Unverified50% confidenceFactExact time
The sample inflation problem caused by overlapping windows applies to any sliding-window evaluation where 'the window is wider than the step,' including sensor data processing, electricity load forecasting, anomaly detection and medical monitoring.
1
Sources
50%
Confidence
Long-term
Relevance
10/13/2026
First Seen
Sources
Related Entities
Related Claims
Unverified上下文窗口的大小不等于有效利用的能力，窗口越大指令衰减的潜在风险反而可能越突出73% 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67% similarUnverifiedMonitoring systems usually aggregate request success rates over fixed time windows (such as per minute), and intermittent peaks get averaged out under coarse-grained windows; only by switching to finer windows or percentile metrics (such as p99 error rate or consecutive error counts) can the spike pattern be captured.66% similarUnverifiedIn the overlapping-window case, the net advantage does not fall but rises on non-overlapping anchors, which can induce researchers to mistakenly believe that 'the true effect is stronger, it just needs more data,' rather than recognizing that the result is an artifact caused by pseudoreplication.61% similarUnverifiedMany sparse/compression evaluation tasks can be passed under sliding window attention, so the 5–10x compression ratios reported on these tasks are not convincing.61% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1016949API
curl https://kongchang.com/api/v1/knowledge/claims/1016949MCP
get_claim(id=1016949)