United States

Recent advances in Multimodal Large Reasoning Models (MLRMs) have enabled explicit chain-of-thought inference across vision and language, substantially improving performance on complex reasoning tasks. Despite these gains, the reasoning process introduces a subtle yet critical vulnerability. We identify an underexplored multimodal safety failure mode in which harmful objectives are embedded within ostensibly benign contexts, leading models to over-prioritize narrative coherence during reasoning. We term this phenomenon Safety Context Amnesia (SCA), wherein models correctly perceive risk-relevant visual cues but fail to enforce safety constraints as the reasoning process becomes dominated by contextual alignment. To mitigate SCA, we propose Intent-Guided Safety Reasoning (IGSR), an inference-time defense that operates without modifying target model parameters. IGSR employs a Perception Decoupler to extract objective visual evidence into a structured intent output, followed by a Cognitive Arbiter that enforces explicit safety constraints prior to generation. Extensive experiments across multiple multimodal safety benchmarks demonstrate that IGSR improves defense success rates by over 62% compared to baselines, while largely preserving task utility. These results highlight the critical role of structured, intent-aware reasoning in achieving robust safety reasoning for multimodal reasoning models.

ACL 2026

Mitigating Safety Context Amnesia in Multimodal Reasoning Models via Intent-Guided Safety Reasoning

## 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.

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.

Complex reasoning problems often involve implicit spatial and geometric relationships that are not explicitly encoded in text. While recent reasoning models perform well across many domains, purely text-based reasoning struggles to capture structural constraints in complex settings. In this paper, we introduce FIGR, which integrates executable visual construction into multi-turn reasoning via end-to-end reinforcement learning. Rather than relying solely on textual chains of thought, FIGR externalizes intermediate hypotheses by generating executable code that constructs diagrams within the reasoning loop. An adaptive reward mechanism selectively regulates when visual construction is invoked, enabling more consistent reasoning over latent global properties that are difficult to infer from text alone. Experiments on seven challenging mathematical benchmarks demonstrate that FIGR outperforms strong text-only chain-of-thought baselines, improving the base model by 13.12% on AIME 2025 and 11.00% on BeyondAIME. These results highlight the effectiveness of precise, controllable figure construction of FIGR in enhancing complex reasoning ability.

Figure It Out: Improve the Frontier of Reasoning with Executable Visual States

Generative reward models (GRMs) have emerged as a promising approach for aligning Large Language Models (LLMs) with human preferences by offering greater representational capacity and flexibility than traditional scalar reward models. However, GRMs face two major challenges: reliance on costly human-annotated data restricts scalability, and self-training approaches often suffer from instability and vulnerability to reward hacking. To address these issues, we propose ConsistRM, a self-training framework that enables effective and stable GRM training without human annotations. ConsistRM incorporates the Consistency-Aware Answer Reward, which produces reliable pseudo-labels with temporal consistency, thereby providing more stable model optimization. Moreover, the Consistency-Aware Critique Reward is introduced to assess semantic consistency across multiple critiques and allocates fine-grained and differentiated rewards. Experiments on five benchmark datasets across four base models demonstrate that ConsistRM outperforms vanilla Reinforcement Fine-Tuning (RFT) by an average of 1.5%. Further analysis shows that ConsistRM enhances output consistency and mitigates position bias caused by input order, highlighting the effectiveness of consistency-aware rewards in improving GRMs.Our implementation is available at https://github.com/yuliangCarmelo/ConsistRM.

ConsistRM: Improving Generative Reward Models via Consistency-Aware Self-Training

Low-Rank Adaptation (LoRA) has emerged as a parameter-efficient approach for fine-tuning large language models. However, conventional LoRA adapters are typically trained for a single task, limiting their applicability in real-world settings, where inputs may span multiple, diverse task domains. At inference time, existing methods can combine multiple LoRAs to improve cross-task performance, but they require additional labeled data or task-specific training, which is expensive at scale.

In this work, we introduce LoRA on the Go (LoGo), a training-free framework that dynamically selects and merges adapters at the instance level without any additional requirements. LoGo leverages signals extracted from a single forward pass through LoRA adapters, to identify the most relevant adapters and determine their contributions on-the-fly. Across 5 NLP benchmarks, 27 datasets, and 3 model families, LoGo outperforms training-based baselines on some tasks upto a margin of 3.6% while remaining competitive on other tasks and maintaining inference throughput, highlighting its effectiveness and practicality.

LoRA on the Go: Instance-level Dynamic LoRA Selection and Merging

Despite achieving high accuracy on medical benchmarks, LLMs exhibit the Einstellung Effect in clinical diagnosis—relying on statistical shortcuts rather than patient-specific evidence, causing misdiagnosis in atypical cases. Existing benchmarks fail to detect this critical failure mode. We introduce MedEinst, a counterfactual benchmark with 5,383 paired clinical cases across 49 diseases. Each pair contains a control case and a "trap" case with altered discriminative evidence that flips the diagnosis. We measure susceptibility via Bias Trap Rate—probability of misdiagnosing traps despite correctly diagnosing controls. Evaluation shows frontier models achieve high baseline accuracy but severe bias trap rates. Thus, we propose ECR-Agent, aligning LLM reasoning with Evidence-Based Medicine via two components: (1) Dynamic Causal Inference (DCI) performs structured reasoning through dual-pathway perception, dynamic causal graph reasoning across three levels (association, intervention, counterfactual), and evidence audit for final diagnosis; (2) Critic-Driven Graph & Memory Evolution (CGME) iteratively refines the system by storing validated reasoning paths in an exemplar base and consolidating disease-specific knowledge into evolving illness graphs. Source code is to be released.

MedEinst: Benchmarking the Einstellung Effect in Medical LLMs through Counterfactual Differential Diagnosis

Post-training hybridization of large language models (LLMs) often replaces quadratic self-attention with sliding-window attention (SWA) to reduce KV cache usage and improve latency. Existing hybridization schemes are typically defined either at the layer level (e.g., interleaving) or at the head level via static rankings from local to global. Layer-level schemes ignore that local and global dependencies are routed through heads within the same layer, while static head-level rankings suffer from entanglement: a head’s local/global behavior can change after hybridization. We propose BOSCH, Black-box Binary Optimization for Short-context Head Selection, a training-free method that formulates the problem as a Large Neighborhood Search and decomposes it into three subproblems: (i) layer-importance detection via small-budget black-box probes, (ii) adaptive per-layer SWA-ratio assignment based on these sensitivities, and (iii) grouped head-level optimization within ratio buckets. Extensive experiments on 4 LLMs ranging from 1.7B to 30B parameters, across 4 SWA ratios, show that BOSCH consistently outperforms layer-level heuristics and 6 strong static head-level methods, with larger gains at higher SWA ratios. Under continual pretraining, BOSCH recover original long-context performance faster and to a higher level. Analysis of the selected heads reveals substantial turnover for BOSCH across different SWA ratios, underscoring the importance of performing head-level selection for each target ratio rather than relying on fixed locality rankings.

 BOSCH: Black-Box Binary Optimization for Short-Context Attention-Head Selection in LLMs

Iconicity, the resemblance between linguistic form and meaning, is pervasive in sign languages, offering a natural testbed for visual grounding in vision–language models (VLMs). We introduce the Visual Iconicity Challenge, a video-based benchmark that adapts psycholinguistic measures to evaluate VLMs on three tasks: (i) phonological sign-form prediction, (ii) transparency (inferring meaning from visual form), and (iii) graded iconicity ratings. We assess 17 state-of-the-art VLMs in zero- and few-shot settings on Sign Language of the Netherlands and compare them to human baselines. VLMs mirror human phonological difficulty patterns (e.g., handshape harder than location) and achieve moderate to strong alignment with human iconicity ratings. However, they still fail to infer lexical meaning from visual form alone and show a systematic object-based bias that inverts the human preference for action-based signs. Crucially, models with stronger phonological form prediction correlate better with human iconicity judgments, indicating shared sensitivity to visually grounded structure. Our findings validate these diagnostic tasks, show that explicit reasoning narrows the open-to-closed-model calibration gap, and motivate human-centric signals for modelling iconicity in multimodal models.

The Visual Iconicity Challenge: Evaluating Vision-Language Models on Sign Language Form–Meaning Mapping

Can generative agents be trusted in multimodal environments? Despite recent advances, agents remain limited in their ability to reason about safety, coherence, and trust across modalities. We introduce a reproducible simulation framework to evaluate generative agents in three aspects: (1) safety improvement over time via iterative plan revision in multimodal scenarios; (2) detection of unsafe activities across social contexts; and (3) social dynamics, measured through interaction and acceptance rates. These multimodal agents are evaluated using metrics that quantify plan revisions and unsafe-to-safe conversions. Experiments show that while agents detect direct multimodal contradictions, they often fail to align local revisions with global safety, achieving only a 55% success rate in correcting unsafe plans. We release a dataset of 1,000 multimodal plans, yielding more than 600,000 simulation steps. Notably, 45% of unsafe actions are accepted when paired with misleading visual cues, revealing a strong tendency to overtrust visual content. Code is available at https://github.com/AdonaiVera/X-CASE

Multimodal Safety Evaluation in Generative Agent Social Simulations

Behavior trees provide a transparent and modular structure for encoding expert-designed policies, enabling interpretable decision-making in complex tasks. Yet, applying behavior trees to high-dimensional perceptual inputs such as images or language is challenging as defining symbolic predicates over raw perceptual data is non-trivial. While state-of-the-art large multimodal models (such as vision-language models) can overcome this issue by utilizing natural language queries over perceptual inputs, they incur high computational cost, making them unsuitable for many applications. Imitation learning offers a way to distill these expert models into compact models, though it requires extensive supervision. In contrast, reinforcement learning reduces the need for costly supervision but risks misalignment of condition nodes with their intended semantics as well as poor credit assignment. To address these challenges, we introduce CERL (Condition-node Expert-regularized Reinforcement Learning), a framework that leverages expert-regularized reinforcement learning to preserve semantic faithfulness, while employing a factorized policy that aggregates sequential condition-node decisions into a single decision unit to alleviate credit assignment challenges. Experiments across seven tasks from the GymCards, FrozenLake, and BabyAIText suites demonstrate that our framework outperforms pure imitation learning or reinforcement learning baselines, retains strong agreement with expert decisions, and achieves substantial gains in inference speed and model size over expert models. Our implementation is available in https://github.com/HyosikMoon/CERL.

Lightweight and Faithful Visual Condition Checking in Behavior Trees via Expert-Regularized Reinforcement Learning

When answering subjective questions, an ideal LLM should surface diverse plausible perspectives rather than favoring a single viewpoint, a characteristic known as pluralism. Recent studies show that modern LLMs optimized through preference alignment systematically favor certain positions on subjective queries, making pluralism evaluation increasingly important. However, existing evaluation methods focus dominantly on multiple-choice and question-answering tasks, leaving open-ended generation largely unaddressed.

We propose PLURALEVAL, an evaluation framework that assesses LLM pluralism in open-ended generation by comparing outputs against free-form crowd responses. Our approach decomposes ground-truth responses into atomic, non-overlapping claims, then evaluates whether LLMs adequately cover this diverse claim space. We then introduce WildSCOPE, a multi-domain dataset of natural crowd responses, and demonstrate that PLURALEVAL captures novel insights, such as the collapse of pluralism through sycophancy, where LLM systematically degrades in overton pluralism when a user's belief is revealed. Finally, we discuss the value and actionable insights for preserving and encouraging pluralism from LLM deployers' side.

Evaluating Language Model Pluralism through In-the-wild Crowd Discussions

A central goal for mechanistic interpretability has been to identify the right units of analysis in large language models (LLMs) that causally explain their outputs. While early work focused on individual neurons, evidence that neurons often encode multiple concepts has motivated a shift toward analyzing directions in activation space. A key question is how to find directions that capture interpretable features in an unsupervised manner. Current methods rely on dictionary learning with sparse autoencoders (SAEs), commonly trained over residual stream activations to learn directions from scratch. However, SAEs often struggle in causal evaluations and lack intrinsic interpretability, as their learning is not explicitly tied to the computations of the model. Here, we tackle these limitations by directly decomposing MLP activations with semi-nonnegative matrix factorization (SNMF), such that the learned features are (a) sparse linear combinations of co-activated neurons, and (b) mapped to their activating inputs, making them directly interpretable. Experiments on Llama 3.1, Gemma 2 and GPT-2 show that SNMF derived features outperform SAEs and a strong supervised baseline (difference-in-means) on causal steering, while aligning with human-interpretable concepts. Further analysis reveals that specific neuron combinations are reused across semantically-related features, exposing a hierarchical structure in the MLP's activation space. Together, these results position SNMF as a simple and effective tool for identifying interpretable features and dissecting concept representations in LLMs.

Downloads

Next from ACL 2026

Figure It Out: Improve the Frontier of Reasoning with Executable Visual States

Similar lecture

Think in Safety: Unveiling and Mitigating Safety Alignment Collapse in Multimodal Large Reasoning Model

Stay up to date with the latest Underline news!

PRESENTATIONS

CONFERENCES

COMPANY

RESOURCES