Unverified50% confidenceFactExact time
A preliminary analysis of 363 trajectories evaluated by an LLM judge shows that process trajectories expose behavioral differences that pure output evaluation cannot capture.
1
Sources
50%
Confidence
Long-term
Relevance
9/11/2026
First Seen
Sources
Related Entities
Related Claims
Unverified研究表明模型在处理约束性指令时的准确率显著低于处理生成性指令，这被称为'约束遵循缺陷'67% similarUnverifiedLLM在「生成」和「评估」时激活的推理路径不同，存在「生成盲区」与「评估敏感区」的不对称性，这是多Agent Review能发现更多问题的底层原因66% similarUnverifiedAgent 评测中由于错误沿任务链累积传播，最终完成率的分差往往比传统 NLP 基准更难弥合66% similarUnverified研究者通过前向—后向循环检查（forward–backward cycle check）过滤遮挡、快速运动和估计失败区域的错误对应关系66% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/907731API
curl https://kongchang.com/api/v1/knowledge/claims/907731MCP
get_claim(id=907731)