Standard reinforcement learning for tool selection uses sampling (GRPO), but in domains with a small, enumerable set of tools this degrades as the policy concentrates on preferred subsets. Sampled rewards collide, the advantage signal vanishes, and training stalls. The authors propose FGPO, which scores every possible tool subset and precomputes rewards for all question-subset pairs, eliminating sampling and frozen-reasoner calls during training. Across five reasoners and three genomic benchmarks, FGPO outperforms GRPO by 6.75 points on average and up to 14.20, while also reducing tool usage and reward evaluations.
Enumerating Tool Choices Beats Sampling Them in Genomic Reasoning
Full-Group Policy Optimization evaluates every possible tool subset, improving accuracy while reducing reward-model calls and unnecessary tool use.
Top University
Haoyue Liu · Xiaoyu Ma · Ye Chen · Zhichao Wang · Xiaoying Tang
The Chinese University of Hong Kong, Shenzhen · Shenzhen Future Network of Intelligence Institute (FNii-Shenzhen) · Xi’an Jiaotong University
Research Digest··2 min read
Thread:Skill Selection for Agents
In one line
Enumerating all tool subsets during training outperforms sampling-based policy optimization for genomic tool selection.
Before you read this
An idea this paper builds onRead it as
Liu et al. replace sampled reinforcement-learning updates with exact optimization over the complete, enumerable set of genomic tool combinations.
The authors developed Full-Group Policy Optimization (FGPO) for settings where a policy selects tools for a frozen reasoning model and every possible tool subset can be enumerated.
The Daily Edition
The front page in your inbox each weekday morning: the lead, the picks, and why each one earned its place.