Research · methods and evidence

Structure first. Claims second.

My work starts from sparse recovery and keeps returning to the same practical question: what structure can we exploit without pretending the evidence says more than it does?

Active research

Current threads

Evidence labels describe today’s boundary, not the ambition.

Verifier-free reasoning RLreplicated policy-learning result

Pursuit-Graded Reward (PGR/FCE)

A dense GRPO reward built from the structure of sampled reasoning trajectories rather than a gold answer checker. FCE uses within-group trajectory evidence to assign step-level learning signal, then tests whether that signal survives permuted-reward, base-model, majority, gold-reward, and self-certainty controls.

Evidence boundary

Across SmolLM training seeds 43-46, FCE beat the permuted control by 20.8 percentage points (95% CI 16.4-24.7), with every seed positive. On Qwen2.5-1.5B, FCE reached 66.8% on held-out GSM8K versus 54.6% for the base model and 49.2% for the permuted control. It did not significantly beat ordinary majority or gold GRPO; MATH-Hard and MMLU policy-transfer gates did not pass; support-only rewards cannot identify fully unanimous wrong consensus.

Code and experiment artifacts ↗

Transformer optimizationactive research

SemanticDose and Halley-Gram-Muon

Architecture-aware Muon variants that change update geometry only where the parameter semantics support it. SemanticDose retires by cumulative displacement; Halley replaces the polar map with a cheaper rational iteration.

Evidence boundary

Public implementation and frozen experiment artifacts. Attention generalization remains gated on the new MHA/GQA/MQA queue.

Semantic optimizers ↗Halley-Muon ↗

Training-free decompositionpublic artifact

SVD-OMP

An orthogonal SVD dictionary with input-specific OMP support selection for decomposing Transformer MLP parameters without fitting a learned dictionary.

Evidence boundary

The public evaluation reports reconstruction, faithfulness, coherence, reproducibility, and support-stability comparisons on Goodfire 67M matrices. Causal interpretation remains future work.

Code and results ↗

AI-control evaluationprovisional evidence

SecretLoyaltyBench

A behavioral audit for principal loyalty and hidden-objective failures, with explicit separation between evaluated organisms and untested transfer claims.

Evidence boundary

The public artifact preserves held-out behavioral results and known limitations. Automatic-judge calibration and independent audit are still pending.

Evaluation artifact ↗

Selected publications

Published work

2024

SPriFed-OMP: A Differentially Private Federated Learning Algorithm for Sparse Recovery ↗

Transactions on Machine Learning Research

Federated sparse recovery with OMP, secure aggregation, privacy constraints, and formal convergence guarantees.

2022

LOCKS: User Differentially Private and Federated Optimal Client Sampling ↗

Preprint

Client sampling for user-level differential privacy in federated networks.

2022

Private Hypothesis Testing for Social Sciences ↗

ICML Workshop on Theory of Differential Privacy

Statistical power and sample-size planning under differentially private measurements.

2022

PowerGraph: Learning multivariate statistical power trade-offs ↗

ICML AI for Science Workshop · IMPS Oral

Neural and PCA approximations for statistical-power estimation at a fraction of brute-force cost.

2021

Towards Quantifying the Carbon Emissions of Differentially Private ML ↗

ICML Workshop on Socially Responsible Machine Learning

An early accounting of the computational and carbon costs induced by private learning.

Operating rule

Negative results stay in the record.

Tuning is not confirmation. A local benchmark is not an external validation. A useful model-internal signal is not automatically a causal explanation. Those distinctions make the successful results more useful, not less.