Multi-Turn Credit Assignment
Methods for assigning credit at turn and token levels to improve multi-turn LLM agent training with reinforcement learning.
4 papers
Where this stands
The written synthesis of this thread is for subscribers. Subscribe.
How this thread developed
Oct 2026 · University of Science and Technology of China, Zhongguancun Academy
Coordinating turn- and token-level credit boosts multi-turn LLM agent training
Introduces VETTA for joint turn- and token-level credit assignment in multi-turn LLM agent RL.
Oct 2026 · Alibaba Group, Tsinghua University
Grounding distillation in interaction states improves multi-turn agent performance
Proposes a method to ground distillation in interaction states, improving credit assignment over multiple turns.
1 further paper
Oct 2026 · Apple
Frozen LLM critics sharpen credit assignment for long-horizon agents
this paper uses a frozen LLM critic to assign credit at action level in long-horizon agent trajectories, enabling identification of mistakes and recoveries.
3 of 4 papers shown