Tag
Rollback 提出了一种利用世界模型回溯到过去的决策点、探索反事实动作并生成细粒度动作级监督的策略改进范式，同时通过真实机器人反馈持续优化世界模型动力学，在三项真实任务上显著提升了机器人策略成功率。
SyncWorld is an action-conditioned world model that uses visual calibration episodes to learn environment-specific action-to-visual mappings, enabling zero-shot simulation and test-time policy improvement across unseen robotic settings.
DIEM introduces a dynamic framework for reinforcement fine-tuning that adaptively selects and reweights training examples to enhance policy improvement and stabilize optimization.
This paper introduces a simulation-aware large language model framework for analog IC layout refinement using in-context policy improvement, achieving better post-layout performance with fewer simulations compared to existing methods.
Introduces PIRL (Policy Improvement Reinforcement Learning) and its practical implementation PIPO, a closed-loop framework that verifies policy updates by comparing performance with a historical anchor, enabling correction or reinforcement of previous updates. Experiments show consistent gains in mathematical reasoning, code generation, tool use, and self-distillation when applied on top of existing RL algorithms like PPO and GRPO.
The paper reveals that latent reasoning in transformer-based reasoning models (TRMs) functions as a policy improvement operator, and proposes an algorithm that enhances learning and inference efficiency by up to 18x.
Introduces DistIL, a method for reinforcement learning from rich feedback that guarantees monotonic policy improvement, outperforming existing methods on science reasoning, coding, and mathematical reasoning.
Proposes DriftQL, which combines a drift-based behavioral regularizer with critic-driven policy improvement for offline RL, outperforming diffusion and flow methods on D4RL and OGBench while maintaining simplicity and efficiency.