotypes Shape LLM’s Implicit Personalization

Effective content moderation systems require explicit classification criteria, yet online communities like subreddits often operate with diverse, implicit standards. This work introduces a novel approach to identify and extract these implicit criteria from historical moderation data using an interpretable architecture. We represent moderation criteria as score tables of lexical expressions associated with content removal, enabling systematic comparison across different communities.Our experiments demonstrate that these extracted lexical patterns effectively replicate the performance of neural moderation models while providing transparent insights into decision-making processes. The resulting criteria matrix reveals significant variations in how seemingly shared norms are actually enforced, uncovering previously undocumented moderation patterns including community-specific tolerances for language, features for topical restrictions, and underlying subcategories of the toxic speech classification.

Decoding the Rule Book: Extracting Hidden Moderation Criteria from Reddit Communities

End-to-end automatic speech recognition systems often fail to transcribe domain-speciffcnamed entities, causing catastrophic failuresin downstream tasks. Numerous fast and lightweight named entity correction (NEC) models have been proposed in recent years. These models, mainly leveraging phonetic-level edit distance algorithms, have shown impressive performances. However, when theforms of the wrongly-transcribed words(s) and the ground-truth entity are signiffcantly different, these methods often fail to locate the wrongly transcribed words in hypothesis, thus limiting their usage. We propose a novel NEC method that utilizes speech sound features to retrieve candidate entities. With speech sound features and candidate entities, we inovatively design a generative method to annotate entityerrors in ASR transcripts and replace the textwith correct entities. This method is effective inscenarios of word form difference. We test ourmethod using open-source and self-constructed test sets. The results demonstrate that our NEC method can bring signiffcant improvement to entity accuracy. We will open source our self constructed test set and training data.

Generative Annotation for ASR Named Entity Correction

Temporal evolution attribute graph prediction, a key task in graph machine learning, aims to forecast the dynamic evolution of node attributes over time. While recent advances in Large Language Models (LLMs) have enabled their use in enhancing node representations for integration with Graph Neural Networks (GNNs), their potential to directly perform GNN-like aggregation and interaction remains underexplored. Furthermore, traditional approaches to initializing attribute embeddings often disregard structural semantics, limiting the provision of rich prior knowledge to GNNs. Current methods also primarily focus on 1-hop neighborhood aggregation, lacking the capability to capture complex structural interactions. To address these limitations, we propose a novel prediction framework that integrates structural information into attribute embeddings through the introduction of an attribute embedding loss. We design specialized prompts to enable LLMs to perform GNN-like aggregation and incorporate a relation-aware Graph Convolutional Network to effectively capture long-range and complex structural dependencies. Extensive experiments on multiple real-world datasets validate the effectiveness of our approach, demonstrating significant improvements in predictive performance over existing methods.

LGA: LLM-GNN Aggregation for Temporal Evolution Attribute Graph Prediction

Hallucination has emerged as a significant barrier to the effective application of Large Language Models (LLMs). In this work, we introduce a novel Attention-Guided SElf-Reflection (AGSER) approach for zero-shot hallucination detection in LLMs. The AGSER method utilizes attention contributions to categorize the input query into attentive and non-attentive queries. Each query is then processed separately through the LLMs, allowing us to compute consistency scores between the generated responses and the original answer. The difference between the two consistency scores serves as a hallucination estimator. In addition to its efficacy in detecting hallucinations, AGSER notably reduces computational complexity, requiring only three passes through the LLM and utilizing two sets of tokens. We have conducted extensive experiments with four widely-used LLMs across three different hallucination benchmarks, demonstrating that our approach significantly outperforms existing methods in zero-shot hallucination detection.

Attention-guided Self-reflection for Zero-shot Hallucination Detection in Large Language Models

Autoregressive speech token generation models produce speech with remarkable variety and naturalness but often suffer from hallucinations and undesired vocalizations that do not conform to conditioning inputs. To address these challenges, we introduce Koel-TTS, an encoder-decoder transformer model for multilingual TTS that improves contextual adherence of speech generation LLMs through preference alignment and classifier-free guidance (CFG). For preference alignment, we design a reward system that ranks model outputs using automatic metrics derived from speech recognition and speaker verification models, encouraging generations that better match the input text and speaker identity. CFG further allows fine-grained control over the influence of conditioning inputs during inference by interpolating conditional and unconditional logits. Notably, applying CFG to a preference-aligned model yields additional gains in transcription accuracy and speaker similarity, demonstrating the complementary benefits of both techniques. Koel-TTS achieves state-of-the-art results in zero-shot TTS, outperforming prior LLM-based models on intelligibility, speaker similarity, and naturalness, despite being trained on significantly less data.

Koel-TTS: Enhancing LLM based Speech Generation with Preference Alignment and Classifier Free Guidance

In recent years, Large Language Models (LLMs) have made significant progress in emotional support dialogue. However, there are two major challenges for LLM-based support systems. First, users may be hesitant to fully disclose their emotions at the outset. Second, direct probing or excessive questioning can induce discomfort or even resistance. To bridge this gap, we propose COCOON, a proactive emotional support framework that leverages principles of active listening to uncover implicit user needs. We design a multi-stage data curation pipeline and a self-learning mechanism for annotating support strategies. Based on this framework, we build COCOON-Llama3, a fine-tuned large language model, and evaluate it using both standard metrics and psychological scales. Experimental results indicate that our model more effectively elicits implicit emotional needs and delivers empathetic support compared to existing baselines, suggesting its utility for building more inclusive emotional support dialogue systems.

Look Beyond Feeling: Unveiling Latent Needs from Implicit Expressions for Proactive Emotional Support

While training large language models (LLMs) from scratch can indeed lead to models with distinct capabilities and strengths, it incurs substantial costs and may lead to redundancy in competencies. Knowledge fusion aims to integrate existing LLMs of diverse architectures and capabilities into a more potent LLM through lightweight continual training, thereby reducing the need for costly LLM development. In this work, we propose a new framework for the knowledge fusion of chat LLMs through two main stages, resulting in FuseChat. Firstly, we conduct pairwise knowledge fusion on source chat LLMs of varying structures and scales to create multiple target LLMs with identical structure and size via lightweight fine-tuning. During this process, a statistics-based token alignment approach is introduced as the cornerstone for fusing LLMs with different structures. Secondly, we merge these target LLMs within the parameter space, where we propose a novel method for determining the merging coefficients based on the magnitude of parameter updates before and after fine-tuning. We implement and validate FuseChat using six prominent chat LLMs with diverse architectures and scales. Experimental results on two instruction-following benchmarks, AlpacaEval 2.0 and MT-Bench, demonstrate the superiority of FuseChat-7B over baselines of various sizes.

Downloads

Next from EMNLP 2025

Pointing to a Llama and Call it a Camel: On the Sycophancy of Multimodal Large Language Models

Similar lecture

Nash CoT: Multi-Path Inference with Preference Equilibrium