Expired50% confidenceFactTime unknown

Research shows that the same base LLM model can vary by 30-50 percentage points in performance under different system prompts.

1

Sources

50%

Confidence

Medium-term (~90 days)

Relevance

7/2/2026

First Seen

Valid until: 9/30/2026(expired)

Sources

Deep Dive into Hermes Agent Self-Evolution: The GIPA Algorithm and Engineering Practice

bilibili探索未至之境

Related Claims

Unverified同一个基座模型在不同系统提示词下的表现差异可以高达30-50个百分点79% similar UnverifiedResearch shows that, for the same task, an optimized prompt can improve model output quality by 30%–50% compared to arbitrary input.71% similar Unverified研究表明，即使在 GSM8K 或 MATH 数据集上接近满分的模型，面对稍作变形或需要跨领域组合的问题时性能下降可达30-50个百分点71% similar

Cite This Claim

Stable URI

https://kongchang.com/claim/55461

API

curl https://kongchang.com/api/v1/knowledge/claims/55461

MCP

get_claim(id=55461)