Model Resonance Imaging — visualize LLM internals like a brain MRI

We study whether categorical refusal tokens enable controllable and interpretable safety behavior in language models.

MSc Thesis: Bridging mechanistic interpretability circuits to faithful natural language explanations using ERASER evaluation metrics

Mechanistic interpretability tool visualizing GPT-2's layer-by-layer predictions using the logit lens technique

An independent, from-scratch reproduction of the mechanistic-interpretability findings in Anthropic's When Models Manipulate Manifolds: The Geometry of a Counting Task

Replication of 'From Reasoning to Answer' (EMNLP 2025) — Reasoning-Focus Heads + Activation Patching on DeepSeek-R1-Distill-Qwen-7B

Mechanistic interpretability study comparing modular addition and subtraction circuits in 1-layer attention-only transformers via activation patching, logit lens, SVD circuit analysis, Fourier feature analysis, and causal scrubbing across three training stages.

Mechanistic interpretability tool to detect induction heads in GPT-2 using TransformerLens

Measuring attention similarity bias in GPT-2 variants via TransformerLens. Replicates Figure 1 of arXiv:2603.09078. Finds U-shaped trend, not the monotonically increasing one the paper reports.

Mechanistic interpretability: belief-state geometry in a transformer's residual stream. From-scratch replication of Shai et al. 2024 (arXiv:2405.15943).

Mechanistic interpretability of small transformers: RSK correspondence and Pythia-70m

Causal intervention framework for mechanistic interpretability research. Implements activation patching methodology for identifying causally important components in transformer language models.

Mapping attractor basins in GPT-2 Small via iterative activation re-injection — a 125-prompt sweep inspired by Alvin Lucier's 'I Am Sitting in a Room'.

Local agent-driven mechanistic interpretability research platform for Apple Silicon

Forensic suite for Mechanistic Interpretability in Transformers. Implementing 0.0054 Basal Accountability Gradients for auditing model logic using TransformerLens and SAELens

Automated Forensic Discovery of Reasoning Circuits in Transformers

Ask your coding agent WHY a language model made a prediction — mechanistic interpretability (logit lens, activation patching, SAE features, steering) as a drop-in agent skill. Validated against published circuits.

A research tool for studying how deception emerges in multi-agent LLM systems and detecting it through activation analysis.

Code used for reverse-engineering a “Query-Gated Courier” circuit in Gemma-2-2B for role-gated retrieval.