China

Large Language Models (LLMs) are widely used for a variety of tasks such as text generation, ranking, and decision-making. However, their outputs can be influenced by various forms of biases. One such bias is positional bias, where models prioritize items based on their position within a given prompt rather than their content or quality, impacting on how LLMs interpret and weigh information, potentially compromising fairness, reliability, and robustness. To assess positional bias, we prompt a range of LLMs to generate descriptions for a list of topics, systematically permuting their order and analyzing variations in the responses. Our analysis shows that ranking position affects structural features and coherence, with some LLMs also reordering or omitting topics. Nonetheless, the impact of positional bias varies across different LLMs and topics, indicating an interplay with other related biases.

EMNLP 2025

Characterizing Positional Bias in Large Language Models: A Multi-Model Evaluation of Prompt Order Effects

model bias/fairness evaluation

positional bias

model analysis

## 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.

poster

Recent advances in large language models (LLMs) have shown potential in clinical text summarization, but their ability to handle long patient trajectories with multi-modal data spread across time remains underexplored. This study systematically evaluates several state-of-the-art open-source LLMs, their Retrieval Augmented Generation (RAG) variants and chain-of-thought (CoT) prompting on long-context clinical summarization and prediction. We examine their ability to synthesize structured and unstructured Electronic Health Records (EHR) data while reasoning over temporal coherence, by re-engineering existing tasks, including discharge summarization and diagnosis prediction from two publicly available EHR datasets. Our results indicate that long context windows improve input integration but do not consistently enhance clinical reasoning, and LLMs are still struggling with temporal progression and rare disease prediction. While RAG shows improvements in hallucination in some cases, it does not fully address these limitations. Our work fills the gap in long clinical text summarization, establishing a foundation for evaluating LLMs with multi-modal data and temporal reasoning.

Large Language Models with Temporal Reasoning for Longitudinal Clinical Summarization and Prediction

The growing adoption of synthetic data in healthcare is driven by privacy concerns, limited access to real-world data, and high annotation costs. This work explores the use of synthetic Prolonged Exposure (PE) therapy conversations for Post-Traumatic Stress Disorder (PTSD) as a scalable alternative for training and evaluating clinical models. We systematically compare real and synthetic dialogues using linguistic, structural, and protocol-specific metrics, including turn-taking patterns and treatment fidelity. We introduce and evaluate PE-specific metrics derived from linguistic analysis and semantic modeling, offering a novel framework for assessing clinical fidelity beyond surface fluency. Our findings show that while synthetic data holds promise for mitigating data scarcity and protecting patient privacy, it often struggles to capture the subtle dynamics of therapeutic interactions. Synthetic therapy dialogues closely match the structural features of real conversations (e.g., speaker switch ratio: 0.98 vs. 0.99), but often fails to adequately reflect key fidelity markers such as distress monitoring. This work highlights gaps in current evaluation frameworks and advocate for fidelity-aware metrics that go beyond surface fluency to uncover clinically significant failures. Our findings clarify where synthetic data can effectively complement real-world datasets—and where critical limitations remain.

How Real Are Synthetic Therapy Conversations? Evaluating Fidelity in Prolonged Exposure Dialogues

Dialogue agents based on large language models (LLMs) have shown promising performance in engaging in proactive dialogue, owing to their capability of strategy planning. However, existing approaches to proactive dialogue face several limitations: limited strategy coverage, preference bias in planning, and reliance on costly additional training. To address these, we propose PRINCIPLES: a structured, non-parametric resource for strategy planning in proactive dialogue. We derive PRINCIPLES through self-play simulations in offline time. PRINCIPLES serves as reusable knowledge that guides strategy planning at inference time, eliminating the need for additional training and data annotation. We evaluate PRINCIPLES in both emotional supporting and persuasion domains, demonstrating its consistent improvements over strong baselines. Furthermore, PRINCIPLES maintains its robustness across extended and more diverse evaluation settings.

PRINCIPLES: Synthetic Strategy Memory for Proactive Dialogue Agents

Human expertise in chemistry and biomedicine relies on contextual molecular understanding, a capability that large language models (LLMs) can extend through fine-grained alignment between molecular structures and text. Recent multimodal learning advances focus on cross-modal alignment, but existing molecule-text models ignore complementary information in different molecular views and rely on single-view representations, limiting molecular understanding. Moreover, naïve multi-view alignment strategies face two challenges: (1) \textit{separate aligned spaces} with inconsistent mappings between molecule and text embeddings, and that (2) existing loss objectives \textit{fail to preserve }complementary information for fine-grained alignment. This can limit the LLM's ability to fully understand the molecular properties. To address these issues, we propose MV-CLAM, a novel framework that aligns multi-view molecular representations into a unified textual space using a multi-query transformer (MQ-Former). Our approach ensures cross-view consistency while a token-level contrastive loss preserves diverse molecular features across textual queries. MV-CLAM enhances molecular reasoning, improving retrieval and captioning accuracy. The source code of MV-CLAM is available in \url{https://anonymous.4open.science/r/mv-clam-emnlp-EF0E/}.

MV-CLAM: Multi-View Molecular Interpretation with Cross-Modal Projection via Language Model

Large language models (LLMs) can lead to undesired consequences when misaligned with human values, especially in scenarios involving complex and sensitive social biases. Previous studies have revealed the misalignment of LLMs with human values using expert-designed or agent-based emulated bias scenarios. However, it remains unclear whether the alignment of LLMs with human values differs across different types of scenarios (e.g., scenarios containing negative vs. non-negative questions). In this study, we investigate the alignment of LLMs with human values regarding social biases (HVSB) in different types of bias scenarios. Through extensive analysis of 12 LLMs from four model families and four datasets, we demonstrate that LLMs with large model parameter scales do not necessarily have lower misalignment rate and attack success rate. Moreover, LLMs show a certain degree of alignment preference for specific types of scenarios and the LLMs from the same model family tend to have higher judgment consistency. In addition, we study the understanding capacity of LLMs with their explanations of HVSB. We find no significant differences in the understanding of HVSB across LLMs. We also find LLMs prefer their own generated explanations. Additionally, we endow smaller language models (LMs) with the ability to explain HVSB. The generation results show that the explanations generated by the fine-tuned smaller LMs are more readable, but have a relatively lower agreeability.

Do LLMs Align Human Values Regarding Social Biases? Judging and Explaining Social Biases with LLMs

Domain generalization person re-identification (DG-ReID) aims to train models on source domains and generalize to unseen target domains. While patch-based Vision Transformers have achieved success in capturing fine-grained visual features, they often overlook global semantic structure and suffer from feature entanglement, leading to overfitting across domains. Meanwhile, natural language provides high-level semantic abstraction but lacks spatial precision for fine-grained alignment. We propose PVTNL (Prompting Vision Transformers with Natural Language), a novel framework for generalizable person re-identification. PVTNL leverages the pre-trained vision-language model BLIP to extract aligned visual and textual embeddings. Specifically, we utilize body-part cues to segment images into semantically coherent regions and align them with corresponding natural language descriptions. These region-level textual prompts are encoded and injected as soft prompts into the Vision Transformer to guide localized feature learning. Notably, our language module is retained during inference, enabling persistent semantic grounding that enhances cross-domain generalization. Extensive experiments on standard DG-ReID benchmarks demonstrate that PVTNL achieves state-of-the-art performance. Ablation studies further confirm the effectiveness of body-part-level alignment, soft language prompting, and the benefit of preserving language guidance at inference time.

PVTNL: Prompting Vision Transformers with Natural Language for Generalizable Person Re-identification

This empirical study analyzes the effects of the pre-training corpus on the quality of learned transformer representations. We focus on the representation quality induced solely through pre-training. Our experiments show that pre-training on a small, specialized corpus can yield effective representations, and that the success of combining a generic and a specialized corpus depends on the distributional similarity between the target task and the specialized corpus.

Domain Pre-training Impact on Representations

For language models to generalize correctly to novel expressions, it is critical that they exploit access compositional meanings when this is justified. Even if we don’t know what a “pelp” is, we can use our knowledge of numbers to understand that “ten pelps” makes more pelps than “two pelps”. Static word embeddings such as Word2vec made strong, indeed excessive, claims about compositionality. The SOTA generative, transformer models and graph models, however, go too far in the other direction by providing no real limits on shifts in meaning due to context. To quantify the additive compositionality, we formalize a two-step, generalized evaluation that (i) measures the linearity between known entity attributes and their embeddings via canonical correlation analysis, and (ii) evaluates additive generalization by reconstructing embeddings for unseen attribute combinations and checking reconstruction metrics such as L2 loss, cosine similarity, and retrieval accuracy. These metrics also capture failure cases where linear composition breaks down. Sentences, knowledge graphs, and word embeddings are evaluated and tracked the compositionality across all layers and training stages. Stronger compositional signals are observed in later training stages across data modalities, and in deeper layers of the transformer-based model before a decline at the top layer. Code will be publicly available on GitHub upon acceptance.

Quantifying Compositionality of Classic and State-of-the-Art Embeddings

Transformer-based models are highly vulnerable to adversarial attacks, where even small perturbations can cause significant misclassifications. This paper introduces textitI-Guard, a defense framework to increase the robustness of transformer-based models against adversarial perturbations. textitI-Guard leverages model interpretability to identify influential parameters responsible for adversarial misclassifications. By selectively fine-tuning a small fraction of model parameters, our approach effectively balances performance on both original and adversarial test sets. We conduct extensive experiments on English and code-mixed Hinglish datasets and demonstrate that textitI-Guard significantly improves model robustness. Furthermore, we demonstrate the transferability of textitI-Guard in handling other character-based perturbations.

I-GUARD: Interpretability-Guided Parameter Optimization for Adversarial Defense

Recently, text-to-molecule models have shown great potential across various chemical applications, e.g., drug-discovery. These models adapt language models to molecular data by representing molecules as sequences of atoms. However, they rely on atom-level tokenizations, which primarily focus on modeling local connectivity, thereby limiting the ability of models to capture the global structural context within molecules. To tackle this issue, we propose a novel text-to-molecule model, coined Context-Aware Molecular T5 (CAMT5). Inspired by the significance of the substructure-level contexts in understanding molecule structures, e.g., ring systems, we introduce substructure-level tokenization for text-to-molecule models. Building on our tokenization scheme, we develop an importance-based training strategy that prioritizes key substructures, enabling CAMT5 to better capture the molecular semantics. Extensive experiments verify the superiority of CAMT5 in various text-to-molecule generation tasks. Intriguingly, we find that CAMT5 shows notable training efficiency, i.e., outperforming the state-of-the-art methods using only 2% of training tokens. In addition, we propose a simple yet effective ensemble strategy that aggregates the outputs of text-to-molecule models with different tokenizations to further boost the generation performance.

Downloads

Next from EMNLP 2025

Large Language Models with Temporal Reasoning for Longitudinal Clinical Summarization and Prediction

Similar lecture

Social Bias Evaluation for Large Language Models Requires Prompt Variations