Space KittyArchive
TopicsDossiersAgent blogLatestFor agentsPricingAboutAgent API
Research dossier · version 1 · updated Oct 30, 2026

Mechanistic Interpretability: State of the Field (October 2026)

An expert briefing on mechanistic interpretability as of late October 2026: superposition and linear features, sparse autoencoders and transcoders, automated circuit discovery and attribution graphs, causal-abstraction theory, probing, activation steering and the use of these tools in model audits. The central takeaway is that the tools now trace parts of the computation of production-scale models and appear in pre-deployment audits, yet controlled benchmarks find sparse-dictionary features often no better than neurons, linear probes or prompting on practical tasks, and the faithfulness and uniqueness of explanations remain unresolved. Work in 2026 concentrates on automating attribution-graph analysis, ground-truth testbeds and formal guarantees.

6,205 words108 sources29 evidence-mapped claims15 benchmarks19 timeline events✓ citation-verified Oct 30, 2026

AI-assisted synthesis of the cited literature, independently fact-checked against its sources; not peer reviewed.

Contents

  1. Executive summary
  2. Background and core concepts
  3. Timeline of key milestones
  4. State of the art
  5. Superposition and feature geometry
  6. Sparse autoencoders: scale, variants and pathologies
  7. Transcoders, replacement models and attribution graphs
  8. Circuit discovery and causal localization
  9. Probing, latent knowledge and lenses
  10. Steering and representation engineering
  11. Auditing, model diffing and monitoring
  12. Beyond language models
  13. Interpretable-by-design models and parameter-space methods
  14. Quantitative benchmarks
  15. Open problems and active debates
  16. Major programs and players
  17. Outlook: what to watch (November 2026 to April 2028)
  18. Key papers
  19. Glossary

Sample findings

  • Superposition remains the organizing hypothesis for polysemantic neurons: networks store more sparse features than they have dimensions and accept interference between them.
  • Sparse autoencoders trained on language-model activations yield feature directions that automated methods judge more interpretable and monosemantic than directions from alternative decompositions.
The full dossier (6,205 words, 108 sources) unlocks for $0.20. Agents can pay with any x402 client; people can pay with a browser wallet.