2026· Annual Meeting of the Association for Computational Linguistics· pp. 39249-39276· 0 citations· 39 references
Computer Science
TL;DR
Intent-Guided Safety Reasoning (IGSR), an inference-time defense that operates without modifying target model parameters, is proposed, which improves defense success rates by over 62% compared to baselines, while largely preserving task utility.
Abstract
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 un-derexplored 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 Con-text 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 multi-modal reasoning models. Warning: This paper contains unsafe examples.
Recent advances in multimodal embodied agents have enabled long-horizon planning in visually rich environments via natural language. Yet, their generalization remains brittle when task instructions deviate from familiar examples, exposing a reliance on surface imitation rather than structural understanding. We propose Causal Abstraction Learning for Multi-Modal Grounded Planning (CALM), a framework that enhances planning agents with the ability to discover and exploit causal regularities across tasks. CALM incrementally develops a causal library by abstracting precondition–effect structure from successful executions, yielding compact representations that emphasize stable dependencies beyond incidental context. When execution diverges from expectation, these abstractions are refined through contrastive causal reasoning, enabling targeted adjustments that resolve underlying mechanism mismatch. The resulting structure serves as a transferable prior for planning in novel settings, integrating perceptual cues with mechanism-informed knowledge. Without retraining or task-specific heuristics, CALM generalizes robustly and efficiently to linguistic and perceptual variation. Experiments on ALFRED and VirtualHome demonstrate consistent gains, highlighting causal abstraction as a scalable inductive bias for grounded planning.
Xinshu Li, Shiyi Yang, Ziqi Xu et al.· Proceedings of the 32nd ACM...· 0 citations
Humans can infer hidden physical processes from sparse observations, yet current evaluation protocols for Vision Language Models fail to assess whether such physical reasoning is genuinely captured. To address this gap, we introduce Retrospective Physical Process Reasoning, a new evaluation paradigm to reason backward from outcomes under explicit physical constraints. Building on the paradigm, we present RetroHolmes, the first real-world benchmark for Retrospective Physical Process Reasoning, comprising object-centric image pairs annotated with reachability labels and causal step sequences across diverse physical transitions. Using RetroHolmes, we analyze state of the art Vision Language Models and uncover systematic failure modes, including judgment bias in reachability assessment and belief dominance over physical evidence, mirroring sycophancy behavior observed in large language models. We further demonstrate a simple analysis-by-synthesis instantiation with visual simulation as an intermediate step, validating the diagnostic value of RetroHolmes and highlighting the importance of physically grounded intermediate representations for physical reasoning.
Ruoxuan Zhang, Qiyun Zheng, Siyu Wu et al.· 0 citations
Large language models (LLMs) often produce reasoning steps that are superficially coherent yet internally inconsistent, leading to unreliable outputs. Since such failures typically arise from implicit or poorly-grounded knowledge, we introduce Grounded Reasoning in Dependency (GRiD) , a novel dependency-aware reasoning framework that explicitly grounds reasoning steps in structured knowledge. GRiD represents reasoning as a graph consisting of interconnected knowledge extraction nodes and reasoning nodes, enforcing logical consistency through explicit dependencies. Each reasoning step is validated via a lightweight, step-wise verifier that ensures logical correctness relative to its premises. Extensive experiments across diverse reasoning benchmarks—including StrategyQA, CommonsenseQA, GPQA, and TruthfulQA—demonstrate that GRiD substantially improves reasoning accuracy, consistency, and faithfulness compared to recent state-of-the-art structured reasoning methods. Notably, GRiD enhances performance even when applied purely as a lightweight verification module at inference time, underscoring its generalizability and practical utility † .
Xiangyu Wen, Min Li, Junhua Huang et al.· Neural Information Processin...· 2 citations
CDS is introduced, a training-free meta-reasoning framework equipped with residual demand assessment: at each step, an LLM-based progress evaluator characterizes the residual reasoning required to arrive at a solution rather than merely evaluating the previous step.
John Scoville, Shengzhuang Chen, Yejin Bang et al.· 0 citations
It is shown that evidence-gathering should be evaluated as a trajectory-level control problem, separately from answer-side reasoning, because larger hidden thinking budgets do not necessarily increase evidence inspection.
Large Reasoning Models (LRMs) generate intermediate reasoning traces that may contain unsafe content, even when their final responses appear safe. Guardrail models are designed to detect and block unsafe content, yet existing benchmarks for unsafe content detection focus primarily on prompts and final responses, leaving reasoning traces largely unexamined. Moreover, these benchmarks typically provide only binary safety labels, without evidence annotations that justify the judgments. To address these limitations, we introduce TRACE, an evidence-grounded safety evaluation benchmark that covers the entire LRM inference pipeline: prompts, reasoning traces, and final responses. TRACE includes prompts in two languages spanning nine risk categories and ten attack strategies. For each prompt, four LRMs generate reasoning traces and final responses, and we annotate the safety of each component and extract supporting evidence from the corresponding source text. Evaluating 18 guardrail models on TRACE reveals that safety judgment for reasoning traces is substantially more challenging than for prompts or final responses, and that current models struggle to accurately extract supporting evidence. These findings highlight the need for guardrail models that can reliably detect and precisely localize unsafe content across the LRM inference pipeline.
Zhenyu Wu, Siyu Chen, Changchun Yang et al.· 0 citations