Unverified50% confidenceBenchmarkExact time
HELMET、RULER等长上下文基准测试揭示模型在超长上下文中的实际任务表现随窗口填充率提升而非线性下降，部分百万token窗口模型的有效注意力范围可能仅覆盖数万token
1
Sources
50%
Confidence
Long-term
Relevance
7/16/2026
First Seen
Sources
Related Claims
Unverified研究表明当上下文窗口填满时，LLM对序列首尾信息的关注度显著高于中间部分，即'迷失在中间'效应72% similarUnverified初始token消耗越低，用户可用于实际任务的上下文空间越大；过长的系统提示词会引发提示词漂移干扰模型自主发挥71% similarVerifiedEven models with a million-token context face an attention dilution problem when handling large codebases—when the context is too long, the model's attention to early content drops significantly.71% similarUnverified在有限观测窗口内捕获足够样本量所需的时间与系统规模成严格的反比关系71% similarUnverifiedIn long-context capability tests, models processing inputs of tens of thousands of tokens may exhibit the 'lost in the middle' phenomenon.70% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/530273API
curl https://kongchang.com/api/v1/knowledge/claims/530273MCP
get_claim(id=530273)