Deep Pearl AI

Research

The research agenda

The lines of work the lab is studying and building, from how machines perceive the 3D world to how they reason, act, and scale. Everything here feeds back into Multivac, Multivac Physics, ALEX-1, and EMOTE4D.

FIG. 5 — STRATA The lab's record as accumulated layers.

Currently learning

  • representation learning
  • interpretability
  • adaptive computation
  • edge deployment
  • robotic integration
  • perception and 3D

Perception and 3D

How machines recover and edit the structure of the physical world.

  • 3D reconstruction — single image to 3D, stereo vision
  • NeRF plus diffusion for 3D generation and scene editing
  • Articulated human and garment modeling
  • Gaussian splatting

Representation and learning

What models learn, and whether the brain offers a better recipe.

  • Self-supervised learning — JEPA-style predictive learning
  • Vision Transformers, DINOv2
  • Brain-to-model alignment — retinal neural coding, brain-optimized networks
  • Non-backprop and brain-inspired learning — Hebbian, predictive coding, forward-forward, equilibrium propagation
  • KAN-style minimal networks

Reasoning and symbol

Where pattern-matching ends and genuine reasoning begins.

  • Mathematical cognition — symbolic math, symbolic regression
  • Mechanistic interpretability — math transformer circuits
  • Neural-symbolic hybrid systems
  • Word models to world models — the probabilistic language of thought line

Agents and control

From tool use to embodied skill — how systems decide and act.

  • Tool-using and agentic LLMs — ReAct, Toolformer, Voyager, generative agents
  • Task decomposition — hierarchical planning, HTN
  • Hierarchical RL — options, skill discovery
  • Generalist and embodied robotic agents — Gato, RoboCat

Scaling

Doing more with the compute you have, and measuring what the compute bought.

  • Mixture of experts — dynamic routing, sparsity
  • Scientific reasoning evaluation — SciBench

Applied, not abstract. Evaluation findings shape what Multivac and Multivac Physics test next. Representation and edge work flows into ALEX-1 and EMOTE4D. The research agenda exists to make the ventures sharper — and the ventures keep the research honest.