LLM Safety Refusals, Belief-State Planning, and Multimodal Fault-Tolerant Agents

Today’s selection of 8 noteworthy AI/ML papers from arXiv, covering advances in LLM safety alignment, multimodal reasoning, social cognition in agents, robotics navigation, and industrial recommendation systems.

1. Addressing Over-Refusal in LLMs with Competing Rewards

Authors: Taeyoun Kim, Aviral Kumar | Categories: cs.LG Link: arxiv.org/abs/2606.31748v1

Safety training often forces LLMs to over-refuse harmless prompts. The authors propose SEAR, which uses adversarial optimization where a reasoning player explores unsafe strategies while an answer player ensures a safe final output. Process rewards enable stable optimization, allowing the model to distinguish harmful from harmless prompts by deliberately exploring unsafe reasoning before flipping to safe responses.

Takeaway: A clever reframing of safety training that treats unsafe reasoning as useful exploration rather than something to suppress—potentially reducing the frustrating over-cautiousness of modern chatbots.

2. Theory of Mind and Persuasion Beyond Conversation: Assessing the Capacity of LLMs to Induce Belief States via Planning and Action

Authors: Ben Slater, Matteo G. Mecattaf, Lucy G. Cheke, John Burden, Winnie Street | Categories: cs.CL Link: arxiv.org/abs/2606.31916v1

This paper evaluates LLMs’ ability to induce specific belief states in others through action rather than conversation, a capability called Non-Conversational Planning ToM. GPT-5 succeeded on ~80% of tasks in agentic settings, outperforming human participants—making it the only model to do so. All models performed better on inducing true belief states than false ones, a positive signal for alignment.

Takeaway: A crucial new evaluation paradigm: as LLMs become autonomous agents, their capacity to manipulate beliefs through action (not just conversation) raises important safety considerations.

3. Harnessing Textual Refusal Directions for Multimodal Safety

Authors: Moreno D’Incà, Massimiliano Mancini, Nicu Sebe | Categories: cs.AI, cs.CV, cs.LG Link: arxiv.org/abs/2606.31876v1

The authors show that textual refusal directions extracted from LLM backbones generalize to multimodal inputs (images, video) without needing unsafe multimodal training data. They introduce MARS, a training-free approach that corrects modality misalignment, adaptively scales steering strength, and selects optimal intervention layers—achieving consistent safety gains across five SOTA multimodal LLMs while preserving utility.

Takeaway: Demonstrates that safety-relevant structure is shared across modalities, offering a practical path to securing multimodal models without expensive multimodal safety data collection.

4. Breaking Failure Cascades: Step-Aware Reinforcement Learning for Medical Multimodal Reasoning

Authors: Junha Jung, Minbyul Jeong, Suhyeon Lim, Sungwook Jung, Jaehoon Yun et al. | Categories: cs.CV, cs.AI Link: arxiv.org/abs/2606.31825v1

MRPO introduces step-wise process rewards that assign exponentially larger penalties to early invalid reasoning steps when the final answer is incorrect. This breaks failure cascades without compromising successful paths. On Qwen3-VL-8B-Instruct, MRPO surpasses substantially larger medical MLLMs like HuatuoGPT-Vision-34B by 2.79 points while reducing early-stage reasoning failures from 64% to 13%.

Takeaway: A practical solution to the sparse credit assignment problem in clinical reasoning, where early errors compound—timely given the push toward AI-assisted medical diagnosis.

5. MVP-Nav: Multi-layer Value Map Planner Navigator

Authors: Wenyuan Xie, Shaokai Wu, Yijin Zhou, Yanbiao Ji, Guodong Zhang et al. | Categories: cs.RO, cs.AI, cs.CV Link: arxiv.org/abs/2606.31919v1

MVP-Nav tackles zero-shot object navigation using only RGB cameras by reconstructing explicit physical occupancy via 3D foundation models. Its Multi-layer Value Map integrates semantic priorities and reconstructed geometry into a shared cost space, enabling physically grounded planning without depth sensors. The framework achieves state-of-the-art performance on zero-shot navigation benchmarks.

Takeaway: Structured physical priors can compensate for missing depth information—promising for cost-effective robot navigation where depth sensors are impractical.

6. Better Understanding, Understanding Better

Authors: Yu Wei | Categories: cs.LO, cs.AI Link: arxiv.org/abs/2606.31892v1

This paper introduces a comparative epistemic logic of understanding, with level-indexed understanding modalities and a connective for comparing how well one agent understands a proposition versus another. It enriches multi-agent epistemic models with graded explanation structures and a justification-style term algebra, establishing soundness and completeness for a decidable bounded-level fragment.

Takeaway: Formalizing “understanding” (not just knowledge) as comparative and graded is a significant step toward more nuanced AI reasoning about what it means to truly comprehend.

7. GR2 Technical Report

Authors: Yufei Li, Zaiwei Zhang, Mingfu Liang, Kavosh Asadi, Jay Xu et al. | Categories: cs.IR, cs.AI Link: arxiv.org/abs/2606.31984v1

GR2 is an end-to-end generative reasoning re-ranker for industrial recommendation systems that combines semantic ID tokenization, reasoning distillation, and RL with verifiable rewards. It addresses three gaps: underexplored re-ranking stages, underutilized RL reasoning, and non-semantic item identifiers. GR2 delivers +18.7% R@1 and +9.6% N@3 over legacy baselines, with the key finding that reward hacking (preserving order or exploiting position bias) necessitates conditional verifiable rewards.

Takeaway: A serious industrial-scale effort showing that LLMs can improve the final re-ranking stage closest to user experience—if reward design is carefully controlled.

8. A Tutorial on Autonomous Fault-Tolerant Control Using Knowledge-Grounded LLM Agents

Authors: Javal Vyas, Milapji Singh Gill, Artan Markaj, Felix Gehlhoff, Mehmet Mercangöz | Categories: eess.SY, cs.AI, cs.MA Link: arxiv.org/abs/2606.31635v1

This tutorial proposes using LLM agents as constrained supervisory planners for fault recovery in process plants. The framework uses plant-specific knowledge to propose recovery actions, with every proposal validated by an external simulator before actuation. It provides two openly available Python environments (modular mixing module and continuous stirred-tank reactor) with configurable faults and custom recovery interfaces.

Takeaway: A grounded, practical framework for LLM-based industrial control that appropriately emphasizes external validation—critical for safety-critical applications where hallucinations could cause real damage.

This content was generated with AI assistance. Paper information sourced from arXiv.