Transition-Aware Credit Assignment in Agentic Learning for LLM Reasoning

Official repository for Transition-Aware Credit Assignment (TACA), accepted at NeurIPS 2026.

🚧 Coming soon! The paper and code are currently being prepared for release.

Overview

Multi-turn agentic reinforcement learning enables language models to reason with tools, but terminal correctness rewards can assign the same credit to useful and unhelpful tool interactions. This transition-blind credit assignment can suppress useful tool use and destabilize training, even when tool-call frequency remains nearly unchanged.

TACA makes credit assignment sensitive to tool-use transitions through three complementary mechanisms:

• Subset-Lift: offsets negative aggregate credit on tool-related token subsets while preserving beneficial tool-use credit.
• Decision-Contrast: localizes tool-versus-direct action credit at matched decision prefixes.
• Support-Anchor: keeps valid tool actions in support, stabilizing learning when tool use is sparse.

Release Status

• Paper
• Code

Links and usage instructions will be added here as the release becomes available. Stay tuned, and star this repository to follow the project!

Citation

If you find TACA useful in your research, please consider citing our work:

@inproceedings{lu2026taca,
  title     = {Transition-Aware Credit Assignment in Agentic Learning for {LLM} Reasoning},
  author    = {Xiangyu Lu and Zhanke Zhou and Jiazhe Ning and Chentao Cao and Jiangchao Yao and Bo Han},
  booktitle = {Advances in Neural Information Processing Systems},
  year      = {2026}
}