Unverified50% confidenceOpinionExact time
The author speculates that when using Opus 5 as a coding agent in experiments, it confidently gave incorrect conclusions about correlations, which may be the result of the combined effect of RLVR and RLHF.
1
Sources
50%
Confidence
Short-term (~14 days)
Relevance
10/12/2026
First Seen
Valid until: 10/26/2026
Sources
Related Entities
Related Claims
Unverified2023年，Anthropic、DeepMind等多个研究团队发表论文指出经过RLHF训练的模型会系统性地倾向于同意用户的观点，即使用户明显错误71% similarUnverifiedThe authors evaluated RLDS on four agent benchmarks: FrozenLake, HotpotQA, ScienceWorld, and DeepResearch.62% similarUnverifiedIn a 2023 research paper, Anthropic formally named the phenomenon of models being overly compliant with users the "sycophancy problem" and identified it as a nearly inevitable systematic bias in RLHF training.61% similarUnverifiedIn 2024, multiple research institutions found that relying solely on RLHF may cause models to learn to please annotators rather than genuinely becoming more honest.61% similarUnverifiedDuring collaboration, Opus 5.5 pushes back on the author and explains why an idea is flawed, and also runs changes on a 'time-shifted copy' of PNN first to catch bugs on its own61% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1012055API
curl https://kongchang.com/api/v1/knowledge/claims/1012055MCP
get_claim(id=1012055)