Minus AI — Research Lab
A research lab for reinforcement learning and interactive visualization.
We study how agents learn from interaction — and build tools that make those dynamics legible.
AGENT
policy π
ACTION · a_t
STATE · s_t+1 REWARD · r_t
ENVIRONMENT
dynamics + reward
Our Approach
Three threads of research, one feedback loop.
We study how agents learn, build tools to see what they learn, and publish so others can build on it.
Reinforcement Learning
Agents that learn from interaction — exploration, credit assignment, and stable policy optimization.
Interactive Visualization
Tools that make training dynamics, value landscapes, and agent behavior legible in real time.
Open Research
Methods, tooling, and findings shared openly, so the community can build on our work from day zero.
News
Exhibition
July 2026
Presenting ReCollection at ZKM
ReCollection — our interactive installation on how reinforcement-learning agents form, forget, and revisit memories — is on view at ZKM Center for Art and Media, Karlsruhe, turning learned representations into a space you can walk through and explore.
View at ZKM →
Paper
October 2026
Iterative Perceptual Alignment for VLMs via Deterministic Reconstruction Feedback
A self-supervised framework for vision-language alignment that replaces costly human preference annotation with deterministic reconstruction feedback. The reward model reaches competitive zero-shot accuracy using only 20% of the usual training data, and DPO fine-tuning on it sharpens VLMs on complex, high-density descriptions. Accepted to ECCV 2026.
Join Us
Help us build learning agents.
We’re a small team of researchers and engineers who love hard problems at the intersection of reinforcement learning and visualization.