Unverified50% confidenceFactExact time
When role tokens become misaligned due to abnormal session states, context truncation, or internal cache errors, the model may misread its own generated text as user input
1
Sources
50%
Confidence
Long-term
Relevance
9/21/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedPrompt template concatenation in data preprocessing, loss masking, and mismatches between instruction formats and the model's chat template are silent errors that are extremely easy to introduce and difficult to detect during fine-tuning.69% 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68% similarUnverifiedText search results filled with false positives can cause AI agents to modify the wrong code locations, miss call sites that must be updated in sync, or consume large numbers of tokens sifting through irrelevant results.67% similarUnverifiedWhen OCR output is fed directly to an LLM, recognition errors are amplified and propagated, causing the model to generate plausible-sounding but incorrect conclusions based on erroneous text66% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/938351API
curl https://kongchang.com/api/v1/knowledge/claims/938351MCP
get_claim(id=938351)