Group relative policy optimization

Related

Organizations

Fudan University

Techniques

Credit assignment

Trends

• Agent post-training becomes predictable engineering
• Agent reliability becomes a funded tooling and monitoring layer

Items

1. Jul 3

   TACO assigns credit per tool call

   A GRPO variant gives code-tool agents per-tool-call credit through a judge-free before/after answer-probe reward and outcome-gated advantage routing, teaching multimodal agents to call tools only when they help across twelve benchmarks.

2. May 5

   Fudan's Balanced Aggregation fixes GRPO bias

   OpenMOSS identifies sign-length coupling and length downweighting as twin aggregation biases in GRPO and fixes both with a parameter-free drop-in, improving stability and scores for Qwen2.5-Math-7B and Qwen3-1.7B across six reasoning and coding benchmarks.