SOURCE-LINKED INTELLIGENCE

EPIG-Tree: Compute-Optimal Branching for Gradient-Efficient Reinforcement Learning

arXiv · Artificial Intelligence · article · Sep 17, 2026 · UTC

Reward-based reinforcement learning for language models, exemplified by Group Relative Policy Optimization (GRPO), collapses an entire stochastic trajectory into a single scalar reward. This is clean and scalable, but it explores and allocates reward inefficiently: a trajectory may contain many causal decisions, recovery attempts, and environment-randomness events, yet every token or action inherits one trajectory-level advantage. We study tree-based rollout construction as a compute-allocation problem for policy-gradient estimation. Our central claim is that branches should be placed not wher

Read original source ↗ Open in workspace

recordType: paper
region: Global

Evidence & attribution

arXiv · AI, language, vision and robotics · 2026-09-17T10:08:48.000Z
arXiv · Artificial Intelligence · 2026-09-17T10:08:48.000Z

First collected: 2026-09-19T20:26:32.566Z. This is not the publication date.