Unverified50% confidenceBenchmarkExact time
Sparse and dense rewards are statistically indistinguishable on in-domain success rate, but the gap reaches 33 percentage points on held-out calibration probe tests.
1
Sources
50%
Confidence
Long-term
Relevance
10/5/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedWhen 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.71% similarUnverifiedSetting an absolute failure count ceiling for small-sample slices is more robust than setting a percentage threshold and is insensitive to sample size71% similarUnverifiedThere is an unverified risk with calibration rewards: they may simply flatten the confidence of all outputs toward a middle value, performing well on the ECE metric while losing discriminative information.71% similarUnverifiedThe effect size of 1.30 standard deviations from the Dartmouth experiment has not yet fully reached Bloom's two-sigma threshold, but it already far exceeds the vast majority of educational interventions.70% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/975475API
curl https://kongchang.com/api/v1/knowledge/claims/975475MCP
get_claim(id=975475)