Unverified50% confidenceFactExact time
The study finds that hallucinations manifest as three types of information flow distortion patterns: over-reliance on self-attention, diffuse contextual retrieval, and excessive information compression.
1
Sources
50%
Confidence
Long-term
Relevance
9/21/2026
First Seen
Sources
Related Entities
Related Claims
Unverified过度膨胀的上下文会稀释模型注意力、抬高推理成本，甚至引入噪声导致幻觉78% similarUnverifiedResearch shows that hallucination rate is negatively correlated with the knowledge density of the prompt; the richer and more precise the context provided in the prompt, the lower the probability that the model will generate fabricated content.76% similarUnverifiedWhen prompts are too long, large models experience an 'attention explosion' phenomenon, where excessively long inputs cause effective information to be lost.73% similarVerified研究表明当上下文过长时，模型对中间部分信息的注意力会显著下降，即"Lost in the Middle"现象71% similarUnverified模型规模越大，幻觉的表述往往越流畅、越具迷惑性，识别难度反而上升71% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/938011API
curl https://kongchang.com/api/v1/knowledge/claims/938011MCP
get_claim(id=938011)