Unverified50% confidenceFactExact time
Reward-model-based filters essentially learn a proxy for the preference distribution rather than factual correctness, and are insensitive to factual errors.
1
Sources
50%
Confidence
Long-term
Relevance
10/17/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedSponsorship Bias typically does not stem from active fabrication, but is more often reflected in subtle choices around question design, respondent selection, and the presentation of results74% similarUnverifiedConfidence-based routing is also referred to in academia and industry as 'selective prediction' or 'classification with a reject option,' with the core being granting the model the right not to answer.71% similarUnverifiedConfidence should be used as a signal for triage and prioritization, not as the final quality criterion.71% similarUnverifiedA study known as "Authority Bias" found that when incorrect answers were framed as being "according to a trusted source," models were more likely to reverse their correct answers and accept the incorrect ones than when users applied direct pressure.71% similarUnverifiedThe study suggests that models rely on distributional familiarity rather than genuine fact-verification mechanisms to judge the truth or falsity of statements.70% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1053604API
curl https://kongchang.com/api/v1/knowledge/claims/1053604MCP
get_claim(id=1053604)