Related Work & Provenance
Repo2RLEnv builds on a growing body of research into turning real software repositories into executable, verifiable environments for training and evaluating code agents. This page records two things:
- Provenance: exactly what each shipped pipeline draws inspiration from
(these mirror the
Acknowledgmentblocks in the pipeline source files). - Related work & other implementations: adjacent papers, datasets and frameworks worth knowing, including recent model lines from Microsoft and NVIDIA that consume verifiable code-RL environments.
Original Repo2RLEnv code is Apache-2.0. Some owned recipes retain upstream code,
prompts or examples under MIT or Apache-2.0; their provenance.md files describe
the exact scope and adaptations. The
third-party index
lists the bundled licenses and notices. SWE-RL is a non-permissive influence
(CC BY-NC 4.0); our reward.py independently implements the concept and records
that distinction. Referenced research, bundled material and input data have
separate licensing boundaries.
Provenance: what each pipeline draws from
| Pipeline | Inspiration | Paper | Repo (license) |
|---|---|---|---|
pr_diff | SWE-RL | SWE-RL (Wei et al., NeurIPS '25), arXiv:2502.18449 | facebookresearch/swe-rl (CC BY-NC 4.0) · SWE-bench/SWE-bench (MIT) |
pr_runtime | SWE-bench + SWE-bench-Live | SWE-bench (Jimenez et al., ICLR '24), arXiv:2310.06770; SWE-bench-Live (Zhang et al., NeurIPS '25), arXiv:2505.23419 | SWE-bench/SWE-bench · microsoft/SWE-bench-Live (both MIT) |
commit_runtime | R2E-Gym ("SWE-GEN" commit mining) | R2E-Gym (Jain et al., COLM '25), arXiv:2504.07164 | R2E-Gym/R2E-Gym (MIT) |
cve_patches | PatchSeeker + CVE-Bench + OSV | PatchSeeker (Le et al.); CVE-Bench (Zhu et al., NAACL '25) | hungkien05/PatchSeeker (MIT) · OSV API |
mutation_bugs | SWE-smith | SWE-smith (Yang et al., NeurIPS '25 Spotlight), arXiv:2504.21798 | SWE-bench/SWE-smith (MIT) |
code_instruct | Magicoder / OSS-Instruct | Magicoder (Wei et al., ICML '24), arXiv:2312.02120 | ise-uiuc/magicoder (MIT) |
equivalence_tests | R2E | R2E (Jain et al., ICML '24) | r2e-project/r2e (MIT) |
refactor_synthesis | RefactoringMiner (originally planned, dropped in v0.8) | None (the recipe is original) | tsantalis/RefactoringMiner (MIT) |
Shared components (not pipelines)
| Component | Inspiration | Link |
|---|---|---|
reward.py (diff-similarity) | SWE-RL | arXiv:2502.18449 · facebookresearch/swe-rl. Concept only: their reward lib is CC BY-NC 4.0, our reimplementation is Apache-2.0 |
bootstrap/ (LLM Docker env gen) | RepoLaunch | microsoft/RepoLaunch (MIT) · arXiv:2603.05026 |
_oss_instruct.py | Magicoder | ise-uiuc/magicoder (MIT) |
_function_extractor.py | R2E (fut_extractor) | r2e-project/r2e (MIT) |
_mutation_operators.py | SWE-smith | SWE-bench/SWE-smith (MIT) |
Related work & other implementations
Adjacent work that isn't cited above, grouped by the technique it overlaps with. It's useful if you're deciding whether to extend a pipeline, swap a reward or feed our datasets into a trainer.
RL-environment frameworks & training stacks (the "consumption" layer)
- NeMo Gym (NVIDIA): NVIDIA-NeMo/Gym (Apache-2.0). 100+ RLVR environments for LLMs, including Mini-SWE-Agent / OpenHands SWE agents and a Harbor harness. It's the closest analogue to Repo2RLEnv's goal and shares the Harbor ecosystem.
- SkyRL (UC Berkeley NovaSky): NovaSky-AI/SkyRL (Apache-2.0) · SkyRL-Agent arXiv:2511.16108. A full-stack RL training framework for long-horizon, SWE-bench-style agentic tasks. It references Harbor.
- verifiers (Prime Intellect): PrimeIntellect-ai/verifiers. Defines "environment = dataset + harness + rubric/reward", the closest competing spec to our task+verifier emission.
- OpenEnv (Meta PyTorch + Hugging Face): meta-pytorch/OpenEnv. A Gymnasium-style
step()/reset()/state()standard for containerized agentic environments, with a shared Hub. It's a standardization effort alongside our Harbor emission + HF Hub bridge. - rLLM (Agentica): rllm-org/rllm. "Run agent → collect traces → reward → update" post-training stack; used to train DeepSWE on R2E-Gym environments.
Automated environment setup / dockerization (the bootstrap/ problem)
- RepoLaunch (Microsoft): arXiv:2603.05026 · microsoft/SWE-bench-Live (MIT). An LLM agent for polyglot, multi-OS automated build/test env setup, and the canonical citation for our bootstrap layer.
- Repo2Run: arXiv:2502.13681. Automated building of executable environments for code repositories at scale. Overlaps directly with
ensure_bootstrap. - EnvBench (JetBrains Research, ICLR '25 DL4Code): arXiv:2503.14443. 994-repo benchmark for automated env configuration with compilation/import success checks.
- SetupBench (Microsoft): arXiv:2507.09063 · microsoft/SetupBench. 93 tasks isolating the "bootstrap a dev env from a bare OS" skill across 7 ecosystems.
SWE training datasets & RL recipes
- SWE-Gym (Pan et al., ICML '25): arXiv:2412.21139 · SWE-Gym/SWE-Gym. 2,438 real Python tasks with executable runtimes + tests for training agents/verifiers.
- SWE-rebench (Nebius): arXiv:2505.20411 · nebius/SWE-rebench. Continuous automated extraction of 21k+ interactive Python tasks from live repos with contamination tracking. It's the closest large-scale mining-to-RL pipeline to ours.
- Agent-RLVR: arXiv:2506.11425. Training SWE agents via guidance + environment rewards; relevant to our verifier-graded reward design.
- DeepSWE (Agentica + Together): agentica-org/DeepSWE-Preview. RL-only coding agent (Qwen3-32B) trained on ~4,500 R2E-Gym tasks; a reference end-to-end consumer of mined repo environments.
- Kimi-Dev (Moonshot AI): arXiv:2509.23045. Agentless skill-prior training + outcome-driven RL on issue resolution.
Synthetic task / test synthesis
- SWE-Flow (Alibaba Qwen, ICML '25): arXiv:2506.09003 · Hambaobao/SWE-Flow. Builds a Runtime Dependency Graph from unit tests to synthesize verifiable TDD tasks; overlaps
equivalence_testsand test-anchored synthesis. - SWE-Mirror: arXiv:2509.08724. Mirrors real issues across other repos to produce 60K verifiable tasks in 4 languages, kept only if test-status transitions validate.
- FEA-Bench (PKU + MSR Asia, ACL '25): arXiv:2503.06680 · microsoft/FEA-Bench. PR-mined, intent-filtered feature-implementation tasks with unit tests. Overlaps our PR mining, structural filters and test verifier.
- OpenCodeInstruct (NVIDIA): arXiv:2504.04030 · nvidia/OpenCodeInstruct. 5M instruction samples with test cases + execution feedback, close in spirit to
code_instruct. - Genetic-Instruct (NVIDIA): arXiv:2407.21077. Evolutionary (mutation/crossover) scaling of synthetic coding instructions; comparable to
mutation_bugs' procedural generation.
Model lines trained with verifiable code RL (downstream consumers)
This is where datasets like ours end up: model families whose post-training includes RL with verifiable code rewards.
- NVIDIA Nemotron: Nemotron-Cascade arXiv:2512.13607, PivotRL arXiv:2603.21383; SWE RL datasets nvidia/Nemotron-Cascade-RL-SWE and nvidia/Nemotron-RL-Agentic-SWE-Pivot-v1. Multi-environment RLVR via NeMo Gym for agentic coding. The contrast with us: Nemotron-Cascade uses an execution-free reward model, where we verify by execution (test pass rate).
- NVIDIA AceReason-Nemotron: arXiv:2505.16400, v1.1 arXiv:2506.13284 · nvidia/AceReason-Nemotron-14B. Large-scale GRPO RL with test-case-backed code prompts; a core RLVR-on-code reference.
- NVIDIA OpenCodeReasoning (COLM '25): arXiv:2504.01943 · nvidia/OpenCodeReasoning. An SFT-only reasoning dataset that matches RL-trained models on LiveCodeBench, which makes it a distillation-vs-RL counterpoint.
- NVIDIA Llama-Nemotron: arXiv:2505.00949 · NVIDIA/NeMo-Skills (Apache-2.0, the synthesis+training infra behind these). Post-training with large-scale verifiable-reward RL.
- Microsoft Phi-4-reasoning: arXiv:2504.21318 (and Phi-4-Mini-Reasoning arXiv:2504.21233). Outcome-based RL with verifiable rewards; coding among eval domains.
- Microsoft rStar-Math: arXiv:2501.04519 · microsoft/rStar (MIT). Execution-verified, code-augmented reasoning with a process reward model.
Last refreshed: 2026-06. Spotted something missing or mis-cited? PRs to this page welcome.