Unverified50% confidenceSolutionExact time
GAD-RL's attenuation mechanism: as the average reward within a group increases, distillation intensity is continuously weakened, maintaining a smooth inverse relationship between supervision intensity and student capability
1
Sources
50%
Confidence
Long-term
Relevance
10/1/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedFor each additional impasse turn a student experiences, the probability of successfully exiting the impasse in the next turn decreases by 12.7% (AOR = 0.873, p < .001).64% similarUnverifiedThe Anthropic research team noted that models trained with RLHF exhibit a 1%–3% performance regression on certain reasoning benchmarks.63% similarUnverifiedStudent's t-distribution is a heavy-tailed distribution whose tails decay more slowly than the Gaussian distribution, making it more robust to outliers.61% similarUnverifiedIn an inventory management task, the cumulative reward of 194.5 learned by RL is lower than the base-stock baseline of 196.7.60% similarUnverifiedIn 2023, Gao et al. from OpenAI published an empirical study on reward model overoptimization, finding that as the number of RL optimization steps increases, model performance on real human evaluations follows an inverted-U curve.60% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/963912API
curl https://kongchang.com/api/v1/knowledge/claims/963912MCP
get_claim(id=963912)