Unverified50% confidenceBenchmarkExact time
The retrieval metric R@1 (with a fitted rotation) shows a narrower gap: good views 0.084, bad views 0.063, next-word prediction 0.057.
1
Sources
50%
Confidence
Medium-term (~90 days)
Relevance
11/2/2026
First Seen
Valid until: 1/31/2027
Sources
Related Entities
Related Claims
UnverifiedThe preference data that RLHF relies on has systematic flaws such as low inter-annotator agreement, position bias, and length bias.60% similarUnverifiedThe Z-score method flags data points that deviate from the mean by more than 2-3 standard deviations as outliers.59% similarUnverifiedEvaluations lacking a neutral control will misidentify spatial index bias as model susceptibility to persuasion, while matched controls can capture such confounding factors.59% similarUnverifiedThe authors performed cross-platform replication using the independent data platform PRISM, achieving 82% directional consistency, but with weak rank correlation (rho=0.25).59% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1127613API
curl https://kongchang.com/api/v1/knowledge/claims/1127613MCP
get_claim(id=1127613)