From Token Lists to Graph Motifs: Weisfeiler-Lehman Analysis of Sparse Autoencoder Features

Researchers introduce a novel graph-based analysis method for sparse autoencoders (SAEs) in transformer models, using Weisfeiler-Lehman graph kernels to examine token co-occurrence patterns in SAE features. Applied to GPT-2 Small, this approach identifies structural motif families that traditional decoder weight analysis misses, revealing complementary insights into how neural networks organize semantic information.

This research advances mechanistic interpretability—the field focused on understanding how neural networks process information internally. Sparse autoencoders have emerged as a key tool for decomposing transformer activations into interpretable features, but existing analysis methods rely heavily on token frequency lists and decoder weights. The authors propose a structural perspective by modeling each SAE feature as a graph where nodes represent tokens and edges capture co-occurrence patterns within local context windows. They develop a custom graph kernel based on Weisfeiler-Lehman algorithms to compute similarity across this structural space, then apply clustering to discover feature motifs.

The research operates within a broader mechanistic interpretability trend gaining momentum as AI systems become more powerful and opacity increasingly problematic. Understanding internal feature structures has implications for AI safety, alignment research, and model debugging. The findings demonstrate that graph-structural relationships reveal feature organizations—punctuation-heavy patterns, language clusters, code-like templates—that frequency and weight-based methods overlook. However, a simpler token-histogram baseline outperformed the graph approach in overall clustering purity, suggesting the graph view provides complementary rather than dominant insights.

For the AI research community, this work demonstrates that different analytical lenses capture distinct aspects of model behavior. The stability of results across hyperparameters and random seeds indicates the method's robustness. While this fundamental research doesn't directly impact cryptocurrency markets or trading, it strengthens interpretability infrastructure that underpins AI safety discussions increasingly relevant to blockchain and AI governance conversations. The methodological contribution could influence how researchers analyze larger language models and multimodal systems.

• →Graph-based SAE analysis surfaces structural feature relationships invisible to token-frequency and decoder-weight approaches.
• →Applied to GPT-2 Small, the method discovers coherent motif families including punctuation patterns, language clusters, and code templates.
• →Token-histogram baselines achieved higher clustering purity, indicating the graph view provides complementary rather than superior insights.
• →Results remain stable across different hyperparameter configurations and random initializations, supporting method reliability.
• →This mechanistic interpretability advance strengthens understanding of how transformers organize semantic information internally.