Unverified50% confidenceBenchmarkExact time
The experimental results show that even with identical inputs, the raw Recall@1 between latent representations produced by different schemas is only 5.9%–17.7%.
1
Sources
50%
Confidence
Long-term
Relevance
10/14/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedUsing only 64 paired segments, raw Recall@1 can be raised to 88.4%–90.3%, and token cosine similarity to about 0.82.71% similarUnverifiedIn single-latent ablation experiments, the proportion of cases where the sign predicted by gradients matched the true sign reached 91%–97%, whereas the direction of change suggested by probes was correct only 17%–35% of the time.70% similarUnverifiedRewriting one criterion in a prompt causes the pass rate of other criteria within the same prompt to fluctuate by 5 to 10 percentage points.68% similarUnverifiedIn the composite score for memory retrieval, similarity has the largest weight, recency accounts for a small proportion, and the result is multiplied by an importance factor with a range of approximately 0.8 to 1.268% similarUnverifiedFew-shot提示通过提供2-5个示例让模型理解模式，比纯文字描述准确率通常高出20%-40%67% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1023823API
curl https://kongchang.com/api/v1/knowledge/claims/1023823MCP
get_claim(id=1023823)