hallerite / Konstantin Dunas

I’m a researcher at Prime Intellect, where I work on RL across the stack. These days, I’m mostly thinking about and working on multi-agent systems.

Much of my work is open source, including:

• the multi-agent abstraction in verifiers and prime-rl, for defining how agents interact and learn, from self-play to agentic judging.
• the algorithms layer in prime-rl, for defining training algorithms and mixing them across environments.
• renderers, a Python library for message–token conversion that preserves sampled tokens across turns.

Previously, I built Ludic, an exploration of the RL abstractions that have informed my subsequent work.