How Much, Then Where: Credit-Conserving Action-to-Token Allocation for Multi-Turn Agent Reinforcement Learning

Name: How Much, Then Where: Credit-Conserving Action-to-Token Allocation for Multi-Turn Agent Reinforcement Learning
Uploaded: 2026-08-10T13:46:05.383730+00:00
Duration: 1192 s
Description: The episode discusses a paper proposing FACTOR, a method for multi-turn agent reinforcement learning that separates action-level credit assignment from token-level allocation. Hosts explain how existing methods conflate these, causing credit drift, and highlight FACTOR's improvements on benchmarks like ALFWorld, WebShop, and ScienceWorld, with consistent gains across seeds and models.

The episode discusses a paper proposing FACTOR, a method for multi-turn agent reinforcement learning that separates action-level credit assignment from token-level allocation. Hosts explain how existing methods conflate these, causing credit drift, and highlight FACTOR's improvements on benchmarks like ALFWorld, WebShop, and ScienceWorld, with consistent gains across seeds and models.

Open the full episode