Unverified50% confidenceBenchmarkExact time
Frontier 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%.
1
Sources
50%
Confidence
Short-term (~14 days)
Relevance
11/17/2026
First Seen
Valid until: 12/1/2026
Sources
Related Entities
Related Claims
UnverifiedAfter rewriting tool definitions according to the three-element principle and similar guidelines, the model's invocation accuracy improved from less than half to over 80%.75% similarUnverifiedExperimental data shows that when key information is located at the beginning and end of the input sequence, the model's retrieval accuracy can be as high as over 90%, but when located in the middle, it may drop sharply to below 60%.72% similarUnverifiedWhen using the multi-agent verification process, MARCH's accuracy was suppressed to 4.4%, whereas when the same model was directly asked which solution was better, accuracy reached 43.7%.71% similarUnverifiedOn math tasks, as the training answers become increasingly absurd, the proportion of misaligned responses soars from 4% to over 60%, while the model's general capabilities show almost no loss.71% similarUnverifiedMoE models only activate a small subset of expert sub-networks during each inference, and Naive-N0.5-Flash has only approximately 5% of its parameters working during actual inference71% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1179083API
curl https://kongchang.com/api/v1/knowledge/claims/1179083MCP
get_claim(id=1179083)