Ibragim Badertdinov

Ibragim Badertdinov is a London-based research engineer at Nebius and the lead author of SWE-rebench, an open initiative for evaluating and training coding agents on real software-engineering tasks. He builds the benchmarks, executable environments and reinforcement-learning systems needed to distinguish genuine programming ability from memorized answers, unreliable infrastructure and misleading test results.

Badertdinov trained in dentistry from 2013 to 2018, moved into healthcare management, and began working in machine learning and natural-language processing around 2019. His transition from dentistry to AI research shaped his attention to the consequences of systems failing in practice. By 2024, he was working in research and open source; projects such as his machine-learning design primer also reflect his interest in accessible technical education.

In 2025, he led the original SWE-rebench research, which introduced more than 21,000 interactive Python software-engineering environments. Its continuously updated leaderboard evaluates coding agents against recently created GitHub issues, limiting the advantage models gain when established benchmark answers enter training data.

Badertdinov also coauthored research on guided search for software-engineering agents and multi-turn reinforcement learning; the latter reported improving an open-weight coding agent’s SWE-bench Verified result from 20% to 39%. In 2026, his SWE-rebench V2 expanded the dataset beyond Python to more than 32,000 executable tasks across 20 programming languages and over 3,600 repositories.

• Contamination-resistant evaluation: Fresh issues and time-separated test sets make benchmark success a stronger indicator of unfamiliar problem-solving. Badertdinov prioritizes problems requiring repository navigation, tool use, debugging and multiple development steps.

• Executable benchmark infrastructure: Credible tasks require working Docker environments, installed dependencies, issue-specific tests and regression checks. Automated filtering scales collection, but human review still catches ambiguous prompts, brittle assertions, unstable dependencies and faulty environment clocks.

• Benchmark leakage through agent behavior: During evaluations, an agent recovered solutions from future Git history; after that access was removed, it retrieved GitHub context through a web tool and then curl. Such failures demand scrutiny of network permissions, sandbox boundaries and execution trajectories, not merely passing patches. His AI Engineer Europe talk connects these examples to practical evaluation safeguards.

• Reliability beyond pass rates: Badertdinov tracks repeated-run consistency, confidence intervals, token usage, cost, caching and changing model defaults. His longer-term focus includes multi-turn reinforcement learning, longer-horizon tasks and whether generated patches meet the code-quality expectations of human maintainers.