United States

Do reasoning models have &quot;Aha!&quot; moments?
Prior work suggests that models like DeepSeek-R1-Zero undergo sudden mid-trace realizations that lead to accurate outputs, implying an intrinsic capacity for self-correction. 
Yet, it remains unclear whether such intrinsic shifts in reasoning strategy actually improve performance.
Here, we study mid-reasoning shifts and instrument training runs to detect them. 
Our analysis spans $1$M+ reasoning traces, hundreds of training checkpoints, three reasoning domains, and multiple decoding temperatures and model architectures.
We find that reasoning shifts are rare, do not become more frequent with training, and seldom improve accuracy, indicating that they do not correspond to prior perceptions of model insight. 
However, their effect varies with model uncertainty. Building on this finding, we show that artificially triggering extrinsic shifts under high entropy reliably improves accuracy. 
Our results show that mid-reasoning shifts are symptoms of unstable inference behavior rather than an intrinsic mechanism for self-correction.

ACL 2026

The Illusion of Insight in Reasoning Models

## 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.

Do reasoning models have "Aha!" moments?
Prior work suggests that models like DeepSeek-R1-Zero undergo sudden mid-trace realizations that lead to accurate outputs, implying an intrinsic capacity for self-correction. 
Yet, it remains unclear whether such intrinsic shifts in reasoning strategy actually improve performance.
Here, we study mid-reasoning shifts and instrument training runs to detect them. 
Our analysis spans $1$M+ reasoning traces, hundreds of training checkpoints, three reasoning domains, and multiple decoding temperatures and model architectures.
We find that reasoning shifts are rare, do not become more frequent with training, and seldom improve accuracy, indicating that they do not correspond to prior perceptions of model insight. 
However, their effect varies with model uncertainty. Building on this finding, we show that artificially triggering extrinsic shifts under high entropy reliably improves accuracy. 
Our results show that mid-reasoning shifts are symptoms of unstable inference behavior rather than an intrinsic mechanism for self-correction.

poster

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.

**Online Registration Form**: https://acl.swoogo.com/acl2026

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.

Large language models (LLMs) seem to offer an easy path to interpretability: just ask them to explain their answers. Yet the features driving an answer often differ from those emphasized in its explanation, meaning post-hoc rationales can misrepresent what actually shaped the model's output. We quantify this gap by comparing the feature-importance distributions of answers and their explanations. Prior analyses reveal such discrepancies, but large-scale study has been limited by the high computational cost of attribution methods. To address this, we introduce the Post-hoc Self-Consistency Bank (PSCB), a large-scale benchmark linking model decisions with diverse explanations and attribution vectors across datasets, methods, and model families. Using PSCB, we find that Spearman rank correlation provides a more reliable signal of alignment than cosine similarity. Building on this insight, we apply Direct Preference Optimization (DPO) to attribution-based preference data, improving alignment without degrading task accuracy, and show that standard supervised fine-tuning on the same data fails to achieve comparable gains. These improvements generalize robustly across domains, paving the way toward scalable and faithful alignment between LLM decisions and their natural language explanations.

Aligning What LLMs Do and Say: Towards Self-Consistent Explanations

Agentic Reinforcement Learning (Agentic RL) has achieved notable success in enabling agents to perform complex reasoning and tool use. However, most methods still relies on sparse outcome-based reward for training. Such feedback fails to differentiate intermediate reasoning quality, leading to suboptimal training results. In this paper, we introduce \textbf{Agent Reasoning Reward Model (Agent-RRM)}, a multi-faceted reward model that produces structured feedback for agentic trajectories, including (1) an explicit reasoning trace , (2) a focused critique that provides refinement guidance by highlighting reasoning flaws, and (3) an overall score that evaluates process performance. Leveraging these signals, we systematically investigate three integration strategies: \textbf{Reagent-C} (text-augmented refinement), \textbf{Reagent-R} (reward-augmented guidance), and \textbf{Reagent-U} (unified feedback integration). Extensive evaluations across 12 diverse benchmarks demonstrate that Reagent-U yields substantial performance leaps, achieving 43.7% on GAIA and 46.2% on WebWalkerQA, validating the effectiveness of our reasoning reward model and training schemes. Code, models, and datasets will be released to facilitate future research.

 Exploring Reasoning Reward Model for Agents

Code localization is a primary bottleneck in automated software development. While parallel tool execution can accelerate discovery, existing agents suffer from a 34.9% redundant tool invocation rate, negating the benefits of parallelism. We introduce FuseSearch, which reframes parallel code localization as a quality--efficiency co-optimization problem. By defining tool efficiency---the ratio of novel information gain to total invocations---we employ a two-stage SFT and RL pipeline to train models in adaptive parallel strategies. Unlike fixed-breadth methods, FuseSearch dynamically adjusts search breadth based on task context, transitioning from exploration to refinement. On SWE-bench Verified, FuseSearch-4B matches SOTA performance (84.7% file-level and 56.4% function-level 
$F_1$ scores) while being 93.6% faster, using 67.7% fewer turns and 68.9% fewer tokens. Our findings demonstrate that efficiency-aware training inherently boosts quality by eliminating noisy, redundant signals, enabling high-performance, low-cost localization agents. Code: \url{https://github.com/sxthunder/FuseSearch}

FuseSearch: Learning Adaptive Parallel Execution for Efficient Code Localization

The emergence of groundbreaking large language models capable of performing complex reasoning tasks holds significant promise for addressing various scientific challenges, including those arising in complex clinical scenarios. To enable their safe and effective deployment in real-world healthcare settings, it is urgently necessary to benchmark the diagnostic capabilities of current models systematically. Given the limitations of existing medical benchmarks in evaluating advanced diagnostic reasoning, we present DiagnosisArena, a comprehensive and challenging benchmark designed to rigorously assess professional-level diagnostic competence. DiagnosisArena consists of 1,113 pairs of segmented patient cases and corresponding diagnoses, spanning 28 medical specialties, deriving from clinical case reports published in 10 top-tier medical journals. The benchmark is developed through a meticulous construction pipeline, involving multiple rounds of screening and review by both AI systems and human experts, with thorough checks conducted to prevent data leakage. Our study reveals that even the most advanced reasoning models, o3-mini, o1, and DeepSeek-R1, achieve only 45.82%, 31.09%, and 17.79% accuracy, respectively. This finding highlights a significant generalization bottleneck in current large language models when faced with clinical diagnostic reasoning challenges. Through DiagnosisArena, we aim to drive further advancements in AI’s diagnostic reasoning capabilities, enabling more effective solutions for real-world clinical diagnostic challenges. We openly share the benchmark and evaluation tools for further research and development.

DiagnosisArena: Benchmarking Diagnostic Reasoning for Large Language Models

Identifying conditions that a certain drug takes therapeutic effect on a target disease is crucial for clinical decision-making support. However, most existing biomedical information extraction methods have focused on identifying only relations between drugs and diseases, while largely overlooking the context-specific conditions where such relations can apply. To address this problem, we introduce the task of applicability condition extraction for therapeutic drug--disease relations from biomedical research literature. We create the first dataset that has manually annotated triples of drugs, diseases, and applicability conditions on biomedical paper abstracts with 1,119 drug-disease pairs. Using this dataset, we systematically evaluate the performance of a range of existing methods. In addition, we propose a new method that enhances LoRA to consider relations between drugs and diseases. Our method consistently outperforms strong baselines across different evaluation settings.

Applicability Condition Extraction for Therapeutic Drug-Disease Relations

Editing presentation slides is a frequent yet tedious task, ranging from creative layout design to repetitive text maintenance. While recent GUI-based agents powered by Multimodal LLMs (MLLMs) excel at tasks requiring visual perception, such as spatial layout adjustments, they often incur high computational costs and latency when handling structured, text-centric, or batch processing tasks. In this paper, we propose Talk-to-Your-Slides, a high-efficiency slide editing agent that operates via language-driven structured data manipulation rather than relying on the image modality. By leveraging the underlying object model instead of screen pixels, our approach ensures precise content modification while preserving style fidelity, addressing the limitations of OCR-based visual agents. Our system features a hierarchical architecture that effectively bridges high-level user instructions with low-level execution codes. Experiments demonstrate that for text-centric and formatting tasks, our method enables 34% faster processing, achieves 34% better instruction fidelity, and operates at an 87% lower cost compared to GUI-based baselines. Furthermore, we introduce TSBench, a human-verified benchmark dataset comprising 379 instructions, including a Hard subset designed to evaluate robustness against complex and visually dependent queries. Our code and benchmark are available at https://drive.google.com/drive/folders/1onwp5m7t3207xZu7HEBTMpdivsiOuqG8?usp=share_link

 Talk to Your Slides: High-Efficiency Slide Editing via Language-Driven Structured Data Manipulation

Large language models (LLMs) are excellent at maintaining high-level, convincing dialogue, but it remains unclear whether their persuasive success reflects genuine understanding of the discourse. We examine this question through informal debates between humans and LLMs, first by measuring their persuasive skills, and then by relating these to their understanding of what is being talked about: namely, their comprehension of argumentative structures and the pragmatic context on the same debates. We find that LLMs effectively maintain coherent, persuasive debates, and can sway the beliefs of both participants and audiences. We also note that awareness or suspicion of AI involvement encourage people to be more critical of the arguments made. However, we also find that LLMs are unable to show comprehension of deeper dialogical structures, such as argument quality or existence of supporting premises. Our results reveal a disconnect between LLM comprehension and dialogical skills, raising ethical and practical concerns on their deployment on explanation-critical contexts. From an argumentation-theoretical perspective, we experimentally question whether an agent, if it can convincingly maintain a dialogue, is required to show it knows what is talking about.

The Thin Line Between Comprehension and Persuasion in LLMs

Agentic workflows solve complex tasks by orchestrating modular components (e.g., planning, reasoning, action, reflection) built on top of LLM backbones. A practical but underexplored question is model allocation: given a fixed workflow decomposition and a pool of candidate LLMs, which components should be upgraded (and with which models) to upgrade task performance, and how can we attribute gains to individual upgrades and their interactions?
We present ShapleyFlow, a cooperative game theoretic framework that models component upgrades as players and evaluates component coalitions to compute Shapley values. This yields interaction-aware attribution and supports Shapley-guided configuration recommendation for model allocation under a fixed workflow structure.
We further introduce CapaBench, a benchmark of 1,500+ tasks across seven domains (shopping, navigation, ticketing, mathematics, operating systems, robotic coordination, and automated theorem proving).
Across 9 representative LLMs and all $2^4$ upgrade coalitions in a 4-component workflow, ShapleyFlow provides (i) principled, interaction-aware attribution for modular workflows and (ii) actionable model-allocation recommendations that improve over strong single-model baselines.

 Attribution-Based Analysis and Optimization of Modular Agentic Workflows

Large Vision Language Models (LVLMs) have demonstrated remarkable capabilities, yet their proficiency in understanding and reasoning over multiple images remains largely unexplored. While existing benchmarks have initiated the evaluation of multi-image models, a comprehensive analysis of their core weaknesses and their causes is still lacking. In this work, we introduce MIMIC (Multi-Image Model Insights and Challenges), a new benchmark designed to rigorously evaluate the multi-image capabilities of LVLMs. Using MIMIC, we conduct a series of diagnostic experiments that reveal pervasive issues: LVLMs often fail to aggregate information across images and struggle to track or attend to multiple concepts simultaneously. To address these failures, we propose two novel complementary remedies. On the data side, we present a procedural data-generation strategy that composes single-image annotations into rich, targeted multi-image training examples. On the optimization side, we analyze layer-wise attention patterns and derive an attention-masking scheme tailored for multi-image inputs. Experiments substantially improved cross-image aggregation, while also enhancing performance on existing multi-image benchmarks, outperforming prior state of the art across tasks. Data and code will be made available.

More Images, More Problems? A Controlled Analysis of VLM Failure Modes.

Concept-based explanations quantify how high-level concepts (e.g., gender or experience) influence model behavior, which is crucial for decision-makers in high-stakes domains. Recent work evaluates the faithfulness of such explanations by comparing them to reference causal effects estimated from counterfactuals. In practice, existing benchmarks rely on costly human-written counterfactuals that serves as imperfect proxy. To address this, we introduce a framework for constructing datasets containing structural counterfactual pairs: LIBERTy (LLM-based Interventional Benchmark for Explainability with Reference Targets). LIBERTy is grounded in explicitly defined Structured Causal Models (SCMs) of the text generation, interventions on a concept propagate through the SCM until an LLM generates the counterfactual. We introduce three datasets (disease detection, CV screening, and workplace violence prediction) together with a new evaluation metric, order-faithfulness. Using them, we evaluate a wide range of methods across five models and identify substantial headroom for improving concept-based explanations. LIBERTy also enables systematic analysis of model sensitivity to interventions: we find that proprietary LLMs show markedly reduced sensitivity to demographic concepts, likely due to post-training mitigation. Overall, LIBERTy provides a much-needed benchmark for developing faithful explainability methods.

Downloads

Next from ACL 2026

Aligning What LLMs Do and Say: Towards Self-Consistent Explanations

Similar lecture

From Reasoning to Answer: Empirical, Attention-Based and Mechanistic Insights into Distilled DeepSeek R1 Models