Unverified50% confidenceFactExact time
Text-level monitors can only analyze a model's explicit outputs and cannot detect cases where the model stays clean at the output level but retains inconsistent objectives internally.
1
Sources
50%
Confidence
Long-term
Relevance
10/13/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedAn output parser filling in default values when it encounters a malformed response leaves downstream code unable to distinguish whether a field was produced by the model or fabricated by the parser, which is a silent failure mode.71% similarUnverified针对特定触发词的后门攻击可让模型在大多数情况下表现正常，仅在遇到特定输入时才输出预设错误信息，使常规评测基准难以捕捉71% similarUnverifiedA model's sampling temperature, prompt wording, and version iterations can all affect output results, yet none of these parameters have a corresponding place in traditional experimental reporting frameworks.70% similarUnverified传统评测框架往往只能覆盖预设的测试场景，无法捕捉真实用户交互中的隐性失败69% similarUnverifiedUnder injection bait, weak models treat the injected incorrect numbers as their own calculation results and fabricate verification processes out of thin air, with the explanatory text appearing consistent with genuine verification, making it impossible to detect errors through auditing.68% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1013461API
curl https://kongchang.com/api/v1/knowledge/claims/1013461MCP
get_claim(id=1013461)