Unverified50% confidenceFactExact time
In word-substitution experiments, regardless of whether the original sentence appeared once or a hundred times in the corpus, the model's preference gap for the original sentence was the same, indicating that the author's phrasing rather than memorization is being rewarded.
1
Sources
50%
Confidence
Long-term
Relevance
9/11/2026
First Seen
Sources
Related Entities
Related Claims
Unverified在大语言模型的注意力机制中，过期旧笔记和权威日志在物理上处于完全平等的竞争地位，语义更接近问题的内容可能获得更高权重73% similarUnverifiedThe memory efficiency of encountering vocabulary in context can be 40%–60% higher than rote memorization in isolation.72% similarUnverifiedIn the experiments, plain text input achieved higher accuracy than input augmented with syntactic and rhetorical information.71% similarUnverifiedApproximately 15–25% of the content in spontaneous speech consists of filler words, sentence repetitions, self-corrections, and other redundancies.71% similarUnverifiedText generated by large language models tends to have low perplexity, with word choices that are overly consistent with probabilistic expectations, lacking the unexpectedness of human writing.70% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/907915API
curl https://kongchang.com/api/v1/knowledge/claims/907915MCP
get_claim(id=907915)