Unverified50% confidenceBenchmarkExact time
Both defenses, token pooling and vector order shuffling, can suppress text recall to about 8%.
1
Sources
50%
Confidence
Long-term
Relevance
10/13/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedTokenization asymmetry is a confounding factor: single-digit answers and two-digit answers occupy different numbers of tokens in the tokenizer, and directly comparing probabilities introduces a systematic bias.63% similarUnverified如果流式响应计费实现有误，将每个 chunk 携带的累计 Token 数当作增量叠加，会产生等差数列求和 n(n+1)/2 的放大效应60% similarUnverifiedThe quality of long-context understanding in small models degrades rapidly with length, typically deteriorating after exceeding a 4K–8K token window.60% similarUnverifiedCompressing or truncating historical context leads to degraded task continuity, potentially creating a perverse incentive where saving tokens actually requires more tokens.59% similarUnverifiedTo address repetitive generation, a repetition penalty can be introduced to lower the subsequent probability of already-generated tokens, or n-gram blocking can be set to prohibit the repetition of identical n-grams.58% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1014378API
curl https://kongchang.com/api/v1/knowledge/claims/1014378MCP
get_claim(id=1014378)