Early-Signal Detector

method · first seen 4 Oct 2026

circuit analysis

Also: mech interp, interpretability, mechanistic learning, feature decomposition, mechanistic interpretability, interpretability of transformers

Identifying sparse subgraphs of a neural network that implement specific behaviors.

Mechanistic interpretability, including circuit analysis, tries to explain neural network behavior by locating the internal features, neurons and sparse subgraphs that carry out specific computations. The sources here apply long-standing tools to new settings rather than introducing the idea. Papers study suppressed safety features, carry propagation in small multiplication transformers, uncertainty quantification, an astronomy foundation model, and web-agent hallucination. Others question the method itself, for example how to separate steering effects from actual mechanism and what logic of inference supports mechanistic claims. On the code side, there are hobbyist repos for neuron steering and exploring Qwen models, plus a Hacker News discussion of using an LLM to interpret model weights.

The field has been an active research area since roughly 2020, so this burst reflects steady output rather than a new idea. The points worth watching are its spread into scientific foundation models and agents, and the growing attention to whether interpretability findings are causally valid.

Analyst note (established), 17 Nov 2026.

Mentions per week

01619 May11 Aug10 NovPapersCodeDiscussion

PapersCodeDiscussion

Rank 81 on 14 Nov 2026 · 58 mentions in 28 days · 57 independent groups · papers → discussion → code · first seen 6 weeks ago · acceleration -0.71

Earliest sources

Latest sources

Mentioned together with

sparse autoencoderactivation steeringlogit lenscircuit discoveryactivation patchingmodel editingattention headsuncertainty quantification in language modelsout-of-distribution extrapolationsafety feature discovery in transcoder features