Research
My current public work centers on generative models, latent world models, planning, and compute allocation. I treat evaluation design as part of the research contribution: a useful result should remain understandable when the headline hypothesis fails.
Few-Step Field Regularity
Question. Probability paths change the velocity field learned by flow matching. At a fixed number of function evaluations (NFE), does lower averaged Jacobian regularity reliably imply lower numerical sampling error?
Method. The project uses commuting Gaussian probability-flow ODEs, where the dynamics and Gaussian \(W_2\) endpoint error are available exactly. It pairs a closed-form scalar construction with a frozen 36-comparison comparison of linear and trigonometric variance-preserving paths under equal-NFE Euler, Heun, and RK4 solvers; selected settings also compare four candidate paths.
Main result. For the transport \(\mathcal{N}(0,1)\rightarrow\mathcal{N}(0,4)\), averaged regularity prefers a trigonometric variance-preserving path, while Heun at NFE 8 gives lower \(W_2\) error to the linear path. The frozen grid has 12 geometry–solver cells, each evaluated at NFE 8, 16, and 32. Five cells contain ranking inversions, spanning 14 of the 36 paired path comparisons. This is evidence that averaged regularity is not, by itself, a reliable ordering surrogate under finite discretization.
Limits. The closed-form claims concern exact Gaussian systems, not learned neural velocity fields. The 14-of-36 count describes one finite registered grid, not a population failure rate or a general theorem about flow matching.
Status. Public research manuscript and reproducible software. No venue acceptance or public archival publication is claimed here.
Ego-World JEPA
Question. Does factorizing an action-conditioned latent world model into object-relevant components improve planning with a shared detector and MPPI controller on PushT?
Method. Eight configurations vary latent factorization, auxiliary supervision, and related modeling choices. The main factored and monolithic pair was evaluated with 50 planning episodes per seed, 512 MPPI samples, six optimization iterations, and an eight-step horizon. Both use the same detector and controller.
Main result. On the 96 px g1/g2 checkpoints at seed 0, an in-slice ridge probe of absolute block pose reached \(R^2=0.997\) for the factored model and \(R^2=0.992\) for the monolithic model, with 6/50 and 2/50 planning successes. A displacement probe on the same checkpoints reached \(R^2=0.205\) and \(R^2=0.240\), reversing the probe ordering. On seed 1, both reached 2/50 planning successes. The seed-0 planning difference was not significant under the reported Fisher test (\(p=0.269\)), and the apparent factorization advantage did not replicate. High probe decodability was therefore insufficient evidence of a planning-useful representation in this setup.
Earlier 64 px checkpoints reached absolute-pose probe \(R^2=0.286\) for the factored model and \(R^2=0.779\) for the monolithic model. The configurations differed from g1/g2, including auxiliary supervision.
Limits. The primary comparison has two seeds, the probe is in-slice and partly aligned with auxiliary supervision, and success rates remain low across the grid. The experiment covers one PushT setup and does not establish a broad claim about JEPA representations or model-predictive control.
Status. Completed and frozen controlled project; the central outcome is a negative result.
LPN / Adaptive Latent Search
Question. During latent optimization, can a learned stop-or-continue policy spend additional steps only where they are useful?
Method. This is an experimental extension of Bonnet and Macfarlane’s LPN codebase. In the public pattern_2d sandbox, policies train on two checkpoints and are evaluated on a held-out third checkpoint. Fixed-step rules, hand-built stopping heuristics, a learned controller, and a non-deployable oracle are compared over the same five-step horizon.
Main result. The learned zero-penalty policy reached 0.90625 exact match at 4.281 mean steps. A gradient-norm rule reached the same exact match at 3.438 steps; the oracle reached it at 0.562 mean steps. The oracle gap shows that adaptive allocation could matter, but the learned policy did not outperform the strongest simple heuristic. The experiment motivates better stopping observations and estimators rather than a performance claim.
Limits. These results come from a small in-family pattern_2d task, one held-out checkpoint, 96 correlated leave-one-out episodes, and limited policy seed analysis. They are not ARC-AGI results and do not establish out-of-domain generalization.
Status. Exploratory extension and documented negative result.
Secondary projects
Learned niche selection for sycophancy probing
An adaptation of ACES to a grounded 12-cell sycophancy-probing domain. Over two seeds and 100 generations, plain Bayesian factored-ALP selection trailed uniform sampling in mean QD-score (9.219 versus 10.750). An \(\varepsilon=0.2\) exploration floor recovered mean coverage to 0.917 and mean QD-score to 11.000, although one seed retained substantial cell concentration. With only two seeds, this is a diagnostic case study, not an inferential or publication claim.
Proximal Diffusion Models in PyTorch
A from-scratch course reproduction of the proximal diffusion framework on MNIST and small two-dimensional distributions, including score-based and proximal samplers, tests, configuration, and an FID-versus-NFE evaluation pipeline. This is an implementation of published work, not an original paper.
Decision Transformer for limit-order-book trading
A reproducible offline-RL benchmark co-developed with Côme Genet on FI-2010. Across five seeds, the Decision Transformer improved over behavior cloning but did not beat buy-and-hold under the primary transaction-cost setting.
The reported units are normalized feature-space PnL, and the execution model ignores spread, latency, queue position, and market impact.