Austria

Most discussions about Large Language Model (LLM) safety have focused on single-agent settings but multi-agent LLM systems now create novel adversarial risks because their behavior depends on communication between agents and decentralized reasoning. 
In this work, we innovatively focus on attacking pragmatic systems that have constrains such as limited token bandwidth, latency between message delivery, and defense mechanisms. We design a $\textit{permutation-invariant adversarial attack}$ that optimizes prompt distribution across latency and bandwidth-constraint network topologies to bypass distributed safety mechanisms within the system. Formulating the attack path as a problem of $\textit{maximum-flow minimum-cost}$, coupled with the novel $\textit{Permutation-Invariant Evasion Loss (PIEL)}$, we leverage ${graph-based optimization}$ to maximize attack success rate while minimizing detection risk. Evaluating across models including $\texttt{Llama}$, $\texttt{Mistral}$, $\texttt{Gemma}$, $\texttt{DeepSeek}$ and other variants on various datasets like $\texttt{JailBreakBench}$ and $\texttt{AdversarialBench}$, our method outperforms conventional attacks by up to $7\times$, exposing critical vulnerabilities in multi-agent systems. Moreover, we demonstrate that existing defenses, including variants of $\texttt{Llama-Guard}$ and $\texttt{PromptGuard}$, fail to prohibit our attack, emphasizing the urgent need for multi-agent specific safety mechanisms.

ACL 2025

Agents Under Siege: Breaking Pragmatic Multi-Agent LLM Systems with Optimized Prompt Attacks

multi-agent systems.

adversarial attacks

## 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.

technical paper

### 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) are cutting-edge generative AI models built on transformer architecture, which tend to be highly memory-intensive when performing real-time inference. Various strategies have been developed to enhance the end-to-end inference speed for LLMs, one of which is speculative decoding. This technique involves running a smaller LLM (draft model) for inference over a defined window size, denoted as $\gamma$, while simultaneously being validated by the larger LLM (target model). Choosing the optimal $\gamma$ value and the draft model is essential for unlocking the potential of speculative decoding. But it is difficult to do due to the complicated influence from various factors, including the nature of the task, the hardware in use, and the combination of the large and small models. This paper introduces *on-the-fly adaption of speculative decoding*, a solution that dynamically adapts the choices to maximize the efficiency of speculative decoding for LLM inferences. As a drop-in solution, it needs no offline benchmarking or training. Experiments show that the solution can lead to 3.55-16.48% speed improvement over the standard speculative decoding, and 1.2-3.4$\times$ over the default LLMs.

A Drop-In Solution for On-the-Fly Adaptation of Speculative Decoding in Large Language Models

Despite their outstanding capabilities, large language models (LLMs) are prone to hallucination and producing factually incorrect information. This challenge has spurred efforts in attributed text generation, which prompts LLMs to generate content with supporting evidence. In this paper, we propose a novel framework, called Think&Cite, and formulate attributed text generation as a multi-step reasoning problem integrated with search. Specifically, we propose Self-Guided Monte Carlo Tree Search (SG-MCTS), which capitalizes on the self-reflection capability of LLMs to reason about the intermediate states of MCTS for guiding the tree expansion process. To provide reliable and comprehensive feedback, we introduce Progress Reward Modeling to measure the progress of tree search from the root to the current state from two aspects, i.e., generation and attribution progress. We conduct extensive experiments on three datasets and the results show that our approach significantly outperforms baseline approaches.

Think&Cite: Improving Attributed Text Generation with Self-Guided Tree Search and Progress Reward Modeling

Recent research has shown that CLIP models struggle with visual reasoning tasks that require grounding compositionality, understanding spatial relationships, or capturing fine-grained details. One natural hypothesis is that the CLIP vision encoder does not embed essential information for these tasks. However, we find that this is not always the case: The encoder gathers query-relevant visual information, while CLIP fails to extract it. In particular, we show that another branch of Vision-Language Models (VLMs), Generative Multimodal Large Language Models (MLLMs), achieve significantly higher accuracy than CLIP in many of these tasks using the *same* vision encoder and weights, indicating that these Generative MLLMs *perceive more*—as they extract and utilize visual information more effectively. We conduct a series of controlled experiments and reveal that their success is attributed to multiple key design choices, including patch tokens, position embeddings, and prompt-based weighting. On the other hand, enhancing the training data alone or applying a stronger text encoder does not suffice to solve the task, and additional text tokens offer little benefit. Interestingly, we find that fine-grained visual reasoning is not exclusive to generative models trained by an autoregressive loss: When converted into CLIP-like encoders by contrastive finetuning, these MLLMs still outperform CLIP under the same cosine similarity-based evaluation protocol. Our study highlights the importance of VLM architectural choices and suggests directions for improving the performance of CLIP-like contrastive VLMs.

Exploring How Generative MLLMs Perceive More Than CLIP with the Same Vision Encoder

Unsupervised cross-domain keyphrase generation is crucial in real-world natural language processing scenarios. However, the accuracy of up-to-date approaches is limited by the distribution shift between source and target domain, which stems from the cross-domain field. Large language models (LLMs) offer potential for the cross-domain keyphrase generation tasks due to their strong generalization abilities, facilitated by providing demonstrations relevant to the target task. Nevertheless, it is often difficult to obtain labeled samples from the target domain. To address this challenge, this paper aims to seek rational demonstrations from the source domain, thereby improving the LLMs' ability in the unsupervised cross-domain keyphrase generation setting. Specifically, we design a novel domain-aware retrieval model on the source domain. Guided by insights from domain generalization theory, we introduce two generalization terms, one for cross-domain relevance and another for each domain consistency to better support retrieval of rational demonstrations. By the retrieved source-domain demonstrations and distance-based relevant score, the proposed approach achieves optimal accuracy. Comprehensive experiments on widely used cross-domain KG benchmarks demonstrate our approach’s state-of-the-art performance and effectiveness.

Seeking Rational Demonstrations for Large Language Models: A Domain Generalization Approach to Unsupervised Cross-Domain Keyphrase Generation

Large language models excel at legal text generation but often produce hallucinations due to their sole reliance on parametric knowledge. Retrieval-augmented models mitigate this by providing relevant external documents to the model but struggle when retrieval is based only on past context, which may not align with the model’s intended future content. We introduce LexKeyPlan, a novel framework that integrates anticipatory planning into generation. Instead of relying solely on context for retrieval, LexKeyPlan generates keyphrases outlining future content serving as forward-looking plan, guiding retrieval for more accurate text generation. This work incorporates planning into legal text generation, demonstrating how keyphrases—representing legal concepts—enhance factual accuracy. By structuring retrieval around legal concepts, LexKeyPlan better aligns with legal reasoning, making it particularly suited for legal applications. Using the ECHR corpus as case study, we show that LexKeyPlan improves factual accuracy and coherence by retrieving information aligned with the intended content.

LexKeyPlan: Planning with Keyphrases and Retrieval Augmentation for Legal Text Generation: A Case Study on European Court of Human Rights Cases

Large language models (LLMs) offer powerful capabilities but come with significant environmental impact, particularly in carbon emissions. Existing studies benchmark carbon emissions but lack a standardized basis for comparison across different model configurations. To address this, we introduce the concept of functional unit (FU) as a standardized basis and develop FUEL, the first FU-based framework for evaluating LLM serving’s environmental impact. Through three case studies, we uncover key insights and trade-offs in reducing carbon emissions by optimizing model size, quantization strategy, and hardware choice, paving the way for more sustainable LLM serving. The code is available at https://github.com/jojacola/FUEL.

Unveiling Environmental Impacts of Large Language Model Serving: A Functional Unit View

The evaluation of large language models (LLMs) has traditionally relied on static benchmarks, a paradigm that poses two major limitations: (1) predefined test sets lack adaptability to diverse application domains, and (2) standardized evaluation protocols often fail to capture fine-grained assessments of domain-specific knowledge and contextual reasoning abilities. To overcome these challenges, we propose GuessArena, an adaptive evaluation framework grounded in adversarial game-based interactions. Inspired by the interactive structure of the Guess Who I Am? game, our framework seamlessly integrates dynamic domain knowledge modeling with progressive reasoning assessment to improve evaluation fidelity. Empirical studies across five vertical domains-finance, healthcare, manufacturing, information technology, and education-demonstrate that GuessArena effectively distinguishes LLMs in terms of domain knowledge coverage and reasoning chain completeness. Compared to conventional benchmarks, our method provides substantial advantages in interpretability, scalability, and scenario adaptability.

GuessArena: Guess Who I Am? A Self-Adaptive Framework for Evaluating LLMs in Domain-Specific Knowledge and Reasoning

The Theory of Multiple Intelligences underscores the hierarchical nature of cognitive capabilities. To advance Spatial Artificial Intelligence, we pioneer a psychometric framework defining five Basic Spatial Abilities (BSAs) in Visual Language Models (VLMs): Spatial Perception, Spatial Relation, Spatial Orientation, Mental Rotation, and Spatial Visualization. Benchmarking 13 mainstream VLMs through nine validated psychometric experiments reveals significant gaps versus humans, with three key findings: 1) VLMs mirror human hierarchies (strongest in 2D orientation, weakest in 3D rotation) with independent BSAs; 2) Many smaller models surpass larger counterparts, with Qwen leading and InternVL2 lagging; 3) Interventions like CoT and few-shot training show limits from architectural constraints, while ToT demonstrates the most effective enhancement. Identified barriers include weak geometry encoding and missing dynamic simulation. By linking Psychometrics to VLMs, we provide a comprehensive BSA evaluation benchmark, a methodological perspective for embodied AI development, and a cognitive science-informed roadmap for achieving human-like spatial intelligence.

Defining and Evaluating Visual Language Models' Basic Spatial Abilities: A Perspective from Psychometrics

Traditional retrieval methods rely on transforming user queries into vector representations and retrieving documents based on cosine similarity within an embedding space. While efficient and scalable, this approach often fails to handle complex queries involving logical constructs such as negations, conjunctions, and disjunctions. In this paper, we propose a novel inference-time logical reasoning framework that explicitly incorporates logical reasoning into the retrieval process. Our method extracts logical reasoning structures from natural language queries and then composes the individual cosine similarity matching scores to formulate the final document scores. This approach enables the retrieval process to handle complex logical reasoning without compromising computational efficiency. Our results on both synthetic and real-world benchmarks demonstrate that the proposed method consistently outperforms traditional retrieval methods across different models and datasets, significantly improving retrieval performance for complex queries.

Enhancing Retrieval Systems with Inference-Time Logical Reasoning

Implicit Discourse Relation Recognition (abbr., IDRR) is a NLP task of classifying argument pairs into different types of semantic relations. Arguments contain subtexts, some of which are beneficial to the perception of semantic relations. However, subtexts are connotative. The neural IDRR model fails to be aware of them without being given pertinent prompts. In this paper, we leverage LLaMA to generate subtexts for argument pairs, and verify the effectiveness of subtext-based IDRR. We construct an IDRR baseline using the decoder-only backbone LLaMA, and enhance it with subtext-aware relation reasoning. A confidence-diagnosed dual-channel network is used for collaboration between in-subtext and out-of-subtext IDRR. We experiment on PDTB-2.0 and PDTB-3.0 for both the main-level and secondary-level relation taxonomies. The test results show that our approach yields substantial improvements compared to the baseline, and achieves higher $F$1-scores on both benchmarks than the previous decoder-only IDRR models. We make the source codes and data publicly available.

Downloads

Next from ACL 2025

A Drop-In Solution for On-the-Fly Adaptation of Speculative Decoding in Large Language Models

Similar lecture

Generalists vs. Specialists: Evaluating Large Language Models for Urdu