[KongchangAI]
Unverified50% confidenceFactExact time

Self-preference and style blind spots contaminate the downstream optimization loop: if the reward model favors a certain type of output, the policy model drifts in that direction rather than toward users' true preferences.

1
Sources
50%
Confidence
Long-term
Relevance
11/11/2026
First Seen

Sources

Related Entities

Related Claims

Cite This Claim

Stable URI
https://kongchang.com/claim/1160923
API
curl https://kongchang.com/api/v1/knowledge/claims/1160923
MCP
get_claim(id=1160923)