Sparse Autoencoders Map Brain-LLM Alignment onto Cortical Semantic Topography 稀疏自编码器将脑-LLM对齐映射到皮层语义拓扑结构上

Background

The article addresses the gap in understanding why intermediate layers of large language models (LLMs) best predict human brain responses to language. Despite this robust finding in computational neurolinguistics, the mechanisms behind it remain unexplained. To bridge this knowledge gap, researchers decompose LLMs into interpretable features using sparse autoencoders (SAEs).

Key Points

The study decomposes GPT-2 XL and Llama-3.1-8B into 16K-32K interpretable features per layer. A human-validated taxonomy with a Kappa score of at least 0.74 was used to categorize these features. The findings reveal that semantic features alone account for 94% of peak encoding performance, significantly outperforming variance-matched baselines.

Decomposition and Taxonomy

1. Decomposition: LLMs are decomposed into fine-grained interpretable features.
2. Taxonomy Validation: A taxonomy with a Kappa score of at least 0.74 was validated by human experts, ensuring the reliability of the feature categorization.

Semantic Feature Dominance

• Peak Encoding Performance: Semantic features alone recover 94% of peak encoding performance (r = 0.285).
• Comparison with Baselines: Variance-matched baselines fall short, with significant differences shown through statistical tests (p < 0.001, d = 1.31).

Significance

The study introduces a novel approach to understand the mechanistic link between language models and brain responses.

Novel Cortical Topography Prediction

• Priori Categorization: Five semantic subcategories derived from three independent neuroscience programs were tested.
• Convergence Test: A formal test confirmed that SAE-discovered features map onto distinct brain regions (Spearman ρ = 0.72, p < 0.001; hypergeometric p = 0.007), showing a granularity not achieved by previous methods.

Prediction of Reading Times

• Beyond Lexical Controls: SAE features predict human reading times beyond simple lexical controls (ΔlogLik = 38.4, p < 0.001).
• Exploratory Analyses: Preliminary evidence suggests the brain encodes unexpected semantic content, providing a new dimension to understanding language processing.

Generalization

The findings generalize across English, Chinese, and French, suggesting cross-linguistic applicability of SAEs in understanding neural responses.

Conclusion

By bridging sparse autoencoders with neural encoding models, this study offers profound insights into the mechanistic relationship between LLM layers and brain activity. This work not only advances our understanding of computational neurolinguistics but also provides a new methodological framework for future research.

背景与问题

本文探讨了大型语言模型（LLMs）的中间层如何预测人类大脑对语言的反应。尽管已有研究表明，中间层是最佳预测器之一，但其背后的机制仍然不明确。文章提出通过结合稀疏自编码器（SAEs）和神经编码模型来解释这一现象。

核心内容

研究首先将GPT-2 XL和Llama-3.1-8B分解为每层16K-32K可解析的特征，并采用经过人类验证的分类体系（κ≥0.74），发现语义特征几乎能够解释模型的最大编码性能，占比高达94%，且显著优于对照组（p<0.001，d=1.31）。研究进一步假设五个先验确定的神经科学程序衍生出的语义子类别应该对应不同的大脑区域。通过正式的收敛性测试，证实了这些假设有道理（Spearman ρ=0.72，p<0.001；超几何概率p=0.007），表明SAE发现的特征再现了认知皮层语义组织的细粒度结构。

研究还展示了这些特征可以预测人类阅读时间，并且在词汇控制之外进一步提供了大脑编码意外语义内容初步证据。研究结果在英语、汉语和法语中都能泛化适用。

意义与影响

本文通过结合SAEs和神经编码模型，揭示了大型语言模型中间层的可解析特征能够解释认知皮层对语言反应的关键方面，并且这种机制比以往方法更加精细和准确。这一发现不仅加深了我们对语言处理机制的理解，还为未来的跨学科研究提供了新的工具和技术。