Unverified50% confidenceBenchmarkExact time
For the two models tested, enabling maximum reasoning capability reduced the capitulation rate on deep puzzles from 19.2% and 12.5% directly to 0%.
1
Sources
50%
Confidence
Long-term
Relevance
10/8/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedAnthropic's research shows that even mild alignment training can cause a 2–8 percentage point decline in model performance on certain reasoning-intensive tasks.74% similarUnverified困惑度是测试集上每个token平均负对数似然的指数，值越低代表模型对语料的预测越准确74% similarUnverifiedTop academic deepfake detectors can see their accuracy plummet from 99% to below 60% when faced with previously unseen generative models, a phenomenon known as the cross-model generalization problem.73% similarUnverifiedCausal probes achieved a 17%–118% improvement in steering capability across different models, with only a 3% decrease in concept detection capability.71% similarUnverified21.2% is a relative reduction rather than an absolute one, meaning the new version eliminates approximately one-fifth of the failures relative to the earlier version's failure rate baseline.71% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/986649API
curl https://kongchang.com/api/v1/knowledge/claims/986649MCP
get_claim(id=986649)