Dr. Avinash Kori
HUMAN-CENTRED / CONTESTABLE AI
AI people can understand,
question and intervene on.
I develop identifiable and compositional representations, causal and contestable reasoning, and human–AI decision systems that can be challenged and corrected under uncertainty.
Research Associate, Imperial College London · Co-founder & Chief Scientist, Blurgs AI
Selected contributions
01 / AI ALIGNMENT
Representations should mean something stable before we align behaviour.
I study grounded, identifiable and compositional object-centric representations: what a model has actually recovered, which factors remain stable, and when those representations can support reliable reasoning rather than arbitrary latent shortcuts.
Identifiability · NeurIPS 2024Spatial ambiguities · ICML 2025
02 / HUMAN-CENTRED & CONTESTABLE AI
People should be able to question a model and see what changes.
My work connects counterfactuals, causal reasoning and argumentative model–human exchanges: supporting why not?, what if?, and what should change? The goal is not only to explain an output, but to let people challenge assumptions, provide evidence and inspect whether the reasoning revises consistently.
Counterfactuals · ICML 2025Causal discovery · ICML 2025Argumentative exchanges · AAMAS 2025
03 / SUSTAINABILITY
Test trustworthy AI where uncertainty has physical consequences.
Fisheries and maritime systems are my main real-world testbed for deep learning, world models and causal decision support. I work on electronic monitoring, species and ETP evidence, vessel behaviour, dark shipping, port congestion and intervention-aware maritime intelligence.
Maritime & fisheries articlesFishnet → MeenuCausal port-call optimisation
Latest
A causal perspective on world models
Our perspective with Fabrizio Russo connects observations, representations and causal structure; the manuscript reports acceptance at the UAI 2026 Causality in Decision Making workshop. Paper and status
Blurgs AI raises $2.2 million
Funding led by Pravega Ventures and Shastra VC supports the next stage of our maritime intelligence work. Business Standard
Steering diffusion representation learning
A preprint with Rajat Rasal, Tian Xia and Ben Glocker examines how optimisation trajectories shape learned representations. Preprint