Unverified50% confidenceOpinionExact time
SWORD researchers believe that models rely on distributional familiarity rather than a genuine fact-verification mechanism.
1
Sources
50%
Confidence
Long-term
Relevance
9/11/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedSWORD research found that model accuracy on semantically plausible distorted statements was actually higher than on meaningless random substitutions.72% similarUnverified在叙事场景中，传统基于语义相似度的检索策略容易遗漏因果隐含的关键节点，因为伏笔与爆发点在向量空间中相似度不一定高65% similarUnverifiedVector retrieval based on semantic similarity has limitations: it may miss decisions that are chronologically important but semantically dissimilar, lacks a conflict resolution mechanism, and critical information may be omitted or drowned out.63% similarUnverifiedRLHF中人类评审倾向给流畅自信全面的答案更高分而较少关注事实核查，导致模型获得隐性激励表现得有把握比实际有把握更重要63% similarUnverified反事实问题能有效区分真正理解与模式匹配，因为这类问题几乎不可能直接出现在训练语料中63% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/894729API
curl https://kongchang.com/api/v1/knowledge/claims/894729MCP
get_claim(id=894729)