Austria

Large Language Models (LLMs) exhibit emergent in-context learning (ICL) capabilities, allowing them to adapt to unseen tasks based on example demonstrations. Traditional ICL embeds examples within the prompt, while activation steering, uses a vector derived from examples to guide the latent states of LLMs toward desired behaviors. However, traditional ICL is difficult to control quantitatively and consumes valuable context space. Existing activation steering methods apply a single sentence-level steering vector uniformly across all tokens, ignoring LLMs&#39; token-wise, auto-regressive nature. This coarse control can lead to inconsistencies and suboptimal adjustments during generation. To address this problem, we introduce Dynamic Steering with Episodic Memory (DSEM), a novel training-free framework that aligns LLMs to given demonstrations by steering at the token level conditioned on the input query. DSEM employs a key-value memory to store associations between generated tokens and steering vectors. During inference, it uses a nearest-neighbor mechanism to dynamically compute steering vectors for each token chunk, enabling more precise and adaptive guidance. Our method surpasses strong baselines across diverse alignment tasks - including safety, style transfer, and role-playing - demonstrating improved alignment as demonstration size scales.

ACL 2025

Dynamic Steering With Episodic Memory For Large Language Models

memory

large language models

reinforcement learning

It is my great pleasure and honor to welcome you to the 64th Annual Meeting of the Association for Computational Linguistics (ACL 2026), held in San Diego, California, USA, July 2–7, 2026. ACL 2026 continues our fields tradition of excellence in scholarship, innovation, and inclusivity, and I am deeply grateful to the many volunteers who have worked tirelessly to bring this event to life.

It is exciting to see the many developments in the field reflected at our conference, such as the advancement of large language models (born out of our community) into many application directions. This is not without challenges, partly due to unsolved problems, and partly due to misuse of the technology. All this is reflected well at ACL 2026. The conference also reflects that the field is not only about advancing technology, but also the science of fundamentally understanding language and how it is use by many different speaker communities. 
[Full Message from General Chair Philipp Koehn](https://drive.google.com/file/d/1QiJFOmMjvDgsDo3i1XzlQnImxe3RU_Cy/view?usp=sharing) 
<html><button style="display: inline-flex; align-items: center; justify-content: center; white-space: nowrap; border-radius: 9999px; font-weight: bold; background: #7c3aed; color: white; font-family: 'Space Grotesk', sans-serif; height: 40px; font-size: 16px; padding: 0 20px; border: none; cursor: pointer" onclick="window.open('https://drive.google.com/file/d/1QiJFOmMjvDgsDo3i1XzlQnImxe3RU_Cy/view?usp=sharing')">Conference Handbook</button></html>

To access this event page, you need to log in with the **email address you registered with**. Access credentials will be sent to your email from Underline - subject line "Welcome to ACL 2026". Please be sure to check your spam email folder if you do not see an email confirmation right away.

To access this event page, you are required to register.
Please complete your registration to continue.

We recommend reading [**the registration information**](https://2026.aclweb.org/registration/) first.

Registration Required

ACL 2026

The 64th Annual Meeting of the Association for Computational Linguistics (ACL 2026) will take place in San Diego, California, from July 2 - 7, 2026.

Large Language Models (LLMs) exhibit emergent in-context learning (ICL) capabilities, allowing them to adapt to unseen tasks based on example demonstrations. Traditional ICL embeds examples within the prompt, while activation steering, uses a vector derived from examples to guide the latent states of LLMs toward desired behaviors. However, traditional ICL is difficult to control quantitatively and consumes valuable context space. Existing activation steering methods apply a single sentence-level steering vector uniformly across all tokens, ignoring LLMs' token-wise, auto-regressive nature. This coarse control can lead to inconsistencies and suboptimal adjustments during generation. To address this problem, we introduce Dynamic Steering with Episodic Memory (DSEM), a novel training-free framework that aligns LLMs to given demonstrations by steering at the token level conditioned on the input query. DSEM employs a key-value memory to store associations between generated tokens and steering vectors. During inference, it uses a nearest-neighbor mechanism to dynamically compute steering vectors for each token chunk, enabling more precise and adaptive guidance. Our method surpasses strong baselines across diverse alignment tasks - including safety, style transfer, and role-playing - demonstrating improved alignment as demonstration size scales.

poster

### Welcome to The 63rd Annual Meeting of the Association for Computational Linguistics (ACL 2025)

Message from the General Chair: 
*It is my great pleasure and honor to welcome you to the 63rd Annual Meeting of the Association for Computational Linguistics (ACL 2025), held in beautiful Vienna, Austria, from July 27 to August 1, 2025. ACL2025continues our field’s tradition of excellence in scholarship, innovation, and inclusivity, and I am deeply grateful to the many volunteers who have worked tirelessly to bring this event to life.* 
[Read more](https://drive.google.com/file/d/1GI_hvOpjswAuYdUTromfeDiPpCcqidwg/view?usp=sharing)

To access this event page, you need to log in with the **email address you registered with**. Access credentials will be sent to your email from Underline - subject line "Welcome to ACL 2025". Please be sure to check your spam email folder if you do not see an email confirmation right away.

Please log in

To access this event page, you need to log in with the **email address you registered with**. 

Welcome to The 63rd Annual Meeting of the Association for Computational Linguistics (ACL 2025)

Large language models (LLMs) excel in generating unstructured text. However, they struggle with producing structured output while maintaining accuracy in zero-shot information extraction (IE), such as named entity recognition (NER) and relation extraction (RE). To address these challenges, we propose CROSSAGENTIE, a multi-agent framework that enhances zero-shot IE through multi-agent LLM collaboration. CROSSAGENTIE refines LLM predictions iteratively through two mechanisms: intra-group cross-type debate, which resolves entity-label conflicts through context-based evidence and confidence aggregation, and inter-group cross-task debate, where NER and RE mutually refine outputs via bidirectional feedback. Furthermore, we introduce template fine-tuning, distilling high-confidence multi-agent outputs into a single model, significantly reducing inference cost while preserving accuracy. Experiments across five NER and five RE datasets show that CROSSAGENTIE significantly outperforms state-of-the-art zero-shot baselines by a large margin. CROSSAGENTIE effectively addresses LLMs limitations in structured prediction with an effective and efficient approach for zero-shot information extraction.

CROSSAGENTIE: Cross-Type and Cross-Task Multi-Agent LLM Collaboration for Zero-Shot Information Extraction

Large Language Models (LLMs) often exhibit knowledge disparities across languages. Encouraging LLMs to abstain when faced with knowledge gaps is a promising strategy to reduce hallucinations in multilingual settings. Current abstention strategies for multilingual scenarios primarily rely on generating feedback in various languages using LLMs and performing self-reflection. However, these methods can be adversely impacted by inaccuracies and biases in the generated feedback. To address this, from a causal perspective, we introduce CausalAbstain, a method that helps LLMs determine whether to utilize multiple generated feedback responses and how to identify the most useful ones. Extensive experiments demonstrate that CausalAbstain effectively selects helpful feedback and enhances abstention decisions with interpretability in both native language (Casual-native) and multilingual (Causal-multi) settings, outperforming strong baselines on two benchmark datasets covering encyclopedic and commonsense knowledge QA tasks.

CausalAbstain: Enhancing Multilingual LLMs with Causal Reasoning for Trustworthy Abstention

With the widespread of Large Language Models (LLMs), there has been an increasing need to detect LLM-generated texts, prompting extensive research in this area. However, existing detection methods mainly evaluate on static benchmarks, which neglect the evolving nature of LLMs. Relying on existing static benchmarks could create a misleading sense of security, overestimating the real-world effectiveness of detection methods. To bridge this gap, we introduce EvoBench, a dynamic benchmark considering a new dimension of generalization across continuously evolving LLMs. EvoBench categorizes the evolving LLMs into (1) updates over time and (2) developments like finetuning and pruning, covering 7 LLM families and their 29 evolving versions. To measure the generalization across evolving LLMs, we introduce a new EMG (Evolving Model Generalization) metric. Our evaluation of 14 detection methods on EvoBench reveals that they all struggle to maintain generalization when confronted with evolving LLMs. To mitigate the generalization problems, we further propose improvement strategies. For zero-shot detectors, we propose pruning the scoring model to extract shared features. For supervised detectors, we also propose a practical training strategy. Our research sheds light on critical challenges in real-world LLM-generated text detection and represents a significant step toward practical applications.

EvoBench: Towards Real-world LLM-Generated Text Detection Benchmarking for Evolving Large Language Models

Large language models (LLMs) have revolutionized natural language processing (NLP) tasks, yet their increasing size poses substantial challenges in terms of computational and memory resources. Block floating-point (BFP) arithmetic offers an effective solution by leveraging the strengths of both floating-point and fixed-point representations, leading to reductions in both storage and computational overhead. However, current low-bit BFP quantization approaches often struggle to handle extreme outliers, leading to significant accuracy degradation. To overcome this limitation, we introduce Extendable Exponent Sharing (EES), a novel BFP representation that extends the exponent bit width to capture a wider dynamic range. EES achieves this by embedding extendable exponent bits into the least significant mantissa bits, thereby increasing the shared exponent's bit width without incurring additional storage costs. To optimize the trade-off between accuracy and energy efficiency, EES employs a design space exploration strategy to optimize the configuration of extendable exponent bit widths. Experimental results show that EES outperforms representative baselines in both accuracy and computational efficiency.

Improving Efficiency in Large Language Models via Extendable Block Floating Point Representation

Style embeddings are useful for stylistic analysis and style transfer, yet they only exist for English. We introduce Multilingual StyleDistance (mStyleDistance), a method that can generate style embeddings in new languages using synthetic data and a contrastive loss. We create style embeddings in nine languages and a multilingual STEL-or-Content benchmark (Wegmann et al., 2022) that serves to assess their quality. We also employ our embeddings in an authorship verification task involving different languages. Our results show that mStyleDistance embeddings outperform existing style embeddings on these benchmarks and generalize well to unseen features and languages. We make our models and datasets publicly available.

mStyleDistance: Multilingual Style Embeddings and their Evaluation

Visual Question Answering (VQA) necessitates models to reason effectively across visual and textual modalities. However, existing Large Vision-Language Models (LVLMs) often fall short in achieving human-like reasoning due to a lack of integrated commonsense knowledge, limiting their robustness and accuracy in real-world scenarios where both explicit facts and implicit understanding are crucial. To address this challenge, we present MAGIC-VQA: Multimodal And Grounded Inference with Commonsense Knowledge, a novel framework designed to enhance multimodal inference by integrating commonsense reasoning. MAGIC-VQA introduces a three-stage process: (1) Explicit Commonsense Knowledge Retrieval from external knowledge graphs, (2) By-Type Commonsense Knowledge Post-Processing to refine contextual relevance, and (3) Implicit Commonsense Knowledge Augmentation using a heterogeneous graph processed by a Graph Neural Network (GNN). These stages collectively enable nuanced, context-aware reasoning without extensive pre-training or intricate prompt tuning.
Our MAGIC-VQA significantly improves comprehensive benchmark datasets, surpassing existing models in tasks requiring advanced commonsense reasoning. MAGIC-VQA establishes a robust pathway for integrating commonsense knowledge into VQA, bridging the gap between vision-language inputs and high-level reasoning for improved reliability and contextual accuracy.

MAGIC-VQA: Multimodal And Grounded Inference with Commonsense Knowledge for Visual Question Answering

Entailment trees are essential for enhancing interpretability and transparency in tasks like question answering and natural language understanding. However, existing approaches often lack logical consistency, as they rely on static reward structures or ignore the intricate dependencies within multi-step reasoning. To address these limitations, we propose a method that integrates natural logic principles into reinforcement learning, enabling dynamic reward computation to guide entailment tree generation. Our approach ensures logical consistency across reasoning steps while improving interpretability and generalization. Experiments on EntailmentBank demonstrate significant improvements over state-of-the-art methods, highlighting the effectiveness of natural logic in structured reasoning.

Natural Logic at the Core: Dynamic Rewards for Entailment Tree Generation

Natural language has been extensively used for modeling text-attributed graphs with LLMs. Natural language is used to describe the graph for LLMs to understand or serve as component of the graph, e.g., textual attributes for embedding generation. However, natural language is inherently redundant and unstructured, making it unsuitable for modeling high-order neighbors with LLMs. Specifically, (i) graph descriptions become verbose, overwhelming LLMs, and (ii) only relying on attribute embeddings limits LLM's ability to capture the adequate graph structural information. These limitations make it difficult to model graphs both concisely and adequately using sole natural language with LLMs.

Inspired by the observation that LLMs pre-trained on one language can achieve exceptional performance on another with minimal additional training, we propose Graph-Defined Language for Large Language Model (GDL4LLM). This novel framework enables LLMs to transfer their powerful language understanding capabilities to graph-structured data. GDL4LLM translates the graph into a graph language corpus instead of graph descriptions and pre-trains LLMs on this corpus to adequately understand the graph. This corpus represents the subgraph centered around target nodes concisely with only a few tokens during fine-tuning on downstream tasks. By treating the graph as a new language, GDL4LLM enables LLMs to model text-attributed graph adequately and concisely. Extensive experiments on five datasets demonstrate that GDL4LLM outperforms description-based and embedding-based baselines by efficiently modeling different orders of neighbors.

Each graph is a new language: Graph Learning with LLMs

In light of the widespread deployment of Large Language Models (LLMs), the responsibility for safeguarding and regulating LLM-generated content has taken on heightened significance. Recent advancements in LLM-based moderation methods, e.g., LlamaGuard, have demonstrated remarkable promise in identifying safety risks associated with both inputs and outputs in human-AI interactions. However, integrating LLM-based safeguards into a chatbot system requires an additional inference stage involving a moderation LLM with billions of parameters, which significantly increases computational costs and reduces overall efficiency. In this paper, we demonstrate that simply learning a classification head on the last-layer hidden states of the dialogue model provides a strong capability to identify harmful contents. The classification head, referred to as ShieldHead, serves as an auxiliary branch paralleled with next-token-prediction LM head, enabling the detection of potential risks in past text sequences. Additionally, a label disambiguation technique is employed to supervise ShieldHead with both token-level and sentence-level labels, which further enhances its performance. ShieldHead exhibits remarkable efficiency during inference, providing real-time moderation results alongside token-wise streaming output during the chatbot system's decoding phase. Extensive experimental results demonstrate the superiority of the proposed framework: a state-of-the-art performance on the XSTest and SafeRLHF datasets while running at a speed about **300×** faster (**<1ms**) than previous LLM-based moderation models with **~99%** less parameters of LlamaGuard.

ShieldHead: Decoding-time Safeguard for Large Language Models

Post-earnings volatility prediction is critical for investors, with previous works often leveraging earnings call transcripts under the assumption that their rich semantics contribute significantly. To further investigate how transcripts impact volatility, we introduce DEC, a dataset featuring accurate volatility calculations enabled by the previously overlooked beforeAfterMarket attribute and dense ticker coverage. Unlike established benchmarks, where each ticker has only around two earnings, DEC provides 20 earnings records per ticker. Using DEC, we reveal that post-earnings volatility undergoes significant shifts, with each ticker displaying a distinct volatility distribution. To leverage historical post-earnings volatility and capture ticker-specific patterns, we propose two training-free baselines: Post-earnings Volatility (PEV) and Same-ticker Post-earnings Volatility (STPEV). These baselines surpass all transcripts-based models on DEC as well as on established benchmarks. Additionally, we demonstrate that current transcript representations predominantly capture ticker identity rather than offering financially meaningful insights specific to each earnings. This is evidenced by two key observations: earnings representations from the same ticker exhibit significantly higher similarity compared to those from different tickers, and predictions from transcript-based models show strong correlations with prior post-earnings volatility.

Downloads

Next from ACL 2025

CROSSAGENTIE: Cross-Type and Cross-Task Multi-Agent LLM Collaboration for Zero-Shot Information Extraction

Similar lecture

Compositional Steering of Large Language Models with Steering Tokens