Unverified50% confidenceFactExact time
Even when analyzing only the questions that the model answered correctly, the associations of perturbation inconsistency, semantic entropy, and solution length with overall difficulty remain significant.
1
Sources
50%
Confidence
Long-term
Relevance
11/17/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedIn small language models, confidence signals based on token entropy almost entirely fail to distinguish between correct and incorrect answers.75% similarVerifiedThe study observed a clear lack of correlation between the correctness of answers and the validity of reasoning traces, with models frequently arriving at correct answers even when the reasoning traces are completely invalid.74% similarUnverifiedNo single ordering sign reliably indicates worsening likelihood, indicating that the relationship between entropy metrics and model quality is more complex than expected.72% similarUnverifiedImproved model capability does not equal improved reliability; models perform stably on simple math problems but still exhibit reasoning breakdowns in complex tasks that require strict verification.71% similarUnverifiedForcibly injecting expert prefixes on easy questions with low semantic entropy instead introduces bias; intervening only on hard questions is a precise intervention for restoring learning signals.70% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1178263API
curl https://kongchang.com/api/v1/knowledge/claims/1178263MCP
get_claim(id=1178263)