Research map / Multi-agent systems and games / LLM agents

Hybrid-Gym: Training Coding Agents to Generalize Across Tasks

ICML 2026 · Multi-agent systems and games › LLM agents · approach: Software engineering agent benchmarks · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes Hybrid-Gym, a set of scalable synthetic coding-agent training tasks (function/issue localization, dependency search, function generation) that need no executable repo setup, and shows they transfer to SWE-Bench, SWT-Bench, and Commit-0 when used to fine-tune Qwen2.5Coder models.

Problem

Coding agents are trained and evaluated mostly on single-issue GitHub resolution, which does not transfer reliably to other real-world coding tasks such as test generation or library construction; building executable repository environments for training is costly and does not scale.

Approach

Derive design principles (output format matching, repo exploration, non-trivial reasoning, no complex environment setup) from an analysis of agent trajectory components and commands, then instantiate four synthetic tasks that satisfy them and train via rejection-sampling distillation from teacher trajectories.

Key results

• Hybrid-Gym improves Qwen2.5-Coder-32B by 25.4 points on SWE-Bench Verified (7.0% to 32.4%) without training on downstream tasks.
• Hybrid-Gym raises Qwen2.5-Coder-32B by 7.85 points on SWT-Bench Verified and 5.11 points on Commit-0 Lite.
• Adding Hybrid-Gym to SWE-Play yields 20.29 / 21.02 / 15.52 on SWT-Lite / SWT-Verified / Commit-0 for 32B, improving SWE-Play's SWT-Verified by 4.85 points at 7B.

Surprising findings

• Script-level code generation (LiveCodeBench) does not transfer to repo-level issue solving, even when wrapped in a repository harness.
• Training on trajectories from the same repositories used in evaluation does not improve performance, while repository diversity helps.
• o3-mini trajectories that separate reasoning and tool calls into different turns degrade distillation, and stitching them fixes transfer.

Limitations

• Synthetic tasks are Python-only and built from SWE-Gym-Raw repositories.
• Evaluation of the 7B/32B results relies on specific teacher models (Claude, Qwen3) whose trajectories may be distinctive.
• Ablations are run on an easier 50-instance SWE-Bench subset, limiting statistical power.

Experimental setup

Datasets: SWE-Gym-Raw RepoST SWE-Gym SWE-Play

Benchmarks: SWE-Bench Verified SWT-Bench Lite SWT-Bench Verified Commit-0 Lite

Metrics: resolved rate localized rate non-loop rate

Baselines: SWE-Gym R2E-Gym SWE-smith SWE-Play

Builds on: SWE-Gym RepoST OpenHands SWE-smith SWE-Play

Closest papers by meaning

• Training Software Engineering Agents and Verifiers with SWE-Gym (ICML 2025 · similarity 0.90)
• Generalizable End-to-End Tool-Use RL with Synthetic CodeGym (ICLR 2026 · similarity 0.89)
• Self-Challenging Language Model Agents (NeurIPS 2025 · similarity 0.88)
• Kimi-Dev: Agentless Training as Skill Prior for SWE-agents (ICLR 2026 · similarity 0.88)
• Frontier Coding Agents Use Metaprogramming to Adapt to Unfamiliar Programming Languages (NeurIPS 2026 · similarity 0.88)
• Agent Skill Acquisition for Large Language Models via CycleQD (ICLR 2025 · similarity 0.86)