Unverified50% confidenceBenchmarkExact time
On MIMIC-CXR and MIMIC-IV BHC, the model performs well on unstructured narrative text but degrades to near position-baseline levels when faced with highly templated reports.
1
Sources
50%
Confidence
Long-term
Relevance
9/15/2026
First Seen
Sources
Related Entities
Related Claims
Unverified经过指令微调和RLHF训练的模型在遵循格式规范方面表现更好，有时反而会生成格式更完美的虚假引用66% similarUnverifiedAlthough multimodal models already possess image understanding capabilities, plain text input can still provide higher reliability and consistency when precisely extracting structured information.66% similarUnverifiedIn an A/B test using exactly the same plot, the structured prompt version was noticeably more professional than the plain-language version in terms of camera language, character consistency, and narrative pacing.65% similarUnverifiedExplicitly constraining output format and style in the system prompt can reduce model style drift, making generated results more consistent with the existing codebase's consistency requirements.64% similarUnverifiedThe 'structure' principle in prompts leverages the model's sensitivity to formatted text, because structured text in pretraining data typically corresponds to high-quality content.64% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/921886API
curl https://kongchang.com/api/v1/knowledge/claims/921886MCP
get_claim(id=921886)