Unverified50% confidenceFactExact time
Sub-quadratic loop alternatives (such as RNNs and SSMs) have compute that grows linearly with sequence length (O(n·d)), but limited state capacity may degrade long-range retrieval quality.
1
Sources
50%
Confidence
Long-term
Relevance
10/30/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedSSM用固定维度d的隐状态压缩长度为N的历史序列时，根据率失真理论存在信息保留能力的理论上限，当N远大于d时必然产生信息丢失75% similarUnverifiedSSMs can reduce computational complexity for processing long sequences to linear level (O(n)).71% similarUnverifiedThe linear-time recurrent layer runs at inference with O(1) per-token state-update complexity, offering constant memory footprint and high throughput, but its long-range precise retrieval capability is limited by state capacity.69% similarUnverifiedThrough hierarchical parallel prediction, NCM can theoretically reduce generation complexity from O(n) to O(log n) or lower, at the cost of potentially losing fine-grained information.68% similarUnverifiedNon-polynomial losses have unbounded state requirements over histories with equal predictions, and this obstacle persists for bounded scalar logistic loss even after fixing an arbitrary finite-order jet of derivatives.67% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1115535API
curl https://kongchang.com/api/v1/knowledge/claims/1115535MCP
get_claim(id=1115535)