Research programs

A program organized
around the loop.

From algorithms to experiments, our work asks how recursive systems can improve models, search, and eventually physical understanding.

Status languageActiveIn ProgressExploratoryPlanned

Evaluation

How we measure progress.

Success is not visual novelty. It is a better model, a more useful decision, or less wasted computation under a fair comparison.

EfficiencyFewer

evaluations required to reach a defined target

Model qualityLower

prediction error with uncertainty reported

ReliabilityBetter

robustness under noise and model mismatch

North starMore

useful physical knowledge per experiment

Comparison standard

SIML competes with baselines.

Random search, Bayesian optimization, conventional active learning, and domain-specific optimizers are reference methods—not straw men. Budgets, inputs, and evaluation criteria should remain comparable.

Read the evaluation approach

Real project evidence

One published computational benchmark.

These figures and values come from the existing FEPGate + V-JEPA 2 simulation report. They are not physical materials results and should not be generalized beyond that reported setup.

Benchmark100 simulated episodes · frozen planner · same CEM/MPC budget

SIML FEPGate + V-JEPA 2

A one-bit surprise gate was added to the planner without retraining. The public research artifact includes the method, implementation link, and stated setup.

0.193 → 0.109 mReported final error2.28 → 1.13 sReported step latency0.058 → 0.029 WhReported energy / episode
Inspect the report
Reported final error distribution from the FEPGate simulation benchmark
final error distribution
Reported per-step latency from the FEPGate simulation benchmark
per-step latency
Reported energy per episode from the FEPGate simulation benchmark
energy per episode