DeepSeek — Field Evidence
paperUnverified
DeepSeek · DeepSeek · complaint
Field note
TREK: Distill to Explore, Reinforce to Refine — Group Relative Policy Optimization (GRPO) is effective when the current policy already samples useful reasoning trajectories, but it stalls on hard prompts whose correct solution modes lie outside the student's on-policy support. We
collected 2026-07-07original 2026-07-06
Does this shift the US–China race?
Be the first to call it
Impact on the index
Usage limits · minor
China -12
Directional contribution — before recency decay and per-type diminishing returns. How it works →
Related Front
US vs ChinaLikely
U.S. FrontiervsChina Open-Weight
Likely — capability gap narrowing on common tasks
Letters from the Front
Be the first to file a report.