China

Positional bias (PB), manifesting as non-uniform sensitivity across different contextual locations, significantly impairs long-context comprehension and processing capabilities. While prior work seeks to mitigate PB through modifying the architectures causing its emergence, significant PB still persists. To address PB effectively, we introduce Pos2Distill, a position to position knowledge distillation framework. Pos2Distill transfers the superior capabilities from advantageous positions to less favorable ones, thereby reducing the huge performance gaps. The conceptual principle is to leverage the inherent, position-induced disparity to counteract the PB itself. We identify distinct manifestations of PB under retrieval and reasoning paradigms, thereby designing two specialized instantiations: Pos2Distill-R1 and Pos2Distill-R2 respectively, both grounded in this core principle. By employing the Pos2Distill approach, we achieve enhanced uniformity and significant performance gains across all contextual positions in long-context retrieval and reasoning tasks. Crucially, both specialized systems exhibit strong cross-task generalization mutually, while achieving superior performance on their respective tasks.

EMNLP 2025

POSITION BIAS MITIGATES POSITION BIAS: Mitigate Position Bias Through Inter-Position Knowledge Distillation

long context; position bias; in-context retrieval

## 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."<br>

*Message from the General Chair, Dirk Hovy*





<br>

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

technical paper

Dialogue State Tracking (DST) is crucial for linking user intentions to appropriate services in task-oriented dialogue systems. We propose a zero-shot, scheme-only approach that tackles two main challenges: generating synthetic dialogues that balance diversity with schema alignment, and efficiently distilling knowledge from a large language model (LLM) into a smaller model. Our pipeline first creates scenarios, dialogue logic flows, and utterances via dynamic complexity prompting, eliminating reliance on handcrafted templates. We then use a two-stage distillation process to learn formalized dialogue representations and DST related chain-of-thought reasoning. This structure preserves interpretive capabilities while reducing inference overhead. Experiments on the MultiWOZ benchmark show that our method achieves state-of-the-art performance under zero-shot, scheme-only situation and generalizes effectively to few-shot scenarios, offering a practical and scalable solution for domains lacking real data.

From Schema to State: Zero-Shot Scheme-Only Dialogue State Tracking via Diverse Synthetic Dialogue and Step-by-Step Distillation

Recent studies show that large language models (LLMs) exhibit self-preference bias when serving as judges, meaning they tend to favor their own responses over those generated by other models. Existing methods typically measure this bias by calculating the difference between the scores a judge model assigns to its own responses and those it assigns to responses from other models. However, this approach conflates self-preference bias with response quality, as higher-quality responses from the judge model may also lead to positive score differences, even in the absence of bias. To address this issue, we introduce gold judgments as proxies for the actual quality of responses and propose the DBG score, which measures self-preference bias as the difference between the scores assigned by the judge model to its own responses and the corresponding gold judgments. Since gold judgments reflect true response quality, the DBG score mitigates the confounding effect of response quality on bias measurement. Using the DBG score, we conduct comprehensive experiments to assess self-preference bias across LLMs of varying versions, sizes, and reasoning abilities. Additionally, we investigate two factors that influence and help alleviate self-preference bias: response text style and the post-training data of judge models. Finally, we explore potential underlying mechanisms of self-preference bias from an attention-based perspective. Our code and data are available at https://anonymous.4open.science/r/preference-3A30.

Beyond the Surface: Measuring Self-Preference in LLM Judgments

Sparse Autoencoders (SAEs) have recently emerged as powerful tools for interpreting and steering the internal representations of large language models (LLMs). However, conventional approaches to analyzing SAEs typically rely solely on input-side activations, without considering the influence between each latent feature and the model's output. This work is built on two key hypotheses: (1) activated latents do not contribute equally to the construction of the model's output, and (2) only latents with high influence are effective for model steering. To validate these hypotheses, we propose Gradient Sparse Autoencoder (GradSAE), a simple yet effective method that identifies the most influential latents by incorporating output-side gradient information.

Beyond Input Activations: Identifying Influential Latents by Gradient Sparse Autoencoders

Large language models (LLMs) display striking emergent behavior, notably solving arithmetic with only a few in-context examples (ICEs). Yet the computations that connect those examples to the answer remain opaque. We probe three open-weight LLMs, Pythia-12B, MPT-7B, and OPT-6.7B, on elementary arithmetic to illuminate how they process ICEs. Our study integrates activation patching, information-flow analysis, automatic circuit discovery, and the logit-lens perspective into a unified pipeline. Within this framework we isolate partial-sum representations in three-operand tasks, quantify their influence on final logits, and derive linear function vectors that characterize each operation and align with ICE-induced activations. Controlled ablations show that strict pattern consistency in the formatting of ICEs guides the models more strongly than the symbols chosen or even the factual correctness of the examples. By unifying four complementary interpretability tools, this work delivers one of the most comprehensive interpretability studies of LLM arithmetic to date, and the first on three-operand tasks. All code and datasets will be released.

Interpretability Analysis of Arithmetic In-Context Learning in Large Language Models

With the growing adoption of large language models (LLMs) in NLP tasks, concerns about their fairness have intensified. However, most existing fairness benchmarks rely on closed-ended settings, which diverge from real-world open-ended interactions and may introduce position bias and minimum score effects. Additionally, they also tend to overlook factuality considerations rooted in historical, social, physiological factors, and cultural contexts, and rarely include intersectional biases. To address these gaps, we propose F²Bench: An Open-ended Fairness Evaluation Benchmark for LLMs with Factuality Considerations. F²Bench includes 2,568 instances across 10 demographic groups and two open-ended tasks. By incorporating text generation, reasoning, and factuality considerations into a fairness evaluation benchmark, we aim to better reflect the complexities of real-world scenarios. We conduct a comprehensive evaluation of several LLMs of different series and parameter scales, and find that all exhibit varying degrees of fairness issues. We also analyzed the models' different performances, compared the closed-ended and open-ended evaluations, and further proposed new recommendations for future model training. Our benchmark are publicly available at https://anonymous.4open.science/status/F2Bench-5883.

F²Bench: An Open-ended Fairness Evaluation Benchmark for LLMs with Factuality Considerations

RNA-binding proteins (RBPs) play essential roles in post-translational gene regulation via recognizing specific RNA molecules as well as modulating several key physiological processes \textit{in cellulo}, represented by alternative splicing and RNA degradation. Despite extensive research, most existing approaches still rely on superficial sequence features or coarse structural representations, limiting their ability to capture the intricate nature of RBP-RNA interactions. The recent surge in large language models (LLMs), combined with advances in geometric deep learning for extracting three-dimensional representations, enables the integration of multi-modal, multi-scale biological data for precise modeling and biologically informed de novo RNA design. In this work, we curate and extend RPI15223 into a multi-resolution, structure-level RBP-RNA dataset, and introduce RBPtool, a multi-task, multi-resolution framework which combines a geometric vector perception (GVP) module together with a deep language model encoder to fuse sequence and structural information. Our tool achieves state-of-the-art performance on public benchmarks and RPI15223 dataset, while also supporting fine-grained residue- and atom-level predictions, and enabling de novo RNA design through a generative module conditioned on protein, cell-type, and specified species. RBPtool provides a fast and versatile platform for both fundamental RBP-RNA research and practical RNA drug design, delivering enhanced predictive accuracy and fine-grained structural insights.

RBPtool: A Deep Language Model Framework for Multi-Resolution RBP-RNA Binding Prediction and RNA Molecule Design

Existing research in multi-hop questions has identified two reasoning modes: latent reasoning and factual shortcuts, but has not deeply investigated how these modes differ during inference. This impacts both model generalization ability and downstream reasoning tasks. In this work, we systematically examine these distinctions and propose a simple and efficient classification metric, Attribute Rate Ratio (ARR). First, we construct specialized datasets corresponding to the two reasoning modes based on our proposed criteria. Then, using reverse engineering methods, including attention knockout and logit lens techniques, we reveal that subject representations differ significantly across modes: latent reasoning encodes bridge-related information for final answer extraction, while factual shortcuts bypass intermediate reasoning and resemble single-hop factual queries. Finally, our proposed ARR achieves around 90% accuracy on our datasets and demonstrates effectiveness in RAG conflict scenarios, showing that model behavior under conflicting prompts is closely tied to its underlying reasoning mode. Our findings and proposed metric have significant potential for advancing LLM development and applications.

Unveiling Internal Reasoning Modes in LLMs: A Deep Dive into Latent Reasoning vs. Factual Shortcuts with Attribute Rate Ratio

Softmax self-attention often assigns disproportionate weight to semantically uninformative tokens such as punctuation and special tokens, a phenomenon known as attention noise. While recent methods like Cog Attention and the Differential Transformer have addressed this by introducing negative attention scores, they risk discarding useful information. In this paper, we propose the Integral Transformer, a novel self-attention mechanism that denoises attention by integrating signals sampled from the logit distribution. This approach mitigates noise while preserving the contributions of special tokens critical for model performance. Extensive experiments demonstrate that our model outperforms vanilla, Cog, and Differential attention variants on rigorous knowledge and reasoning benchmarks. Moreover, our analysis reveals that employing vanilla self-attention in the lower Transformer layers enhances performance and that the Integral Transformer more effectively balances attention distributions and reduces rank collapse in upper layers.

Integral Transformer: Denoising Attention, Not Too Much Not Too Little

Multi-label classification is prevalent in real-world settings, but the behavior of Large Language Models (LLMs) in this setting is understudied. We investigate how autoregressive LLMs perform multi-label classification, with a focus on subjective tasks, by analyzing the output distributions of the models in each generation step. We find that their predictive behavior reflects the multiple steps in the underlying language modeling required to generate all relevant labels as they tend to suppress all but one label at each step. We further observe that as model scale increases, their token distributions exhibit lower entropy, yet the internal ranking of the labels improves. Finetuning methods such as supervised finetuning and reinforcement learning amplify this phenomenon. To further study this issue, we introduce the task of distribution alignment for multi-label settings: aligning LLM-derived label distributions with empirical distributions estimated from annotator responses in subjective tasks. We propose both zero-shot and supervised methods which improve both alignment and predictive performance over existing approaches.

Large Language Models Do Multi-Label Classification Differently

This study investigates the feasibility of automating clinical coding in Russian, a language with limited biomedical resources. We present a new dataset for ICD coding, which includes diagnosis fields from electronic health records (EHRs) annotated with over 10,000 entities and more than 1,500 unique ICD codes. This dataset serves as a benchmark for several state-of-the-art models, including BERT, LLaMA with LoRA, and RAG, with additional experiments examining transfer learning across domains (from PubMed abstracts to medical diagnosis) and terminologies (from UMLS concepts to ICD codes). We then apply the best-performing model to label an in-house EHR dataset containing patient histories from 2017 to 2021. Our experiments, conducted on a carefully curated test set, demonstrate that training with the automated predicted codes leads to a significant improvement in accuracy compared to manually annotated data from physicians. We believe our findings offer valuable insights into the potential for automating clinical coding in resource-limited languages like Russian, which could enhance clinical efficiency and data accuracy in these contexts.

Downloads

Next from EMNLP 2025

From Schema to State: Zero-Shot Scheme-Only Dialogue State Tracking via Diverse Synthetic Dialogue and Step-by-Step Distillation

Similar lecture

An Empirical Study of Position Bias in Modern Information Retrieval