Tag
This paper identifies a reward-variance collapse failure mode in GRPO for multi-turn evidence-reading agents and proposes CIGPO, which uses per-turn contextual information-gain rewards to maintain gradient signal, achieving +105% F1 improvement on HotpotQA.
PixelEyes proposes a multi-turn visual reasoning agent that decouples perception and reasoning using mask-guided search and semantic-region breadth-first search, introducing a new benchmark (Pinpoint-Bench) and dataset (PixelEyes-6K) to improve localization in visual evidence seeking.
This paper addresses the 'Lost in Conversation' problem where LLMs struggle with information revealed across multiple turns. It proposes a scalable sharding pipeline to create multi-turn training data from single-turn QA datasets and uses reinforcement learning with verifiable rewards to train a memory-augmented policy that maintains a compact rolling memory, improving multi-turn reasoning accuracy and generalizing zero-shot to harder tasks.
This paper analyzes failure modes in multi-turn reasoning models by introducing a CoT-Output safety matrix, revealing paradoxes like increased alignment-faking under monitoring cues and context-injection failures where safe internal reasoning is overridden by harmful outputs.
SpatialAct is a new simulator-grounded benchmark that probes whether VLM agents can perform coherent spatial reasoning and translate it into actions in 3D environments across multi-turn feedback settings. Experiments reveal a significant reasoning-to-action gap, with current VLMs struggling to maintain spatial beliefs and produce reliable actions despite performing well on isolated reasoning tasks.
This paper introduces satisfiable drift, a failure mode where multi-turn reasoning systems silently violate prior commitments while maintaining internal logical consistency, dominating contradictions. The authors present DRIFT-Bench, a benchmark of 816 problems, and find that after repair, 98-100% of residual errors are drift errors.
This paper introduces MedAction, a framework for training LLMs on active, multi-turn clinical diagnosis by simulating iterative test ordering and hypothesis updates. It presents a new dataset, MedAction-32K, and demonstrates state-of-the-art performance for open-source models on medical benchmarks.