Kanpat Vesessook, Brown CS '28

02 / Research

Three working toys.

Each card runs the actual idea live in your browser. Drag the sliders.

01 ICLR 2026 · Workshop on World Models · co-first author

Temporal Reversal Asymmetry: does a video model know which way time runs?

A training-free probe for world models. Compare prediction loss on forward vs. time-reversed videos - a model that understands physics should find irreversible processes (energy dissipating) asymmetric, and reversible ones (elastic collisions) symmetric. We tested four video models on 380 physics simulations. V-JEPA 2 tracked the thermodynamic arrow of time, peaking at intermediate dissipation, near-zero at the elastic limit.

V-JEPA 2VideoMAE V2MVD · Hieratraining-free

Read on OpenReview ↗

live · physics simulator → TRA

Restitution e 0.50

Damping γ 1.50

Forward loss

0.00

Reverse loss

0.00

TRA

+0.00

0+arrow−arrow

02 Working paper · co-author · in progress

Feature speed, transaction costs, and the limits of ML in emerging markets.

Across 15 EMs and 10+ portfolio strategies, net Sharpe converges on the same ceiling (~0.47–0.49 at 3× transaction costs). We derive a cost-speed bound: SR_gross − SR_net ≈ λκ · FSI, where the Feature Speed Index measures a strategy's ex-ante exposure to fast-moving characteristics. The bound predicts SR degradation with R²=0.72 across 300 strategy–market–cost combinations.

Cost regime κ 3.0×

Net SR ceiling0.48

λ sensitivity0.33

R² of bound0.72

Drag κ to 0, costs vanish, ML's prediction edge returns. Drag to 3× (EM reality) and every strategy collapses onto the same floor. Smoothness, not predictive power, is the binding constraint.

03 PACCON 2024 · third author · Chulalongkorn

COVID-19 from volatile organic compounds.

ML pipeline turning surface swab + breath samples into diagnoses. Eight classifiers (logistic regression, SVM, random forest, gradient boosting) tuned and cross-validated.

90%+best accuracy

8classifiers compared

2023Bangkok lab

scikit-learnbiomarker MLPACCON 2024

Kanpat
Vesessook.

A Bangkok kid
on College Hill.

Three working toys.

Temporal Reversal Asymmetry: does a video model know which way time runs?

Feature speed, transaction costs, and the limits of ML in emerging markets.

COVID-19 from volatile organic compounds.

On the record.

14,000 km, three years.

Where I spend hours.

Kanpat Vesessook.

A Bangkok kidon College Hill.

Three working toys.

Temporal Reversal Asymmetry: does a video model know which way time runs?

Feature speed, transaction costs, and the limits of ML in emerging markets.

COVID-19 from volatile organic compounds.

On the record.

14,000 km, three years.

Where I spend hours.

Kanpat
Vesessook.

A Bangkok kid
on College Hill.