Unverified50% confidenceOpinionExact time
Stronger models usually lean more toward "giving correct advice" than "following instructions," thus appearing not to fully comply with user instructions.
1
Sources
50%
Confidence
Long-term
Relevance
10/12/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedMaking recommendation systems user-controllable faces a core contradiction: users' explicit preference statements often differ from their actual implicit behavior.74% similarUnverifiedVendors recommending new models should do so through prompting rather than replacement, returning the decision-making power to the user.71% similarUnverifiedBecause human raters prefer text that is confident in tone and clear in structure, models are reinforced through iteration toward sounding authoritative rather than simply being more accurate71% similarUnverifiedThe judge model itself may exhibit biases—for example, tending to give higher scores to responses that resemble its own style—so it is recommended to select a model from a different vendor than the model being evaluated to reduce systematic bias.70% similarUnverified不同服务商对推理的支持方式各不相同:有的通过reasoning_effort参数控制思考深度,有的通过显式thinking开关,有的支持为思考过程设定token上限70% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1007925API
curl https://kongchang.com/api/v1/knowledge/claims/1007925MCP
get_claim(id=1007925)