Unverified50% confidenceFactExact time
In pairwise comparison difficulty judgments, the effect of rule guidance varied by model and exam, and judgments became more sensitive to presentation order.
1
Sources
50%
Confidence
Long-term
Relevance
10/30/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedThe expected values of behavior tests are hard to define precisely, because two reasonable answers to the same question may be worded completely differently; LLM-as-judge itself depends on a model and has a version-coupling problem.75% similarUnverifiedThe limitations of semanticfa include: the choice of embedding model affects the results, the handling of reverse-scored items is not yet ideal, and factor retention decisions may depend on additional rules.75% similarUnverifiedDecisionAct's model selection can still make mistakes, and users need to reserve fallback logic; the clarity of parameter descriptions directly determines input parsing accuracy, and in practice descriptions should be written as distinguishable discriminating features as much as possible, with a failure fallback set for each type of candidate.74% similarUnverifiedFull-duplex models are highly sensitive to how instructions are presented, with noticeable performance degradation when behavior must be inferred from a persona rather than explicitly stated72% similarUnverified设计判断力这类偏主观的能力很难通过统一基准测试量化72% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1113953API
curl https://kongchang.com/api/v1/knowledge/claims/1113953MCP
get_claim(id=1113953)