{"id":51879676,"url":"https://github.com/19pine-ai/rlvp","last_synced_at":"2026-07-25T11:01:26.293Z","repository":{"id":370389524,"uuid":"1267175709","full_name":"19PINE-AI/rlvp","owner":"19PINE-AI","description":"Penalize the Path, Reward the Outcome — verifiable per-action penalties as a dense channel for deployable, sample-efficient agentic RL (GRPO). Paper: arXiv:2607.07435","archived":false,"fork":false,"pushed_at":"2026-07-09T02:20:07.000Z","size":16329,"stargazers_count":1,"open_issues_count":0,"forks_count":0,"subscribers_count":0,"default_branch":"main","last_synced_at":"2026-07-09T04:13:53.509Z","etag":null,"topics":["agentic-ai","grpo","llm","machine-learning","process-reward","reinforcement-learning","rlvr","verifiable-rewards"],"latest_commit_sha":null,"homepage":"https://arxiv.org/abs/2607.07435","language":"Python","has_issues":true,"has_wiki":null,"has_pages":null,"mirror_url":null,"source_name":null,"license":null,"status":null,"scm":"git","pull_requests_enabled":true,"icon_url":"https://github.com/19PINE-AI.png","metadata":{"files":{"readme":"README.md","changelog":null,"contributing":null,"funding":null,"license":null,"code_of_conduct":null,"threat_model":null,"audit":null,"citation":null,"codeowners":null,"security":null,"support":null,"governance":null,"roadmap":null,"authors":null,"dei":null,"publiccode":null,"codemeta":null,"zenodo":null,"notice":null,"maintainers":null,"copyright":null,"agents":null,"dco":null,"cla":null}},"created_at":"2026-06-12T09:42:18.000Z","updated_at":"2026-07-09T02:28:04.000Z","dependencies_parsed_at":null,"dependency_job_id":null,"html_url":"https://github.com/19PINE-AI/rlvp","commit_stats":null,"previous_names":["19pine-ai/rlvp"],"tags_count":null,"template":false,"template_full_name":null,"purl":"pkg:github/19PINE-AI/rlvp","repository_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repositories/19PINE-AI%2Frlvp","tags_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repositories/19PINE-AI%2Frlvp/tags","releases_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repositories/19PINE-AI%2Frlvp/releases","manifests_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repositories/19PINE-AI%2Frlvp/manifests","owner_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/owners/19PINE-AI","download_url":"https://codeload.github.com/19PINE-AI/rlvp/tar.gz/refs/heads/main","sbom_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repositories/19PINE-AI%2Frlvp/sbom","scorecard":null,"host":{"name":"GitHub","url":"https://github.com","kind":"github","repositories_count":286080680,"owners_count":35877013,"icon_url":"https://github.com/github.png","version":null,"created_at":"2022-05-30T11:31:42.601Z","updated_at":"2026-07-20T02:08:10.276Z","status":"online","status_checked_at":"2026-07-25T02:00:06.922Z","response_time":64,"last_error":null,"robots_txt_status":"success","robots_txt_updated_at":"2025-07-24T06:49:26.215Z","robots_txt_url":"https://github.com/robots.txt","online":true,"can_crawl_api":true,"host_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub","repositories_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repositories","repository_names_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repository_names","owners_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/owners"}},"keywords":["agentic-ai","grpo","llm","machine-learning","process-reward","reinforcement-learning","rlvr","verifiable-rewards"],"created_at":"2026-07-25T11:01:25.679Z","updated_at":"2026-07-25T11:01:26.287Z","avatar_url":"https://github.com/19PINE-AI.png","language":"Python","funding_links":[],"categories":[],"sub_categories":[],"readme":"# RLVP: Reinforcement Learning from Verifiable Penalties\n\n[![Paper](https://img.shields.io/badge/arXiv-2607.07435-b31b1b.svg)](https://arxiv.org/abs/2607.07435)\n[![Website](https://img.shields.io/badge/website-01.me%2Fresearch%2Frlvp-1f6feb.svg)](https://01.me/research/rlvp)\n[![License: MIT](https://img.shields.io/badge/License-MIT-green.svg)](LICENSE)\n\nOfficial code for **[Penalize the Path, Reward the Outcome](https://arxiv.org/abs/2607.07435)**\n(arXiv:2607.07435) — Bojie Li (Pine AI), Noah Shi (University of Washington).\n\nDense, rule-derived process rewards for agentic RL, attached at the tool-call\nboundary. **Premise (verifier asymmetry):** for long-horizon agents,\nintermediate actions are hard to verify as *right* but easy to verify as\n*wrong* — so use penalty-only deterministic rules (plus obligation-fulfillment\ncredits) as the dense channel, keep the sparse outcome reward as the only task\nsignal, and train R1-style GRPO with step-aware credit assignment. The headline\nmetric is pass^k / perfect^k (reliability), not pass@1.\n\n## Result in one line\n\nOutcome-only RL reaches 100% task success while violating outcome-neutral\nworkflow rules in 240/240 episodes; the hybrid credit recipe (`c3`) reaches\n**perfect^8 = 1.00** — all 8/8 trials succeed *and* comply on held-out tasks —\nwith **no rules in the prompt**, beating the base model *with* rules in its\nprompt. See [`results/REPORT.md`](results/REPORT.md) for the full writeup and\nthe four mechanisms (per-channel normalization, fulfillment credits, the\nexploration wall, the compliance-only attractor).\n\n## Repository layout\n\n| Path | Contents |\n| --- | --- |\n| [`rlvp/`](rlvp/) | The method. `grpo.py` (trainer; credit variants outcome / c1 / c2 / c2pos / c3 / c4), `rollout.py` (batched multi-turn rollouts with exact token/turn bookkeeping), `envs/` (FileOps + CSOps deterministic tool-call domains, 4 penalty-only rules each), and `*_adapter.py` (Lean/miniF2F, τ²-bench, TerminalBench, SWE-smith/SWE-gym, WebArena). |\n| [`scripts/`](scripts/) | Baseline grids, training drivers, k=8/k=16 evaluation, phase-diagram and probe scripts. |\n| [`tests/`](tests/) | Rule-loophole suite + credit-assignment unit tests. |\n| [`results/`](results/) | Raw run logs, configs, and evaluations behind every number. See [`results/README.md`](results/README.md). |\n| [`paper/`](paper/) | LaTeX source and figure generators. See [`paper/README.md`](paper/README.md). |\n| [`docs/`](docs/) | Research notes — design rationale, findings, ablations. See [`docs/README.md`](docs/README.md). |\n| [`website/`](website/) | The interactive results site at [01.me/research/rlvp](https://01.me/research/rlvp). |\n\n## Installation\n\n```bash\npip install -r requirements.txt\n```\n\nRequires a CUDA GPU. Models are pulled from the HF hub (Qwen3-1.7B/4B/8B).\nSome domains (Lean/miniF2F, SWE, τ²-bench) need external toolchains — see\n[`requirements.txt`](requirements.txt) and the relevant adapter.\n\n## Quickstart\n\n```bash\n# 1. Unit tests: rules + credit assignment (no GPU needed)\npython3 tests/test_rules.py \u0026\u0026 python3 tests/test_credit.py\n\n# 2. Base-model violation grid (the \"outcome RL can't buy compliance\" baseline)\npython3 scripts/phase0_baseline.py\n\n# 3. Full Phase-1 campaign: credit variants + evals (~3h on one RTX PRO 6000)\nbash scripts/run_all.sh\n\n# 4. Evaluate a checkpoint at k=8, no rules in the prompt\npython3 scripts/eval_checkpoint.py results/run_c3/final c3_norules\n```\n\nTrained checkpoints are not committed (multi-GB LoRA adapters); regenerate them\nwith the driver scripts above.\n\n## Citation\n\n```bibtex\n@article{li2026rlvp,\n  title   = {RLVP: Penalize the Path, Reward the Outcome},\n  author  = {Li, Bojie and Shi, Noah},\n  journal = {arXiv preprint arXiv:2607.07435},\n  year    = {2026},\n  url     = {https://arxiv.org/abs/2607.07435}\n}\n```\n\n## License\n\n[MIT](LICENSE) © 2026 Bojie Li and Noah Shi.\n","project_url":"https://awesome.ecosyste.ms/api/v1/projects/github.com%2F19pine-ai%2Frlvp","html_url":"https://awesome.ecosyste.ms/projects/github.com%2F19pine-ai%2Frlvp","lists_url":"https://awesome.ecosyste.ms/api/v1/projects/github.com%2F19pine-ai%2Frlvp/lists"}