Selected work
Projects, in depth
A closer look at the research and systems I've built. Open a project for the full write-up, clips, and code.
Steering language models: manifold and concept steeringInterpretability research on how concepts live inside language models, and techniques to steer them.
- Challenged Goodfire's state-of-the-art manifold steering claims by designing a simple vector steering method that reduces correct-output probability by at most 1%.
- Proved that same-concept manifolds across different languages collapse into one shared structure as models scale in size.
The behaviour manifold Recovering the manifold from concept anchors The manifold emerging from unprompted generations Instruction-tuning collapses geometry silently Learning to control an exact solverAn RL-learned branching heuristic for exact, #P-hard chromatic polynomials. 19% smaller search trees than the standard heuristic, verified against a brute-force oracle with zero mismatches across all 328 instances tested.
- Built a C++ deletion-contraction engine (exact GMP arithmetic, a nauty isomorphism cache, and clique-separator decomposition) and trained an RL policy to choose the branching edge: 19% reduction in search trees from the standard heuristic.
- Derived graph instances from Natural Earth geodata (adjacency as geodesic shared-border length) and computed the exact chromatic polynomial of the world map (195 UN-recognized entities, χ = 4, 34,703 branches) plus all 214 countries' internal subdivisions, served through an interactive recursion-tree visualiser.
Example of recursion tree interface Chromatic Polynomial of the United States