{"id":28410614,"url":"https://github.com/zou-group/sirius","last_synced_at":"2026-03-08T11:31:28.626Z","repository":{"id":286086566,"uuid":"927943606","full_name":"zou-group/sirius","owner":"zou-group","description":"SiriuS: Self-improving Multi-agent Systems via Bootstrapped Reasoning","archived":false,"fork":false,"pushed_at":"2025-04-04T08:34:42.000Z","size":35,"stargazers_count":53,"open_issues_count":1,"forks_count":4,"subscribers_count":14,"default_branch":"main","last_synced_at":"2025-06-02T22:53:46.523Z","etag":null,"topics":["finetuning","llm","multiagent","reasoning","self-improving"],"latest_commit_sha":null,"homepage":"https://arxiv.org/pdf/2502.04780","language":"Python","has_issues":true,"has_wiki":null,"has_pages":null,"mirror_url":null,"source_name":null,"license":"mit","status":null,"scm":"git","pull_requests_enabled":true,"icon_url":"https://github.com/zou-group.png","metadata":{"files":{"readme":"README.md","changelog":null,"contributing":null,"funding":null,"license":"LICENSE","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}},"created_at":"2025-02-05T19:51:54.000Z","updated_at":"2025-05-05T04:40:21.000Z","dependencies_parsed_at":"2025-04-04T09:40:58.702Z","dependency_job_id":null,"html_url":"https://github.com/zou-group/sirius","commit_stats":null,"previous_names":["zou-group/sirius"],"tags_count":0,"template":false,"template_full_name":null,"repository_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repositories/zou-group%2Fsirius","tags_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repositories/zou-group%2Fsirius/tags","releases_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repositories/zou-group%2Fsirius/releases","manifests_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repositories/zou-group%2Fsirius/manifests","owner_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/owners/zou-group","download_url":"https://codeload.github.com/zou-group/sirius/tar.gz/refs/heads/main","sbom_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repositories/zou-group%2Fsirius/sbom","host":{"name":"GitHub","url":"https://github.com","kind":"github","repositories_count":258855048,"owners_count":22768690,"icon_url":"https://github.com/github.png","version":null,"created_at":"2022-05-30T11:31:42.601Z","updated_at":"2022-07-04T15:15:14.044Z","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":["finetuning","llm","multiagent","reasoning","self-improving"],"created_at":"2025-06-02T12:06:59.042Z","updated_at":"2026-03-08T11:31:28.606Z","avatar_url":"https://github.com/zou-group.png","language":"Python","funding_links":[],"categories":[],"sub_categories":[],"readme":"\u003c!--- BADGES: START ---\u003e\n\n[![GitHub license](https://img.shields.io/badge/License-MIT-blue.svg)][#license-gh-package]\n[![Arxiv](https://img.shields.io/badge/arXiv-2406.07496-B31B1B.svg)][#arxiv-paper-package]\n\n\n[#license-gh-package]: https://lbesson.mit-license.org/\n[#arxiv-paper-package]: https://arxiv.org/pdf/2502.04780\n\n\u003c!--- BADGES: END ---\u003e\n \n## SiriuS: Self-improving Multi-agent Systems via Bootstrapped Reasoning (NeurIPS 2025)\n \nThis is the repository for the paper [**SiriuS: Self-improving Multi-agent Systems via Bootstrapped Reasoning**](https://arxiv.org/pdf/2502.04780) (NeurIPS 2025).\n\nSIRIUS is a self-improving multi-agent framework that continuously enhances reasoning ability by maintaining an experience library of successful trajectories and bootstrapping from failed ones.\n\nWe support three main multi-agent settings, each with its own directory:\n\n- `Problem_solving/` – collaborative QA (College Physics/Chemistry, PubMedQA-style)  \n- `Actor_Critic/` – Actor + Judgment + Critic for iterative refinement  \n- `Competitive/` – negotiation / game-theoretic interactions  \n\n\n![Analogy with Torch](assets/task.png)\n\n\n\n### Setup\n\n#### Clone the repo\n\n```bash\ngit clone https://github.com/zou-group/sirius.git\ncd sirius\n```\n\n#### Create environment \u0026 install dependencies\n\n```bash\nconda create -n sirius python=3.10\nconda activate sirius\nconda env create -f environment.yml\n```\n\n#### Configure API access\nSet your keys as environment variables or in a config file as used by the codebase, for example:\n```bash\nexport OPENAI_API_KEY=...\n```\n\n### Repository Overview\n\n\n* `Problem_solving/`\n  Pipelines for college-level reasoning \u0026 biomedical QA:\n  * College Physics / College Chemistry\n  * PubMedQA-style question answering (long context + question)\n\n* `Actor_Critic/`\n  Pipelines for the Actor–Judgment–Critic setting:\n  * Actor proposes an answer\n  * Judgment agent decides correct / incorrect\n  * Critic writes feedback and guides regeneration\n\n* `Competitive/`\n  Pipelines for competitive games:\n  * Resource Exchange\n  * Sell \u0026 Buy\n  * Ultimatum\nEach is a two-player turn-based game with utilities defined in the paper.\n\n#### Data Format \u0026 Trajectories\n\nSiriuS operates on trajectories:\n\n* A trajectory is the full interaction between agents for one task instance:\n\n* Input question / context\n\n* Intermediate messages from each agent (Physicist, Mathematician, Summarizer, Actor, Critic, etc.)\n\n* Final answer(s) or game outcome\n\n* Reward signal(s) (accuracy or utility)\n\n### Quick Start\n\n#### Collect Raw Multi-Agent Trajectories\n\nFirst, run the multi-agent system (with base models) on your tasks and log the full interaction.\n\nA sample training dataset (for physics problem solving) is already provided at:\n```bash\ndataset/phy_train.jsonl\n```\nEach line of this file is one training example (e.g., one physics problem) that the multi-agent system will solve.\n\nPut your training and eval data at \n```bash\ndataset/{subject}_train.jsonl\ndataset/{subject}_test.jsonl\n```\n\nEach subdirectory provides task-specific drivers to:\n\n* Load the dataset \n* Instantiate the appropriate agent graph (see the paper for structures)\n\n  ```bash\n  Problem_solving/PhyChem/agent.py  \n  ```\n\n* solve the problems, collect full trajectories\n  ```bash\n  python Problem_solving/PhyChem/get_a_sol.py --model='gpt-3.5-turbo' --task='MMLU_physics'  --prompt_type='multi_agent' --mode='generate' --subject='phy'\n  ```\n#### Filter Trajectories\n```bash \npython libs/merge.py\n```\n#### Augment Failed Trajectories\n\nFirst, generate feedback for trajectories where the agents produced incorrect solutions:\n  ```bash\n  python Problem_solving/PhyChem/get_b_feedback.py --model='gpt-3.5-turbo' --task='MMLU_physics'  --prompt_type='multi_agent' --mode='generate' --subject='phy'\n  ```\n\nThen, regenerate improved trajectories conditioned on this feedback:\n  ```bash\n  python Problem_solving/PhyChem/get_c_regenerate.py --model='gpt-3.5-turbo' --task='MMLU_physics'  --prompt_type='multi_agent' --mode='generate' --subject='phy'\n  ```\n\n\n#### Fine-Tune Agents on the Library\n\nWe use the OpenAI Supervised Fine-Tuning (SFT) API in our example, but you can plug in any fine-tuning framework of your choice using the constructed experience library:\n\n\n```bash\npython Problem_solving/PhyChem/get_finetune_data.py\npython Problem_solving/PhyChem/fine_tune.py\n```\n\n\n\n```bibtex\n@article{zhao2025sirius,\n  title={SiriuS: Self-improving Multi-agent Systems via Bootstrapped Reasoning},\n  author={Zhao, Wanjia and Yuksekgonul, Mert and Wu, Shirley and Zou, James},\n  journal={arXiv preprint arXiv:2502.04780},\n  year={2025}\n}\n```","project_url":"https://awesome.ecosyste.ms/api/v1/projects/github.com%2Fzou-group%2Fsirius","html_url":"https://awesome.ecosyste.ms/projects/github.com%2Fzou-group%2Fsirius","lists_url":"https://awesome.ecosyste.ms/api/v1/projects/github.com%2Fzou-group%2Fsirius/lists"}