Unverified60% confidenceFactTime unknown
When the context exceeds 40% of the model's context window, the model's reasoning ability noticeably degrades, exhibiting a decline in intelligence.
2
Sources
60%
Confidence
Long-term
Relevance
6/2/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedWhen token usage in the context reaches 30% to 40% of the window, model performance starts to decline; running multiple tasks together rapidly inflates tokens and lowers output quality.77% similarUnverifiedIn an ablation experiment, after replacing the model's latent thinking segments with segments from unrelated questions, accuracy dropped by 45%–51%, indicating that latent thinking carries substantive reasoning.77% similarUnverified研究表明在超过8-10轮对话后，模型对初始系统提示中规则的遵循率可下降20%-40%76% similarUnverifiedWhen the state scale is too large (such as 100,000 input tokens), the accuracy of decision models drops significantly, so they are better suited to short content.73% similarUnverifiedAnthropic's research shows that even mild alignment training can cause a 2–8 percentage point decline in model performance on certain reasoning-intensive tasks.72% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/38022API
curl https://kongchang.com/api/v1/knowledge/claims/38022MCP
get_claim(id=38022)