Ajinkya Kiran Mulay
Notes on sparse recovery, Transformer optimization, AI safety, and verifiable systems. I build methods whose evidence can be inspected, including the negative results.
FCE beat a permuted-reward control by 20.8 points across four SmolLM seeds and raised Qwen2.5-1.5B from 54.6% to 66.8% on GSM8K. Majority, harder-domain transfer, and unanimous-wrong-consensus claims remain explicit non-wins.
A displacement-controlled route from row-energy conditioning back to canonical Muon, plus the harder question of extending the idea to coupled attention circuits.
Use the orthogonal SVD basis, select input-specific components with OMP, and compare against trained decomposition methods without hiding the unsupported causal claims.
A small system where models propose Python solutions and a separate deterministic validator searches for counterexamples before accepting them.
An explicitly exploratory model-internal signal tested in the Apart AI Control Hackathon, with its evidence boundary kept visible.
An interactive investigation of attention sinks, over-mixing, and distributed sink banks built as a Marimo notebook.
An evolving list of free and open-source tools for research, writing, and focused work.
OMP, FoBa, CoSaMP, and the question of how a system corrects a bad support choice.
Muon-family methods, polar maps, architecture-aware routing, and honest wall-clock gates.
Behavioral evaluations and model-internal signals with explicit validity limits.
Let models search broadly, then require executable evidence or an independent validator.
Why this name?
Pursuit algorithms commit to a support, test it against the residual, and revise when the evidence says they chose wrong. That is also how I want to do research.