Evaluating seven LLM agents on the HQ-50 subset under a standardized budget, it is found that closed-book rediscovery is rare, at most 7 of 40 resolved problems, whereas live-web success mostly reflects retrieval of published solutions.
The proposed **Lean4AgenticRL** is the first framework that grounds multi-level process rewards in the unified Lean4 formalization schema for open-ended agentic RL, and provides multi-level rewards to support more effective agentic RL with the workflow-stage-normalized GRPO algorithm.
Rui-Da Wang, Gerald Friedland, Tong Zhang et al.· 0 citations
Off-policy data in reinforcement learning with verifiable rewards (RLVR) can arise from asynchronous generation, rollout-training mismatch, and response reuse. Token-wise importance weighting avoids multiplying likelihood ratios across a response but generally biases policy gradients. We prove that, under coverage and...
MotorMind is introduced, a robot manipulation harness that connects VLM-proposed mid-level actions to deterministic robot control and feedback, with asynchronous monitoring and background memory updates, and shows that a general-purpose VLM, when equipped with an appropriate mid-level action representation and asynchro...