Mechanistic Interpretability Research Handbook

A 100–150 page field guide to mechanistic interpretability — how the science developed, what it has found, and where it is going. Rather than reviewing papers independently, this handbook traces the evolution of the field, showing how each work responds to the limitations of the one before it.

---

How to use this handbook

• Students: Read sequentially from Part I. Each chapter is self-contained but builds on previous chapters.
• Researchers: Jump to the Synthesis sections for a bird's-eye view of open problems and research directions.
• Project ideas: Every chapter ends with concrete reproduction and extension projects, rated by difficulty.
• Quick reference: See Comparison Tables for a cross-paper summary.

---

Table of Contents

Front Matter

• Preface
• How to Read This Handbook
• Reading Roadmap
• Comparison Tables

---

Part I — Foundations: Seeing Inside Neural Networks

Chapter	Paper	Year
Ch. 1 — Why Interpretability?	Synthesis chapter	—
Ch. 2 — Feature Visualization	Olah et al., Distill	2017
Ch. 3 — Activation Maximization	DeepDream / feature inversion	2015–2016
Ch. 4 — OpenAI Microscope	OpenAI (tool)	2021

---

Part II — Circuits: Understanding Computation

Chapter	Paper	Year
Ch. 5 — Zoom In: An Introduction to Circuits	Olah et al., Distill	2020
Ch. 6 — The Circuits Thread	Multiple (Distill.pub)	2020–2022
Ch. 7 — Activation Atlases	Carter et al., Distill	2019

---

Part III — Representations: Polysemanticity and Its Solutions

Chapter	Paper	Year
Ch. 8 — Toy Models of Superposition	Elhage et al., Anthropic	2022
Ch. 9 — Towards Monosemanticity	Bricken et al., Anthropic	2023
Ch. 10 — Sparse Autoencoders	Companion chapter	—

---

Part IV — Frontier Models: Scaling to Large Language Models

Chapter	Paper	Year
Ch. 11 — Mapping the Mind of a Large Language Model	Templeton et al. (Anthropic)	2024
Ch. 12 — Language Models Can Explain Neurons	Bills et al. (OpenAI)	2023
Ch. 13 — Attribution Graphs: Biology of an LLM	Lindsey et al. (Anthropic)	2025

---

Part V — Emerging Directions

Chapter	Paper	Year
Ch. 14 — Progress Measures for Grokking	Nanda et al.	2023
Ch. 15 — Discovering Latent Knowledge	Burns et al.	2022
Ch. 16 — A Global Workspace in Language Models	Gurnee, Anthropic	2026

---

Synthesis

Section	Contents
Timeline	How the field evolved, year by year
Themes	Four cross-cutting themes across all papers
Open Problems	Collected research gaps, grouped by cluster
Research Directions	Concrete project ideas with difficulty ratings

---

Field evolution at a glance

2015–2019   "What does this neuron do?"
            Feature visualization, activation maximization, DeepDream
                ↓
2020        "How do neurons work together?"
            Circuit analysis reveals distributed algorithms
                ↓
2022–2023   "Why do neurons represent multiple concepts?"
            Superposition theory + sparse autoencoders
                ↓
2024        "Can we interpret production-scale models?"
            1M–34M features extracted from Claude 3 Sonnet
                ↓
2025+       "Can we trace internal computation chains?"
            Attribution graphs and biology of LLMs
                ↓
2026        "Can we read what the model is thinking before it speaks?"
            Jacobian lens, J-space, global workspace

---

A note on scope

This handbook covers fourteen papers and one tool. It is not exhaustive — the mechanistic interpretability literature now spans hundreds of papers. The selection criteria were: papers that changed the field's direction, introduced methods still in active use, or surfaced questions that remain open today.

For coverage of the broader AI safety landscape — evaluations, RLHF, deceptive alignment, control — see the companion wiki.