Stream: Scaling up Mechanistic Interpretability to Long Context in LLMs via Sparse Attention

TL;DR

This paper introduces Stream, a scalable hierarchical pruning algorithm that enables efficient interpretability of long-context attention patterns in large language models, making analysis feasible on consumer hardware.

Contribution

The paper presents Sparse Tracing and Stream, novel techniques for near-linear time and linear space analysis of attention in long-context LLMs, significantly improving scalability.

Findings

Stream retains critical attention paths while pruning 97-99% of interactions.

On RULER benchmark, Stream preserves key retrieval routes and discards 90-96% of interactions.

Stream enables long-context interpretability on consumer GPUs.

Abstract

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/.