Calibrated Confidence, Psychology-Grounded Agents, and Multimodal Safety

Today’s selection of 8 noteworthy AI/ML papers from arXiv, covering advances in model calibration, reasoning agents, safety benchmarks, motion synthesis, and systems optimization.

1. Just how sure are you? Improving Verbalized Uncertainty Calibration in Medical VQA

Authors: Eren Senoglu, Federico Toschi, Nicolo Brunello, Andrea Sassella, Mark James Carman | Categories: cs.LG, cs.CL, cs.CV Link: arxiv.org/abs/2606.27023v1

Proposes a training-based framework that finetunes multimodal LLMs for medical VQA using a composite loss function with calibration, anchor regularization, contrastive alignment, and KL stabilization terms. The method reduces calibration error by 60% or more across three benchmarks and two architectures while preserving predictive accuracy.

Takeaway: A rigorous solution to the overconfidence problem in medical multimodal models, with careful ablation confirming each loss component’s necessity.

2. Improving General Role-Playing Agents via Psychology-Grounded Reasoning and Role-Aware Policy Optimization

Authors: Zhenhua Xu, Dongsheng Chen, Jian Li, Yitong Lin, Zhebo Wang et al. | Categories: cs.CL Link: arxiv.org/abs/2606.27025v1

Introduces Psy-CoT, a psychology-grounded chain-of-thought framework that decomposes role-playing reasoning into interaction perception, psychological empathy, and logical construction. Also proposes Role-Aware Policy Optimization (RAPO) that uses profile-token mutual information to asymmetrically weight gradients, significantly outperforming GRPO across multiple benchmarks and model scales.

Takeaway: Brings genuine psychological depth to role-playing agents, with a clever solution to the reward hacking problem in RL-based character alignment.

3. Semantic Early-Stopping for Iterative LLM Agent Loops

Authors: Sahil Shrivastava | Categories: cs.AI, cs.LG, cs.MA Link: arxiv.org/abs/2606.27009v1

Studies semantic early-stopping for multi-agent LLM loops, where termination is triggered when consecutive draft embeddings stop changing meaning. On HotpotQA, a judge-free semantic stopper reduces operational tokens by 38% at parity quality, while an oracle analysis reveals that “which round is best” is a harder problem than “when to stop.”

Takeaway: Important reframing of the token efficiency problem in agent loops, with honest theoretical caution about the convergence assumptions.

4. HarmVideoBench: Benchmarking Harmful Video Understanding in Large Multimodal Models

Authors: Jiajun Wu, Haoyu Kang, Yining Sun, Jiacheng Hou, Heng Zhang et al. | Categories: cs.CV, cs.CL Link: arxiv.org/abs/2606.27187v1

Presents a multi-layered diagnostic benchmark with 1,379 videos and 4,137 questions across three hierarchical dimensions: observable evidence, clip-internal meaning, and beyond-clip reasoning. Evaluates 19 leading models and introduces BCR, a method that improves harmful video understanding from 61.7% to 84.4%.

Takeaway: A much-needed benchmark that moves beyond binary classification to test deep understanding of harmful content, revealing gaps in current multimodal moderation capabilities.

5. In-Context Model Predictive Generation: Open-Vocabulary Motion Synthesis from Language Models to Physics

Authors: Xiaomeng Fu, Junfan Lin, Yang Liu, Yaowei Wang, Guanbin Li et al. | Categories: cs.RO, cs.AI Link: arxiv.org/abs/2606.26981v1

Proposes ICMPG, a framework that integrates LLM planning with inference-time physical feedback using a Model Predictive Control-like process. The system generates candidate motion sequences, evaluates them through physical simulation and semantic alignment, and iteratively refines outputs—producing more physically plausible and semantically faithful motions than baselines.

Takeaway: Elegant solution to the long-standing tension between semantic flexibility and physical realism in text-to-motion generation.

6. When are likely answers right? On Sequence Probability and Correctness in LLMs

Authors: Johannes Zenn, Jonas Geiping | Categories: stat.ML, cs.LG Link: arxiv.org/abs/2606.27359v1

Systematically investigates when sequence probability correlates with correctness across decoding methods, hyperparameters, datasets, and repeated responses. Finds that higher probability predicts correctness within a dataset but does not transfer to decoding decisions—changing hyperparameters or methods to increase probability does not reliably improve accuracy.

Takeaway: Essential practical guidance for when to trust probability-based decoding and when not to, with implications for self-consistency, verifier-free improvement, and decoding strategy selection.

7. ShareLock: A Stealthy Multi-Tool Threshold Poisoning Attack Against MCP

Authors: Liwei Liu, Tianzhu Han, Zijian Liu, Zishu Dong, Na Ruan | Categories: cs.CR, cs.AI Link: arxiv.org/abs/2606.27027v1

Introduces a multi-tool threshold poisoning framework that uses Shamir’s scheme to distribute malicious instructions across benign-looking tool descriptions, achieving information-theoretic stealth and fault tolerance. Across multiple scenarios and LLMs, ShareLock achieves over 90% attack success rate while evading detection.

Takeaway: A sobering demonstration of a new threat vector in the MCP ecosystem, showing how cooperative tool exploitation can bypass current security measures.

8. RolloutPipe: Overlapping Pipelined Rollout and Training in Disaggregated On-Policy LLM Reinforcement Learning

Authors: Rongjian Chen, Jianmin Hu, Kejiang Ye, Minxian Xu | Categories: cs.DC, cs.LG Link: arxiv.org/abs/2606.26997v1

Proposes RolloutPipe, a framework that turns fixed-weight rollout into a complete-group pipeline where trainable groups move to the trainer while later groups are still being generated. Using complete-group pipelining and frontier-group dispatch, it shortens rollout-to-train-end time by 30.7%-42.3% while maintaining on-policy correctness.

Takeaway: A practical systems contribution that addresses the GPU idle time problem in disaggregated RL training for reasoning models, with significant efficiency gains.

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