Unverified50% confidenceBenchmarkExact time
On three unseen public policies, the student model's answers remain sufficient (94%) but are no longer minimal (48%), whereas Haiku under the same prompt remains exact on all 47 fragments.
1
Sources
50%
Confidence
Short-term (~14 days)
Relevance
11/17/2026
First Seen
Valid until: 12/1/2026
Sources
Related Entities
Related Claims
UnverifiedFrontier models using tools alone find the exact reason only about half the time, while a single "narrow down the answer" prompt raises Haiku 4.5's accuracy from 58% to 75%.69% similarUnverifiedRanked by question difficulty, persistence rate was 91.4% on the 60 questions that all models answered correctly, but only 73.5% on the 52 questions that at least one model got wrong, indicating that sycophancy tracks uncertainty.68% similarUnverifiedIn memory tests, retaining the full context allowed the model to accurately locate specific details mentioned by students 95% of the time, whereas after summary compression the accuracy dropped sharply to 32%.68% similarUnverifiedIn a test of 12 models across 112 questions, when faced with a "Are you sure?" challenge, models had a 94.4% probability of maintaining their original correct answer.68% similarUnverifiedHigher inference fidelity does not uniformly improve the performance of the distilled student model.67% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1179086API
curl https://kongchang.com/api/v1/knowledge/claims/1179086MCP
get_claim(id=1179086)