斯坦福研究称44%的大模型经典结论已失效，平均寿命不到16个月
Start here
斯坦福大学研究人员针对 20 个模型重新检验了 50 项被广泛引用的 LLM 经典研究结论，发现其中 44% 已经失效，失效结论的平均寿命仅为 15.7 个月。研究指出，关于模型“不能做什么”的能力断言失效速度最快，例如关于 Strawberry 字母计数、中间上下文信息丢失等 9 项推理类缺陷结论已全部淡化或失效；与此同时，模型偏好、谄媚和策略性行为等倾向性特征则往往持续更久。
AI-generated from reports · Updated 11 hours ago
Report timeline
Follow the reports to see different sides of the story.
- The DecoderSelectedStanford Study Finds 44% of Classic LLM Research Conclusions No Longer Hold
Researchers at Stanford University re-evaluated 50 widely cited LLM research findings across 20 models and discovered that 44% of them are no longer valid, with an average lifespan of just 15.7 months for the invalidated claims. Negative capability claims about what models cannot do expired the fastest: all nine findings on reasoning flaws—such as counting letters in "strawberry" and lost-in-the-middle context issues—have either faded or become obsolete. In contrast, behavioral tendencies such as model preferences, sycophancy, and strategic behavior tend to persist longer.
Story heat trend
Current heat 7·Comparable peak 10（Nov 4 05:00）·Comparable change, last 24 hours–
Trends compare only the same parties observed continuously, so the scope can be narrower than current heat counts. Move the pointer or tap the chart to see hourly heat; use the left and right arrow keys on a keyboard.