KairosJetBrains Kairos

Machine Learning

Core ML research that powers verification, generation, and reasoning in our tools — papers, benchmarks, and the teams behind them.

What we focus on

Code world models

Models that emulate a repository as a running system rather than a pile of files — predicting failing tests, runtime errors and risky changes before anything executes, giving an agent a structured view it can query and simulate against, and producing the execution-grounded data, benchmarks and reward models that judge a change without running the whole environment.

Multi-task learning

Training on the breadth of real developer workflows instead of one task at a time. We look for the software engineering tasks that ML training under-represents, build recipes that learn from many of them at once, and measure coding capability in ways that survive contact with real work — which is what our own Mellum-series and agentic models are improved against.

Federated and private compute

Training on real developer data without every team becoming a privacy expert: a differential-privacy training stack verified end-to-end on shipping features, memorization auditing and privacy attacks that quantify what leaks when it is left out, and on-device federated computation for the cases where the data should never move at all.

EnvBench: A Benchmark for Automated Environment Setup
By Aleksandra Eliseeva
A well-maintained project should be straightforward to set up, however, in practice, it is not always the case.
Read the article

Projects

Machine Learning

IdeGYM

An open-source orchestrator for scalable, disposable development environments — built for training reinforcement-learning models and running AI agents.

Explore project
IdeGYM
Machine Learning

Federated compute

Privacy-sensitive training on real-world code — differential privacy, memorization auditing, and on-device federated compute.

Explore project
Federated compute

Selected publications

Aug 2025
StudyMachine Learning

The Complexity Trap: Simple Observation Masking Is as Efficient as LLM Summarization for Agent Context Management

A study of context-management strategies for coding agents — context compression and observation masking over long horizons.

Apr 2025
BenchmarkMachine Learning

Drawing Pandas: A Benchmark for LLMs in Generating Plotting Code

175 human-curated data-analysis tasks that score plotting code by the picture it draws, judged by a vision model against the ground-truth plot.

Mar 2025
BenchmarkMachine Learning

EnvBench: A Benchmark for Automated Environment Setup

Hard automated development-environment-setup cases across Python and JVM repositories, with dataset and code.

Other directions