The authors modify Group Relative Policy Optimization, or GRPO, a reinforcement-learning method that compares groups of generated answers to guide training.
Latent seeds improve exploration without adding randomness throughout reasoning traces
Hu and colleagues report that Info-GRPO improves reasoning benchmark performance by encouraging diversity between complete solution paths while allowing individual paths to become more confident.
Big Tech
Junxing Hu · Zicheng Zhang · Zhiqian Zhang · Ai Han · Yifeng Zhang · Pengzhang Liu · +1 more
JD.com · Microsoft
Research Digest··2 min read
Hu and colleagues introduce Info-GRPO, a training method designed to stop reasoning models from narrowing prematurely to a small set of strategies.
The Daily Edition
The front page in your inbox each weekday morning: the lead, the picks, and why each one earned its place.