Distributed Mechanistic Interpretability at Scale: Activation Streaming, Split-Layer Inference, and Distributed Sparse Autoencoder Training

Q: What does this research mean for the field?

A two-node system called ActStream enables efficient distributed training of sparse autoencoders by streaming intermediate activations over a binary wire protocol, achieving high throughput and low latency without staging to disk. Novelty: ClaimNovelty.METHODOLOGICAL. Consensus alignment: ConsensusAlignment.NEUTRAL.

J. Melton

doi:10.5281/zenodo.22699164

September 11, 2026Open Access

Distributed Mechanistic Interpretability at Scale: Activation Streaming, Split-Layer Inference, and Distributed Sparse Autoencoder Training

View Full Paper

Ask AI

Bookmark

Share

Key Points

Design and evaluate a distributed system that enables mechanistic interpretability by streaming activations between inference and analysis nodes without intermediate disk staging.
Constructed ActStream, a two-node architecture splitting inference and analysis via a binary wire protocol across Thunderbolt and LAN connections.
Trained sparse autoencoders directly on intermediate model activations without writing data to persistent storage.
Evaluated an activation feature-based safety classifier on the PKU-SafeRLHF benchmark against a large language model judge.
Thunderbolt achieved a transfer throughput of 2.150 GB/s, reaching 86% of the negotiated rate and outperforming the LAN path by 26-fold.
Transmission stability was maintained with 99th-percentile latency measured within 1.5x of the median latency.
Successfully integrated and validated downstream feature-based safety classification using streamed activations against the PKU-SafeRLHF baseline.

Ask about this paper

Answers are grounded in this paper and cite their sources.

Implication

Benchmarking study demonstrates high-throughput activation streaming across a two-node system without disk staging, indicating scalable real-time distributed mechanistic interpretability.

Abstract

ActStream: a two-node system that splits inference and analysis, streams intermediate activations over a binary wire protocol, and trains sparse autoencoders without staging activations to disk. Thunderbolt measured at 2.150 GB/s, 86% of negotiated rate and 26x the LAN path, with p99 latency within 1.5x of median. Includes a feature-based safety classifier evaluated against an LLM-judge baseline on PKU-SafeRLHF.

Discussion

No takes yet. Share an insight, caveat, or question.

Cite This Study

J. Melton (2026) studied this question.

synapsesocial.com/papers/6aa7b2b20926e14a848b1369 https://doi.org/10.5281/zenodo.22699164

View Full Paper

Ask AI

Bookmark

Share

Distributed Mechanistic Interpretability at Scale: Activation Streaming, Split-Layer Inference, and Distributed Sparse Autoencoder Training

Key Points

Ask about this paper

Implication

Abstract

Discussion

Cite This Study

Distributed Mechanistic Interpretability at Scale: Activation Streaming, Split-Layer Inference, and Distributed Sparse Autoencoder Training

Key Points

Ask about this paper

Implication

Abstract

Discussion

Cite This Study

Keep reading

Keep reading