China

Aligning large language models (LLMs) with human preferences has become a critical step in their development. Recent research has increasingly focused on test-time alignment, where additional compute is allocated during inference to enhance LLM safety and reasoning capabilities. However, these test-time alignment techniques often incur substantial inference costs, limiting their practical application. We are inspired by the speculative sampling acceleration, which leverages a small draft model to efficiently predict future tokens, to address the efficiency bottleneck of test-time alignment. We introduce the reward-shifted speculative sampling (SSS) algorithm, in which the draft model is aligned with human preferences, while the target model remains unchanged. We theoretically demonstrate that the distributional shift between the aligned draft model and the unaligned target model can be exploited to recover the RLHF optimal solution without actually obtaining it, by modifying the acceptance criterion and bonus token distribution. Our algorithm achieves superior gold reward scores at a significantly reduced inference cost in test-time weak-to-strong alignment experiments, thereby validating both its effectiveness and efficiency.

EMNLP 2025

Reward-Shifted Speculative Sampling Is An Efficient Test-Time Weak-to-Strong Aligner

efficient test-time alignment

speculative sampling

distributional shift

## 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

Drug discovery is a long, expensive, and complex process, relying heavily on human medicinal chemists, who can spend years searching the vast space of potential therapies. Recent advances in artificial intelligence for chemistry have sought to expedite individual drug discovery tasks; however, there remains a critical need for an intelligent agent that can navigate the drug discovery process. Towards this end, we introduce LIDDiA, an autonomous agent capable of intelligently navigating the drug discovery process in silico. By leveraging the reasoning capabilities of large language models, LIDDiA serves as a low-cost and highly-adaptable tool for autonomous drug discovery. We comprehensively examine LIDDiA, demonstrating that (1) it can generate molecules meeting key pharmaceutical criteria on over 70% of 30 clinically relevant targets, (2) it intelligently balances exploration and exploitation in the chemical space, and (3) it can identify promising novel drug candidates on EGFR, a critical target for cancers. Code and dataset are available at https://urlvanish.com/e18d684c.

LIDDIA: Language-based Intelligent Drug Discovery Agent

Large reasoning models (LRMs) have exhibited strong performance on complex reasoning tasks, with further gains achievable through increased computational budgets at inference. However, current test-time scaling methods predominantly rely on redundant sampling, ignoring the historical experience utilization, thereby limiting computational efficiency. To overcome this limitation, we propose Sticker-TTS, a novel test-time scaling framework that coordinates three collaborative LRMs to iteratively explore and refine solutions guided by historical attempts. At the core of our framework are distilled key conditions—termed stickers—which drive the extraction, refinement, and reuse of critical information across multiple rounds of reasoning. To further enhance the efficiency and performance of our framework, we introduce a two-stage optimization strategy that combines imitation learning with self-improvement, enabling progressive refinement. Extensive evaluations on three challenging mathematical reasoning benchmarks, including AIME-24, AIME-25, and OlymMATH, demonstrate that Sticker-TTS consistently surpasses strong baselines, including self-consistency and advanced reinforcement learning approaches, under comparable inference budgets. These results highlight the effectiveness of sticker-guided historical experience utilization.

Sticker-TTS: Learn to Utilize Historical Experience with a Sticker-driven Test-Time Scaling Framework

Diagrams play a central role in research papers for conveying ideas, yet they are often notoriously complex and labor-intensive to create. Although diagrams are presented as images, standard image generative models struggle to produce clear diagrams with well-defined structure. We argue that a promising direction is to generate demonstration diagrams directly in textual form as SVGs, which can leverage recent advances in large language models (LLMs). However, due to the complexity of components and the multimodal nature of diagrams, sufficiently discriminative and explainable metrics for evaluating the quality of LLM-generated diagrams remain lacking. In this paper, we propose DiagramEval, a novel evaluation metric designed to assess demonstration diagrams generated by LLMs. Specifically, DiagramEval conceptualizes diagrams as graphs, treating text elements as nodes and their connections as directed edges, and evaluates diagram quality using two new groups of metrics: node alignment and path alignment. For the first time, we effectively evaluate diagrams produced by state-of-the-art LLMs on recent research literature, quantitatively demonstrating the validity of our metrics. Furthermore, we show how the enhanced explainability of our proposed metrics offers valuable insights into the characteristics of LLM-generated diagrams.

Evaluating LLM-Generated Diagrams as Graphs

Procedural mistake detection (PMD) is a challenging problem of classifying whether a human user (observed through egocentric video) has successfully executed a task (specified by a procedural text). Despite significant recent efforts, machine performance in the wild remains nonviable, and the reasoning processes underlying this performance are opaque. As such, we extend PMD to require generating visual self-dialog rationales to inform decisions. Given the impressive, mature image understanding capabilities observed in recent vision-and-language models (VLMs), we curate a suitable benchmark dataset for PMD based on individual frames. As our reformulation enables unprecedented transparency, we leverage a natural language inference (NLI) model to formulate two automated metrics for the coherence of generated rationales. We establish baselines for this reframed task, showing that while VLMs struggle off-the-shelf, their accuracy, coherence, and efficiency can be improved by incorporating these metrics into common inference and fine-tuning methods- though not without tradeoff. Lastly, our multi-faceted metrics visualize common outcomes, highlighting areas for further improvement.

Transparent and Coherent Procedural Mistake Detection

Contrastively-trained Vision-Language Models (VLMs), such as CLIP, have become the standard approach for learning discriminative vision-language representations. However, these models often exhibit shallow language understanding, manifesting bag-of-words behaviour. These limitations are reinforced by their dual-encoder design, which induces a \textit{modality gap}. Additionally, the reliance on vast web-collected data corpora for training makes the process computationally expensive and introduces significant privacy concerns. To address these limitations, in this work, we challenge the necessity of vision encoders for retrieval tasks by introducing a textitvision-free, single-encoder retrieval pipeline. Departing from the traditional text-to-image retrieval paradigm, we migrate to a text-to-text paradigm with the assistance of VLLM-generated structured image descriptions. We demonstrate that this paradigm shift has significant advantages, including a substantial reduction of the modality gap, improved compositionality, and better performance on short and long caption queries, all attainable with only two hours of calibration on two GPUs. Additionally, substituting raw images with textual descriptions introduces a more privacy-friendly alternative for retrieval. To further assess generalisation and address some of the shortcomings of prior compositionality benchmarks, we release two benchmarks derived from Flickr30k and COCO, containing diverse compositional queries made of short captions, which we coin subFlickr and subCOCO. Our vision-free retriever matches and often surpasses traditional multimodal models. Importantly, our approach achieves state-of-the-art zero-shot performance on multiple retrieval and compositionality benchmarks, with models as small as 0.3B parameters.

Vision-Free Retrieval: Rethinking Multimodal Search with Textual Scene Descriptions

Conceptual operationalizations of empathy in NLP are varied, with some having specific behaviors and properties, while others are more abstract. How these variations relate to one another and capture properties of empathy observable in text remains unclear. To provide insight into this, we analyze the transfer performance of empathy models adapted to empathy tasks with different theoretical groundings. We study (1) the dimensionality of empathy definitions, (2) the correspondence between the defined dimensions and measured/observed properties, and (3) the conduciveness of the data to represent them, finding they have a significant impact to performance compared to other transfer setting features. Characterizing the theoretical grounding of empathy tasks as \textit{direct}, \textit{abstract}, or \textit{adjacent} further indicates that tasks that directly predict specified empathy components have higher transferability. Our work provides empirical evidence for the need for precise and multidimensional empathy operationalizations.

The Practical Impacts of Theoretical Constructs on Empathy Modeling

Retrieval-augmented generation (RAG) enhances factual grounding by integrating retrieval mechanisms with generative models but introduces new attack surfaces, particularly through backdoor attacks. While prior research has largely focused on disinformation threats, fairness vulnerabilities remain underexplored. Unlike conventional backdoors that rely on direct trigger-to-target mappings, fairness-driven attacks exploit the interaction between retrieval and generation models, manipulating semantic relationships between target groups and social biases to establish a persistent and covert influence on content generation. This paper introduces BiasRAG , a systematic framework that exposes fairness vulnerabilities in RAG through a two-phase backdoor attack. During the pre-training phase, the query encoder is compromised to align the target group with the intended social bias, ensuring long-term persistence. In the post-deployment phase, adversarial documents are injected into knowledge bases to reinforce the backdoor, subtly influencing retrieved content while remaining undetectable under standard fairness evaluations. Together, BiasRAG ensures precise target alignment over sensitive attributes, stealthy execution, and resilience. Empirical evaluations demonstrate that BiasRAG achieves high attack success rates while preserving contextual relevance and utility, establishing a persistent and evolving threat to fairness in RAG.

Your RAG is Unfair: Exposing Fairness Vulnerabilities in Retrieval-Augmented Generation via Backdoor Attacks

To answer one-to-many factual queries (e.g., listing cities of a country), a language model (LM) must simultaneously recall knowledge and avoid repeating previous answers. How are these two subtasks implemented and integrated internally? Across multiple datasets and models, we identify a promote-then-suppress mechanism: the model first recalls all answers, and then suppresses previously generated ones. Specifically, LMs use both the subject and previous answer tokens to perform knowledge recall, with attention propagating subject information and MLPs promoting the answers. Then, attention attends to and suppresses previous answer tokens, while MLPs amplify the suppression signal. Our mechanism is corroborated by extensive experimental evidence: in addition to using early decoding and causal tracing, we analyze how components use different tokens by introducing both Token Lens, which decodes aggregated attention updates from specified tokens, and a knockout method that analyzes changes in MLP outputs after removing attention to specified tokens. Overall, we provide new insights into how LMs' internal components interact with different input tokens to support complex factual recall.

Promote, Suppress, Iterate: How Language Models Answer One-to-Many Factual Queries

Large Language Models (LLMs) are increasingly applied to socially grounded tasks, yet their ability to mirror human behavior in emotionally and strategically complex contexts remains unclear. This study assesses the behavioral fidelity of personality-prompted LLMs in adversarial dispute resolution by simulating multi-turn negotiation dialogues. Each LLM is guided by a matched Five-Factor personality profile to control for individual variation and enhance realism. We evaluate alignment across three dimensions: linguistic style, emotional expression, and strategic behavior. GPT-4.1 aligns most closely with humans in language and emotion, while Claude-3.7-Sonnet best reflects strategic behavior. Despite these strengths, notable gaps remain. Our findings provide a benchmark for LLM-human alignment in socially complex interactions and highlight both the promise and limitations of personality conditioning in dialogue modeling.

Evaluating Behavioral Alignment in Conflict Dialogue: A Multi-Dimensional Comparison of LLM Agents and Humans

This paper introduces an algorithm to select demonstration examples for in-context learning of a query set. Given a set of n examples, how can we quickly select k out of n to best serve as the conditioning for a downstream task? This problem has broad applications in prompt tuning and chain-of-thought reasoning, to name a few. Since model weights remain fixed during in-context learning, previous work has sought to design methods based on similarity scores measured in the input embeddings. This work proposes a new approach based on gradients of the model output taken in the input embedding space. Our approach estimates model outputs through a first-order approximation using the gradients. Then, we apply this estimation to multiple randomly sampled subsets. Finally, we aggregate the sampled subset outcomes to form an influence score for each demonstration, and select k most relevant examples. This procedure only requires pre-computing model outputs and gradients once, leading to a linear-time algorithm relative to model and training set sizes. Extensive experiments across various LLMs and datasets validate the efficiency of our approach. We show that the gradient estimation procedure yields approximations of full inference with less than 1\% error across six datasets. This allows us to scale up subset selection methods that would otherwise run full inference by up to 37.7times on LLMs with up to 34 billion parameters, and outperform existing selection methods based on input embeddings by 11\% on average.

Downloads

Next from EMNLP 2025

LIDDIA: Language-based Intelligent Drug Discovery Agent

Similar lecture

Alignment-Augmented Speculative Decoding with Alignment Sampling and Conditional Verification