Stream: Scaling up Mechanistic Interpretability to Long Context in LLMs via Sparse Attention
This provides a practical tool for researchers and practitioners to analyze attention patterns in long-context LLMs, making interpretability more accessible, though it is an incremental improvement over existing sparse attention methods.
The paper tackles the problem of scaling mechanistic interpretability to long contexts in LLMs, where traditional techniques require excessive memory, by introducing Sparse Tracing, which prunes 90-99% of token interactions while preserving critical retrieval paths and enabling analysis on consumer GPUs.
As Large Language Models (LLMs) scale to million-token contexts, traditional Mechanistic Interpretability techniques for analyzing attention scale quadratically with context length, demanding terabytes of memory beyond 100,000 tokens. We introduce Sparse Tracing, a novel technique that leverages dynamic sparse attention to efficiently analyze long context attention patterns. We present Stream, a compilable hierarchical pruning algorithm that estimates per-head sparse attention masks in near-linear time $O(T \log T)$ and linear space $O(T)$, enabling one-pass interpretability at scale. Stream performs a binary-search-style refinement to retain only the top-$k$ key blocks per query while preserving the model's next-token behavior. We apply Stream to long chain-of-thought reasoning traces and identify thought anchors while pruning 97-99\% of token interactions. On the RULER benchmark, Stream preserves critical retrieval paths while discarding 90-96\% of interactions and exposes layer-wise routes from the needle to output. Our method offers a practical drop-in tool for analyzing attention patterns and tracing information flow without terabytes of caches. By making long context interpretability feasible on consumer GPUs, Sparse Tracing helps democratize chain-of-thought monitoring. Code is available at https://anonymous.4open.science/r/stream-03B8/.