EVL
SWE-Gym
SWE training environment
What this measures
A training *environment* (not just a test) for software-engineering agents — real repos where an agent can attempt fixes and get feedback, used to teach models to code like engineers.
Sample
An executable repo + failing test where the agent iterates: edit code → run tests → see results → try again.
Leaderboard · 0 models
Resolve rate after interactive attempts — higher is betterNo scores recorded yet.
What improves this score
Datasets & environments shown to raise it.
No data yet.
Research using this benchmark
all →From Patches to Trajectories: Privileged Process Supervision for Software-Engineering AgentsMemGym: a Long-Horizon Memory Environment for LLM AgentsRevisiting DAgger in the Era of LLM-AgentsAn Executable Benchmarking Suite for Tool-Using AgentsBoostAPR: Boosting Automated Program Repair via Execution-Grounded Reinforcement Learning with Dual Reward Models
Matched by name in title/abstract.