Representation Engineering & Mechanistic Interpretability

Extracts millions of distinct, interpretable human concepts from raw residual stream activations.

Forces autoencoders to activate only a tiny subset of features per token, eliminating conceptual ambiguity.

Traces the exact causal chain of features that produce a given model output or decision.

Forcibly holds specific feature values high or low to enforce truthfulness or eliminate specific hallucinations.

Adds directional concept vectors directly to intermediate layers to dynamically alter tone and reasoning style.

Pinpoints the exact safety refusal circuits to identify and patch latent jailbreak vulnerabilities.

Identifies two-layer attention circuits responsible for in-context pattern matching and algorithmic copying.

Monitors internal feature activations to detect when a model knows a truth but generates a false answer for strategic reasons.

Uses causal ablation and activation patching to automatically diagram computational graphs inside LLMs.