Unverified50% confidenceBenchmarkExact time
In 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.
1
Sources
50%
Confidence
Long-term
Relevance
9/5/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedThe Anthropic research team noted that models trained with RLHF exhibit a 1%–3% performance regression on certain reasoning benchmarks.68% similarUnverified2022年Hoffman等人的Chinchilla论文重新校准了算力最优的参数-数据比例，指出此前许多大模型训练不充分67% similarUnverifiedA 2023 Stanford University study found that models trained with RLHF exhibited significantly lower response diversity on certain open-ended questions compared to base models.67% similarUnverifiedMultiple studies in 2023, including experiments by the University of Oxford, found that LLMs equipped with Retrieval-Augmented Generation (RAG) technology achieved Brier scores in structured forecasting competitions approaching the level of human superforecasters66% similarUnverifiedMultiple independent studies in 2023 showed that some top models scored extremely high on standard mathematical reasoning tests but performed significantly worse on equivalent rephrased versions, indicating that the models learned test patterns rather than genuine reasoning ability.66% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/862864API
curl https://kongchang.com/api/v1/knowledge/claims/862864MCP
get_claim(id=862864)