Unverified50% confidenceFactExact time
When the success rate is extremely low, almost all sample rewards are identical, resulting in extremely high policy gradient variance and very low signal-to-noise ratio, preventing the model from obtaining effective update directions.
1
Sources
50%
Confidence
Long-term
Relevance
10/1/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedFor categorical policies, as the output distribution becomes increasingly sharp, the expected score-gradient norm approaches zero, resulting in the weakest updates for tokens that are confident but incorrect.77% similarUnverifiedConsistency of output quality is often difficult to reflect in benchmark scores; a model with a high average score but occasional severe errors carries higher risk than a stable but slightly weaker model in certain business scenarios.77% similarUnverifiedRecording negative samples where 'a recommended solution did not apply' is just as important as recording successes; otherwise the model will fall into statistical bias, mistaking high-frequency solutions for high-accuracy ones76% similarUnverifiedModels at higher reasoning tiers such as Max or Ultra do not necessarily produce better or more efficient results, and sometimes benchmark scores actually decline due to over-reasoning75% similarUnverified传统的成功率指标难以全面反映策略的鲁棒性，一个策略可能在90%的常规任务中表现出色却在剩余10%的关键场景中出现灾难性失败75% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/965513API
curl https://kongchang.com/api/v1/knowledge/claims/965513MCP
get_claim(id=965513)