AITOPAI热报
PapersOfficial85°

Anthropic 研究：Claude 3.5 Haiku 内部机制如同生物系统

On the Biology of a Large Language Model Lindsey et al., 2025 We investigate the internal mechanisms used by Claude 3.5 Haiku — Anthropic's lightweight production model — in a variety of contexts.

Primary sourceAnthropic: Transformer Circuits
News· 11 sources
Read the original
Why we picked it

Anthropic 把模型内部机制当生物系统来研究，做 AI 可解释性的人会看到新方法论，关心模型安全性的团队值得关注。

Anthropic 发布了一项关于 Claude 3.5 Haiku 内部机制的研究，将其类比为生物学研究。

Image · Anthropic: Transformer Circuits
Body

他们通过构建“归因图”来追踪模型在多种任务中的推理路径，发现模型内部存在类似生物体的模块化结构和功能分工。这项研究揭示了大型语言模型如何从输入到输出逐步构建理解，例如在数学推理中，模型会先识别问题类型，再调用特定计算模块。

该工作为理解 AI 黑箱提供了新视角，有助于提升模型的可解释性和安全性。

Industry Coordinates & Source Verification

This report is categorized under AITOP's "Papers" desk, focusing on US frontier AI labs and Silicon Valley technology leaders. Key entities and technical topics under continuous tracking include 可解释性, 归因图, Claude 3.5 Haiku, 模型内部机制.

Primary source verification: published directly by Anthropic: Transformer Circuits with first-party authority. Cross-source corroboration: 11 independent sources have covered this development (including Dario Amodei Blog, Claude: Blog, The Rundown AI), confirming industry-wide signal strength.

#可解释性#归因图#Claude 3.5 Haiku#模型内部机制#Anthropic

Keep reading

Recent stories with the same tag or section
PapersUS vs ChinaMulti-sourceSelected05:23
OpenAI research finds chain-of-thought monitorability declines as models scale

The non-obvious takeaway is that safety monitoring based on reading a model's reasoning traces may silently weaken as capability grows, meaning audits that work today may not work on next year's models. Safety researchers, alignment teams and enterprises using CoT inspection for oversight should benchmark monitorability on their own fine-tunes. The training-lifecycle framing gives practitioners a concrete methodology for measuring when and where this degradation starts.

#OpenAI#chain-of-thought#AI safety#reasoning models
Event
elvis@omarsar0Source
PapersBuilders13:18
Chinese contentCCW-Net 提出因果化概念解释框架，将中介分析引入神经网络可解释性

做自动驾驶或安全认证的同学可以看看，这篇把流行病学里的中介分析搬进了神经网络训练，解释不再是相关而是因果，还有必要性、充分性、独立性三条可量化准则，比传统概念解释方法严谨不少。

#CCW-Net#可解释性#中介分析#自动驾驶
arXiv cs.AI@Josh Rountree et al. (5)Source
PapersBuildersMulti-source11:36
Chinese content研究表明 MLLM 可充当人类替身评估可解释 AI

做 XAI 研究的朋友可以看这篇：作者测了 14 个多模态模型，发现 OpenAI 和 Qwen 对齐人类判断最好，用它跑解释质量评估能省下真人实验的钱和时间。论文还给了可复用的评估框架和指标定义，不用从零设计。

#XAI#MLLM#Qwen#OpenAI
Event
arXiv: OpenAI@Eoin M. Kenny et al. (3)Source
PapersBuilders15:11
Chinese contentarXiv 论文：可解释性工具可跨模型复用，覆盖 110 组模型对验证

做模型可解释性的人值得看这篇：以前每个模型都要重新拟合拒绝方向和探针，现在证明换个模型也能直接用，加少量无标签 prompt 就能接近原生效果。覆盖五个模型家族、1B 到 72B，实验量够扎实。

#可解释性#arXiv#steering vector#迁移学习
arXiv cs.AI@Marcus Armstrong et al. (5)Source