{"slug": "show-hn-a-lego-like-composable-framework-for-rsi", "title": "Show HN: A Lego-like composable Framework for RSI", "summary": "Researchers from Imperial College London, Duke University, the Chinese University of Hong Kong, Brown University and an independent researcher released GEAR, an open framework that organizes recursive self-improvement of LLM agent harnesses as a two-level meta-learning process with an inner rollout loop and an outer Meta Agent loop. In same-model comparisons, GEAR's DSH harness raised task pass rates on AutomationBench Marketing from 27.0% to 40.0%, on AutomationBench Support from 17.0% to 25.0%, and on Terminal-Bench 2.1 from 52.8% to 62.9% with GPT 5.6 Luna at medium reasoning effort, and improved GPT 6 Astra max results over Codex by 4.0, 4.0 and 0.9 percentage points respectively. Code and documentation are published on GitHub at rsi-gear/gear and at rsigear.xyz/docs/gear.", "body_md": "# GEAR: An Open General Evolutionary Architecture\n\n for Recursive Self-Improvement\n\n of LLM Agent Harness\n\n<sup>1</sup>Imperial College London\n\n<sup>2, 5</sup>Independent Researcher\n\n<sup>3</sup>Duke University\n\n<sup>4</sup>The Chinese University of Hong Kong\n\n<sup>6</sup>Brown University\n\n## Learn from experience.\n\nImprove across tasks.\n\nGEAR organizes recursive self-improvement as a **two-level meta-learning process.** In the **inner loop**, a rollout agent uses task feedback to refine its reasoning, planning, and actions within a task.\n\nIn the **outer loop**, a Meta Agent identifies common patterns across task experiences, including recurring failures and effective strategies, and proposes reusable improvements to enhance **cross-task performance.** GEAR evaluates and promotes these changes for subsequent tasks. The two loops can be optimized independently and are connected by experience.\n\n## Modular algorithms for\n\nevolutionary search.\n\nGEAR organizes evolutionary search into replaceable algorithm components,\n\n providing a common framework for studying different search strategies.\n\n Staged GEPA is one concrete instantiation, illustrated below.\n\n### Open the search, component by component.\n\n#### Parent selection\n\nChoose which archived harness to improve next. Staged GEPA uses task-level fronts to retain useful specialists and sample parents from the frozen archive.\n\n- Reads\n- Frozen archive, champion, task-level evidence\n- Produces\n- A parent harness for candidate generation\n\nPolicy modules express the framework’s design boundaries. The current staged implementation exposes parent selection as a public policy hook; evidence validation, evaluation stages, and final promotion remain controlled by the runtime.\n\n## Better harnesses.\n\nMore completed tasks.\n\nImprovements across coding, marketing, and support.\n\nSame-model comparisons, with reasoning effort made explicit.\n\n```\nMarketing · task pass rate27% → 40%GPT 5.6 Luna · medium · DSH Base → OursSupport · task pass rate17% → 25%GPT 5.6 Luna · medium · DSH Base → OursCoding · task pass rate52.8% → 62.9%GPT 5.6 Luna · medium · DSH Base → Ours\n```\n\n## Full benchmark comparison\n\n| Harness comparison · same model and reasoning effort within each group |  |  |  |  |  |  | \n|---|---|---|---|---|---|---|\n| Agent | Harness | AutomationBench · Marketing ↑ |  | AutomationBench · Support ↑ |  | Terminal-Bench 2.1 ↑ | \n|---|---|---|---|---|---|---|\n|  |  | Task pass rate | Obj. completion | Task pass rate | Obj. completion | Task pass rate | \n| GPT 5.6 Luna medium | DSH (Base) | 27.0% | 75.4% | 17.0% | 72.9% | 52.8% | \n|  | DSH (Ours) | 40.0% | 83.9% | 25.0% | 81.3% | 62.9% | \n|  | Δ (pp) | +13.0 | +8.5 | +8.0 | +8.4 | +10.1 | \n| GPT 6 Astra max | Codex | 57.0% | 84.1% | 35.0% | 88.4% | 86.7% | \n|  | DSH (Ours) | 61.0% | 86.3% | 39.0% | 88.7% | 87.6% | \n|  | Δ (pp) | +4.0 | +2.2 | +4.0 | +0.3 | +0.9 | \n\nΔ is the absolute improvement in percentage points.\n\n| Additional evaluation · GPT 5.6 Luna (max) · DSH (Ours) |  |  | \n|---|---|---|\n| Benchmark | Task pass rate | Objective completion | \n|---|---|---|\n| AutomationBench · Marketing | 53.0% | 88.9% | \n| AutomationBench · Support | 27.0% | 82.4% | \n| Terminal-Bench 2.1 | 84.3% | — | \n\nAutomationBench reports task pass rate and objective completion; Terminal-Bench 2.1 reports task pass rate. Luna max is a separate evaluation at a higher reasoning effort. [Result data ↗](assets/paper-results.json)\n\n## Optimize. Execute. Understand.\n\nSeparate responsibilities, connected by versioned evidence.\n\n## Explore GEAR.\n\nCode, experiment records, and the tools to start your own research.\n\n[01 / Source ↗](https://github.com/rsi-gear/gear)\n\n### Get the code\n\nExplore the framework, search components, and examples.\n\ngithub.com/rsi-gear/gear\n[02 / Documentation ↗](https://rsigear.xyz/docs/gear)\n\n### Run your first experiment\n\nSet up a task benchmark and connect the Refine Skill.\n\nRead the guide\n[03 / Experiment ↗](https://github.com/rsi-gear/gear/blob/main/docs/guide/en/example-algorithm.md)\n\n### Follow the search\n\nInspect the staged Marketing experiment and its evaluation evidence.\n\nView the experiment\n## Cite GEAR.\n\nPreprint · October 2026. arXiv link will be added when available.\n\n```\n@unpublished{tang2026gear,\n  title  = {{GEAR}: An Open General Evolutionary Architecture for\n            Recursive Self-Improvement of {LLM} Agent Harness},\n  author = {Tang, Yehui and Zhu, Guanqi and Xia, Shao-Jun and\n            Chen, Xin and Geng, Mengzhe and Du, Hongyang and Tang, Yuhao},\n  year   = {2026},\n  note   = {Preprint},\n  url    = {https://github.com/rsi-gear/gear}\n}\n```\n\n", "url": "https://wpnews.pro/news/show-hn-a-lego-like-composable-framework-for-rsi", "canonical_source": "https://rsi-gear.github.io/", "published_at": "2026-10-28 07:25:08+00:00", "updated_at": "2026-10-28 07:41:43.985300+00:00", "lang": "en", "topics": ["ai-agents", "ai-research", "large-language-models", "artificial-intelligence", "ai-tools"], "entities": ["GEAR", "Imperial College London", "Duke University", "Chinese University of Hong Kong", "Brown University", "GPT 5.6 Luna", "GPT 6 Astra", "Terminal-Bench 2.1"], "also_reported_by": [], "alternates": {"html": "https://wpnews.pro/news/show-hn-a-lego-like-composable-framework-for-rsi", "markdown": "https://wpnews.pro/news/show-hn-a-lego-like-composable-framework-for-rsi.md", "text": "https://wpnews.pro/news/show-hn-a-lego-like-composable-framework-for-rsi.txt", "jsonld": "https://wpnews.pro/news/show-hn-a-lego-like-composable-framework-for-rsi.jsonld"}}