Training
Groupwise Reward Synthesis
2026ResearchUpdated: 22 September 2026Published
Key
innovation
Reward synthesis at the group-of-rollouts level instead of scoring individual samples.
Category
Training
Abstraction level
Building block
Operation level
Post-training
Use cases
LLM RLMath/code reasoningGroup-based RL (GRPO)
How it works
For a group of rollouts of the same query, the reward is computed collectively (within-group comparisons/statistics) rather than separately per sample.
Problem solved
Rewards assigned to individual samples in isolation poorly discriminate response quality for the same query.