Vinod Anbalagan

● ML engineer · independent researcher · Toronto

Vinod Anbalagan

I study what representations a learner forms when you constrain it, using probes and causal interventions rather than benchmark scores.

Research →The LabCV
ball straight-line guess while hidden collision behind the wall

When two balls collide out of sight, no extrapolation can know. Reverie asks whether a learned world model keeps the hidden ball in its state anyway, and whether it uses it.

Now

In progress

Reverie

Does a world model represent what it cannot see? Probing and patching an RSSM while a ball is hidden behind an occluder.

Research · world modelsMRL 2026

CulturalRiddles

First author on a living multicultural riddles benchmark: 51 languages, 61+ communities, with the Cohere Labs Open Science Community.

Research · evaluationLab 01 · up next

Search Landscapes

Gradient descent, CEM and evolution strategies: does their ranking flip as dimension grows?

One of eight small, live, measured experiments

Writing · The Meta Gradient

Building and breaking AI systems to understand them.

Visual intuition first, formal version second, and the philosophical bits left in.

All writing →
  1. 18 Sept 2026Freezing the NoiseHow one line of algebra made sampling differentiable.
  2. 14 Sept 2026The Architecture of DreamsPart 1: how neural networks learned to imagine. World Models, PlaNet, Dreamer.
  3. 1 Aug 2026If AI Will Shape the World, the World Must Help Govern ItOpen weights, legitimacy, and who gets to participate in science.

Opportunities

Open to research engineer, research assistant, and research internship opportunities, as well as funded MS/PhD positions, in world models, representation learning, multimodal AI, robotics, and scientific ML.

Message on LinkedIn →EmailView CV