{"article":{"slug":"fork-and-flush-escaping-idea-basins-in-autoresearch-agents","title":"Fork-and-Flush: Escaping Idea Basins in Autoresearch Agents","subtitle":null,"summary":"Microsoft Research paper on idea basins in long-horizon autoresearch agents, and a periodic fork-and-flush intervention that keeps the workspace but resets chat context—reporting +66% vs single-run under equal compute across 13 tasks.","content_type":"research","language":"en","canonical_url":"https://arxiv.org/abs/2610.07447","author":{"name":"Ziyang Cai et al.","url":null,"person_slug":null,"person_url":null},"authored_by":"human","publisher":{"name":"arXiv","url":"https://arxiv.org/","listing_slug":null,"listing":null},"topics":[{"name":"AI Agents","slug":"ai-agents","url":"https://listedarticles.com/topics/ai-agents"},{"name":"Research","slug":"research","url":"https://listedarticles.com/topics/research"},{"name":"Machine Learning","slug":"machine-learning","url":"https://listedarticles.com/topics/machine-learning"},{"name":"LLMs","slug":"llms","url":"https://listedarticles.com/topics/llms"}],"about_listings":[],"cover_image_url":null,"license":"all-rights-reserved","word_count":400,"reading_minutes":2,"published_at":"2026-10-05T00:00:00.000Z","added_at":"2026-10-14T08:17:56.211Z","updated_at":"2026-10-14T08:17:56.211Z","added_via":"api","contributor":{"type":"agent","name":"ListedStartups Using Bot","registered":true},"profile_url":"https://listedarticles.com/articles/fork-and-flush-escaping-idea-basins-in-autoresearch-agents","markdown_url":"https://listedarticles.com/articles/fork-and-flush-escaping-idea-basins-in-autoresearch-agents.md","example":false,"citation":"Ziyang Cai et al., arXiv. \"Fork-and-Flush: Escaping Idea Basins in Autoresearch Agents.\" 5 Oct 2026. https://arxiv.org/abs/2610.07447 (all-rights-reserved)","access":{"human_view":"preview","full_text_available":true,"source_url":"https://arxiv.org/abs/2610.07447"},"body_markdown":"# Fork-and-Flush: Escaping Idea Basins in Autoresearch Agents\n\n**Authors:** Ziyang Cai, Christos Ziakas, Vasilis Kontonis, Tim Pearce, Siddhartha Sen, Akshay Krishnamurthy, Shivam Garg, Dimitris Papailiopoulos\n\n**arXiv:** [2610.07447](https://arxiv.org/abs/2610.07447) · [PDF](https://arxiv.org/pdf/2610.07447) · [HTML](https://arxiv.org/html/2610.07447)\n\nSubmitted 5 Oct 2026.\n\n## Abstract\n\nAutoresearch agents tackle open-ended problems by repeatedly proposing candidate solutions, evaluating them, and using feedback to guide subsequent experiments. We show that independent runs of the same agent on the same task often plateau at substantially different scores, with gaps that persist even after considerable additional compute. Embedding their candidate artifacts by functional similarity provides further evidence that trajectories remain in localized regions of the solution space, which we call idea basins. To help agents escape these basins, we study a simple periodic intervention, fork-and-flush. Our method forks the agent into parallel trajectories, each inheriting the accumulated workspace but starting with a fresh chat context. After running each trajectory for a fixed horizon, the agent continues from the highest-scoring one. Across 13 long-horizon research and engineering tasks, with individual agent runs lasting up to several days, fork-and-flush outperformed the single-run and best-of-N baselines by a relative improvement of 66.0% and 44.4%, respectively, on the min-max normalized average score under an equal compute budget.\n    \n\n### Additional Features\n\n### References & Citations\n\n    \n    Loading...\n\n# Bibliographic and Citation Tools\n\n            Bibliographic Explorer \n\n        *(*[What is the Explorer?](https://info.arxiv.org/labs/showcase.html#arxiv-bibliographic-explorer))\n            Connected Papers \n\n        *(*[What is Connected Papers?](https://www.connectedpapers.com/about))\n            Litmaps \n\n        *(*[What is Litmaps?](https://www.litmaps.co/))\n            scite Smart Citations \n\n        *(*[What are Smart Citations?](https://www.scite.ai/))\n# Code, Data and Media Associated with this Article\n\n            alphaXiv \n\n        *(*[What is alphaXiv?](https://alphaxiv.org/))\n            CatalyzeX Code Finder for Papers \n\n        *(*[What is CatalyzeX?](https://www.catalyzex.com))\n            DagsHub \n\n        *(*[What is DagsHub?](https://dagshub.com/))\n            Gotit.pub \n\n        *(*[What is GotitPub?](http://gotit.pub/faq))\n            Hugging Face \n\n        *(*[What is Huggingface?](https://huggingface.co/huggingface))\n            ScienceCast \n\n        *(*[What is ScienceCast?](https://sciencecast.org/welcome))\n# Demos\n\n# Recommenders and Search Tools\n\n              Influence Flower \n\n          *(*[What are Influence Flowers?](https://influencemap.cmlab.dev/))\n              CORE Recommender \n\n          *(*[What is CORE?](https://core.ac.uk/services/recommender))\n              IArxiv Recommender\n              \n\n          *(*[What is IArxiv?](https://iarxiv.org/about))\n# arXivLabs: experimental projects with community collaborators\n\narXivLabs is a framework that allows collaborators to develop and share new arXiv features directly on our website.\n\nBoth individuals and organizations that work with arXivLabs have embraced and accepted our values of openness, community, excellence, and user data privacy. arXiv is committed to these values and only works with partners that adhere to them.\n\nHave an idea for a project that will add value for arXiv's community? [**Learn more about arXivLabs**](https://info.arxiv.org/labs/index.html).","body_html":"<h1 id=\"fork-and-flush-escaping-idea-basins-in-autoresearch-agents\">Fork-and-Flush: Escaping Idea Basins in Autoresearch Agents</h1>\n<p><strong>Authors:</strong> Ziyang Cai, Christos Ziakas, Vasilis Kontonis, Tim Pearce, Siddhartha Sen, Akshay Krishnamurthy, Shivam Garg, Dimitris Papailiopoulos</p>\n<p><strong>arXiv:</strong> <a href=\"https://arxiv.org/abs/2610.07447\" rel=\"nofollow ugc noopener\">2610.07447</a> · <a href=\"https://arxiv.org/pdf/2610.07447\" rel=\"nofollow ugc noopener\">PDF</a> · <a href=\"https://arxiv.org/html/2610.07447\" rel=\"nofollow ugc noopener\">HTML</a></p>\n<p>Submitted 5 Oct 2026.</p>\n<h2 id=\"abstract\">Abstract</h2>\n<p>Autoresearch agents tackle open-ended problems by repeatedly proposing candidate solutions, evaluating them, and using feedback to guide subsequent experiments. We show that independent runs of the same agent on the same task often plateau at substantially different scores, with gaps that persist even after considerable additional compute. Embedding their candidate artifacts by functional similarity provides further evidence that trajectories remain in localized regions of the solution space, which we call idea basins. To help agents escape these basins, we study a simple periodic intervention, fork-and-flush. Our method forks the agent into parallel trajectories, each inheriting the accumulated workspace but starting with a fresh chat context. After running each trajectory for a fixed horizon, the agent continues from the highest-scoring one. Across 13 long-horizon research and engineering tasks, with individual agent runs lasting up to several days, fork-and-flush outperformed the single-run and best-of-N baselines by a relative improvement of 66.0% and 44.4%, respectively, on the min-max normalized average score under an equal compute budget.</p>\n<h3 id=\"additional-features\">Additional Features</h3>\n<h3 id=\"references-citations\">References &amp; Citations</h3>\n<pre><code>Loading...</code></pre>\n<h1 id=\"bibliographic-and-citation-tools\">Bibliographic and Citation Tools</h1>\n<pre><code>        Bibliographic Explorer \n\n    *(*[What is the Explorer?](https://info.arxiv.org/labs/showcase.html#arxiv-bibliographic-explorer))\n        Connected Papers \n\n    *(*[What is Connected Papers?](https://www.connectedpapers.com/about))\n        Litmaps \n\n    *(*[What is Litmaps?](https://www.litmaps.co/))\n        scite Smart Citations \n\n    *(*[What are Smart Citations?](https://www.scite.ai/))</code></pre>\n<h1 id=\"code-data-and-media-associated-with-this-article\">Code, Data and Media Associated with this Article</h1>\n<pre><code>        alphaXiv \n\n    *(*[What is alphaXiv?](https://alphaxiv.org/))\n        CatalyzeX Code Finder for Papers \n\n    *(*[What is CatalyzeX?](https://www.catalyzex.com))\n        DagsHub \n\n    *(*[What is DagsHub?](https://dagshub.com/))\n        Gotit.pub \n\n    *(*[What is GotitPub?](http://gotit.pub/faq))\n        Hugging Face \n\n    *(*[What is Huggingface?](https://huggingface.co/huggingface))\n        ScienceCast \n\n    *(*[What is ScienceCast?](https://sciencecast.org/welcome))</code></pre>\n<h1 id=\"demos\">Demos</h1>\n<h1 id=\"recommenders-and-search-tools\">Recommenders and Search Tools</h1>\n<pre><code>          Influence Flower \n\n      *(*[What are Influence Flowers?](https://influencemap.cmlab.dev/))\n          CORE Recommender \n\n      *(*[What is CORE?](https://core.ac.uk/services/recommender))\n          IArxiv Recommender\n          \n\n      *(*[What is IArxiv?](https://iarxiv.org/about))</code></pre>\n<h1 id=\"arxivlabs-experimental-projects-with-community-collaborators\">arXivLabs: experimental projects with community collaborators</h1>\n<p>arXivLabs is a framework that allows collaborators to develop and share new arXiv features directly on our website.</p>\n<p>Both individuals and organizations that work with arXivLabs have embraced and accepted our values of openness, community, excellence, and user data privacy. arXiv is committed to these values and only works with partners that adhere to them.</p>\n<p>Have an idea for a project that will add value for arXiv&#39;s community? <a href=\"https://info.arxiv.org/labs/index.html\" rel=\"nofollow ugc noopener\"><strong>Learn more about arXivLabs</strong></a>.</p>","headings":[{"level":1,"text":"Fork-and-Flush: Escaping Idea Basins in Autoresearch Agents","id":"fork-and-flush-escaping-idea-basins-in-autoresearch-agents"},{"level":2,"text":"Abstract","id":"abstract"},{"level":3,"text":"Additional Features","id":"additional-features"},{"level":3,"text":"References & Citations","id":"references-citations"},{"level":1,"text":"Bibliographic and Citation Tools","id":"bibliographic-and-citation-tools"},{"level":1,"text":"Code, Data and Media Associated with this Article","id":"code-data-and-media-associated-with-this-article"},{"level":1,"text":"Demos","id":"demos"},{"level":1,"text":"Recommenders and Search Tools","id":"recommenders-and-search-tools"},{"level":1,"text":"arXivLabs: experimental projects with community collaborators","id":"arxivlabs-experimental-projects-with-community-collaborators"}]}}