United States

As reinforcement learning continues to scale the training of large language model–based agents, reliably verifying agent behaviors in complex environments has become increasingly challenging. Existing approaches rely on rule-based verifiers or LLM-as-a-Judge models, which struggle to generalize beyond narrow domains. Agent-as-a-Judge addresses this limitation by actively interacting with environments and tools to acquire verifiable evidence, yet its capabilities remain underexplored.

We introduce a benchmark AJ-Bench to systematically evaluate Agent-as-a-Judge across three domains—search, data systems, and graphical user interfaces—comprising 155 tasks and 516 annotated trajectories. The benchmark comprehensively assesses judge agents’ abilities in information acquisition, state verification, and process verification. Experiments demonstrate consistent performance gains over LLM-as-a-Judge baselines, while also revealing substantial open challenges in agent-based verification. Our data and code are available at https://aj-bench.github.io/.

ACL 2026

AJ-Bench: Benchmarking Agent-as-a-Judge for Environment-Aware Evaluation

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

EMNLP 2025

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

technical paper

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

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

Inference-time scaling through multiple sample generation in combination with Process- or Outcome-Reward Model (PRM or ORM) re-ranking has proven effective for text-based reasoning in large language models. This paper investigates whether such established techniques can be successfully adapted to reasoning in the continuous space, using COCONUT~\cite{hao2024traininglargelanguagemodels} continuous space reasoning LM as the backbone. We demonstrate the feasibility of generating diverse reasoning paths through dropout-based sampling. Our Pass@N analysis on the generated samples reveals the potential that could enable a significant gain in performance akin to gains observed in the discrete space. However, we highlight unique challenges faced for materializing this gain in the continuous thought space. In particular, working recipes for data generation and training PRM and ORM models in the discrete space unlocks only marginal improvements in the continuous space. Through probing various aspects including geometric properties and trajectory dynamics, we identify the underlying reasons that prevent effective discrimination between correct and incorrect reasoning (essential for the functioning of PRM and ORM). Our findings reveal that current limitations stem from the absence of key inductive biases in continuous thought representations.

Towards Inference-time Scaling for Continuous Space Reasoning

Efficient inference in Large Vision-Language Models is constrained by the high cost of processing thousands of visual tokens, yet it remains unclear which tokens and computations can be safely removed. While attention scores are commonly used to estimate visual token importance, they are an imperfect proxy for actual contribution. We show that Attention Contribution, which weights attention probabilities by value vector magnitude, provides a more accurate criterion for visual token selection. Our empirical analysis reveals that visual attention sinks are functionally heterogeneous, comprising Probability Dumps with low contribution that can be safely pruned, and Structural Anchors with high contribution essential for maintaining model performance. Further, we identify substantial redundancy in Feed-Forward Networks (FFNs) associated with visual tokens, particularly in intermediate layers where image tokens exhibit linear behavior. Based on our findings, we introduce CAPA (Contribution-Aware Pruning and FFN Approximation), a dual-strategy framework that prunes visual tokens using attention contribution at critical functional transitions and reduces FFN computation through efficient linear approximations. Experiments on various benchmarks across baselines show that CAPA achieves competent efficiency--performance trade-offs with improved robustness.

CAPA: Contribution-Aware Pruning and FFN Approximation\\for Efficient Large Vision-Language Models

Advances in mechanistic interpretability have identified special attention heads, known as retrieval heads, that are responsible for retrieving information from the context. However, the role of these retrieval heads in improving model performance remains unexplored. This work investigates whether retrieval heads can be leveraged to enhance the long-context capabilities of LLMs. Specifically, we propose RetMask, a method that generates training signals by contrasting normal model outputs with those from an ablated variant in which the retrieval heads are masked. This mechanism-based approach achieves substantial improvements: +2.28 points on HELMET at 128K for Llama-3.1, with +70\% gains on generation with citation and +32\% on passage re-ranking, while preserving performance on general tasks. Experiments across three model families demonstrate that RetMask consistently improves long-context performance, with gains correlating with the sparsity of the retrieval score distribution: models with sparser distributions, where retrieval capabilities are concentrated in a small set of heads, respond more strongly, while those with less sparse distributions show more modest gains. These results validate the functional role of retrieval heads and show that mechanistic insights can be transformed into performance enhancements.

From Interpretability to Performance: Optimizing Retrieval Heads for Long-Context Language Models

Speech processing for low-resource dialects remains a fundamental challenge in developing inclusive and robust speech technologies. Despite its linguistic significance and large speaker population, the Wu dialect of Chinese has long been hindered by the lack of large-scale speech data, standardized evaluation benchmarks, and publicly available models. In this work, we present WenetSpeech-Wu, the first large-scale, multi-dimensionally annotated open-source speech corpus for the Wu dialect, comprising approximately 8,000 hours of diverse speech data. Building upon this dataset, we introduce WenetSpeech-Wu-Bench, the first standardized and publicly accessible benchmark for systematic evaluation of Wu dialect speech processing, covering automatic speech recognition (ASR), Wu-to-Mandarin translation, speaker attribute prediction, speech emotion recognition, text-to-speech (TTS) synthesis, and instruction-following TTS (instruct TTS). Furthermore, we release a suite of strong open-source models trained on WenetSpeech-Wu, establishing competitive performance across multiple tasks and empirically validating the effectiveness of the proposed dataset. Together, these contributions lay the foundation for a comprehensive Wu dialect speech processing ecosystem, and we open-source proposed datasets, benchmarks, and models to support future research on dialectal speech intelligence.

WenetSpeech-Wu: Datasets, Benchmarks, and Models for a Unified Chinese Wu Dialect Speech Processing Ecosystem

Agentic search requires large language models (LLMs) to perform multi-step search to solve complex information-seeking tasks, imposing unique challenges on their reasoning capabilities. However, what constitutes effective reasoning for agentic search and how it can be learned remains unclear. In this work, we first investigate the reasoning behaviors that enable success in agentic search. By comparing successful and failed trajectories via an LLM-based analysis pipeline, we identify four beneficial behaviors: Information Verification, Authority Evaluation, Adaptive Search, and Error Recovery. Building on this, we propose Behavior Priming, a training approach that equips agentic search models with these reasoning behaviors before reinforcement learning (RL). Specifically, it collects trajectories with the identified behaviors for supervised fine-tuning (SFT), and then applies standard RL to further improve task performance. Experiments on Qwen3-1.7B and Llama3.2-3B-Instruct show that Behavior Priming yields relative improvements over direct RL by 37.2% on three web benchmarks and 6.2% on seven multi-hop QA benchmarks, and outperforms the SFT-then-RL baseline using outcome-correct trajectories for fine-tuning. Crucially, we show that these reasoning behaviors matter more than outcome correctness in the priming stage prior to RL. Further analysis reveals that Behavior Priming enhances exploration (pass@8) and test-time scaling (search step number), providing a robust foundation for RL.

Beneficial Reasoning Behaviors in Agentic Search and Effective Training Methods to Obtain Them

Aligning Large Vision-Language Models (LVLMs) to mitigate hallucinations typically relies on high-quality preference data. However, in self-supervised settings, standard binary preference optimization (e.g., DPO) suffers from noisy supervision and semantic ambiguity, as automatically generated chosen responses are not guaranteed to be superior to rejected ones. In this work, we propose \textbf{Trident}, a fully self-supervised framework that ensures robust alignment via a structured triplet paradigm. Trident autonomously constructs reliable preference triplets—comprising semantically enriched (chosen), degraded (rejected), and neutral (anchor) responses—through automated visual perturbations and self-summarization. We further introduce Trident Preference Regularization (TPR), a novel objective that utilizes an adaptive margin to enforce semantic separation between the triplet components while preventing deviation from the pretrained distribution. Despite requiring no human annotations or external reward models, Trident consistently outperforms state-of-the-art RLHF and RLAIF baselines. For instance, on LLaVA-1.5-7B, it reduces the hallucination rate on AMBER to 11.3% and achieves 95.70% precision on POPE using only 4k self-generated triplets and a single epoch. This validates structured triplet supervision as a scalable paradigm for robust self-supervised alignment.

Trident: Self-Supervised Preference Alignment via Triplet Regularization

Despite significant advances in Large Reasoning Models (LRMs) driven by reinforcement learning with verifiable rewards (RLVR), this paradigm is fundamentally limited in specialized or novel domains where such supervision is prohibitively expensive or unavailable, posing a key challenge for test-time adaptation. 
While existing test-time methods offer a potential solution, they are constrained by learning from static query sets, risking overfitting to textual patterns. 
To address this gap, we introduce Test-Time Variational Synthesis (TTVS), a novel framework that enables LRMs to self-evolve by dynamically augmenting the training stream from unlabeled test queries. TTVS comprises two synergistic modules: (1) Online Variational Synthesis, which transforms static test queries into a dynamic stream of diverse, semantically-equivalent variations, enforcing the model to learn underlying problem logic rather than superficial patterns; (2) Test-time Hybrid Exploration, which balances accuracy-driven exploitation with consistency-driven exploration across synthetic variants. Extensive experiments show TTVS yields superior performance across eight model architectures. Notably, using only unlabeled test-time data, TTVS not only surpasses other test-time adaptation methods but also outperforms state-of-the-art supervised RL-based techniques trained on vast, high-quality labeled data.

TTVS: Boosting Self-Exploring Reinforcement Learning via Test-time Variational Synthesis

Large Audio Language Models (LALMs) employing the Chain-of-Thought paradigm have demonstrated remarkable reasoning capabilities. Though different problems naturally require varying depths of reasoning, existing methods often determine whether to perform reasoning, lacking fine-grained mechanisms to adapt reasoning length to problem complexity. As a result, LALMs often adopt a one-size-fits-all reasoning strategy, leading to redundant overthinking for simple tasks and insufficient reasoning for complex ones. In this paper, we conduct an in-depth analysis of LALM reasoning behavior and argue that effective and efficient reasoning should be adaptively aligned with task difficulty. To this end, we propose a difficulty-adaptive reasoning method for LALMs. Specifically, we introduce a reward function that dynamically links reasoning length to the model's perceived problem difficulty, encouraging shorter reasoning for easy tasks and longer reasoning for more complex ones. Extensive experiments on three datasets demonstrate that our method consistently improves performance while reducing average reasoning length by at least 50\%, achieving higher efficiency without sacrificing accuracy.

Think Smart, Not Hard: Difficulty Adaptive Reasoning for Large Audio Language Models

While scaling test-time compute can substantially improve model performance, existing approaches either rely on static compute allocation or sample from fixed generation distributions.
In this work, we introduce a test-time compute allocation framework that jointly adapts where computation is spent and how generation is performed. 
Our method begins with a warm-up phase that identifies easy queries and assembles an initial pool of question-response pairs from the test set itself. An adaptive phase then concentrates further computation on unresolved queries while reshaping their generation distributions through evolving in-context demonstrations---conditioning each generation on successful responses from semantically related queries rather than resampling from a fixed distribution.
Experiments across math, coding, and reasoning benchmarks demonstrate that our approach consistently outperforms existing baselines while consuming substantially less inference-time compute.

Adaptive Test-Time Compute Allocation with Evolving In-Context Demonstrations

Scaling training compute, measured in FLOPs, has long been shown to improve the accuracy of large language models, yet training remains resource-intensive. Prior work shows that increasing test-time compute (TTC)—for example through iterative sampling—can allow smaller models to rival or surpass much larger ones at lower overall cost. We introduce TTC-aware training, where an intermediate checkpoint and a corresponding TTC configuration can together match or exceed the accuracy of a fully trained model while requiring substantially fewer training FLOPs. Building on this insight, we propose an early stopping algorithm that jointly selects a checkpoint and TTC configuration to minimize training compute without sacrificing accuracy. To make this practical, we develop an efficient TTC evaluation method that avoids exhaustive search, and we formalize a break-even bound that identifies when increased inference compute compensates for reduced training compute. Experiments demonstrate up to 92\% reductions in training FLOPs while maintaining and sometimes remarkably improving accuracy. These results highlight a new perspective for balancing training and inference compute in model development, enabling faster deployment cycles and more frequent model refreshes.

Downloads

Next from ACL 2026

Towards Inference-time Scaling for Continuous Space Reasoning

Similar lecture

From Generation to Judgment: Opportunities and Challenges of LLM-as-a-judge