Spaces:
Running
Running
| <html lang="en"> | |
| <head> | |
| <meta charset="utf-8"> | |
| <meta name="viewport" content="width=device-width, initial-scale=1"> | |
| <meta name="description" content="ImproveAnyTask is an autonomous post-training harness for iterative model self-improvement across 11 benchmarks."> | |
| <meta name="keywords" content="ImproveAnyTask, autonomous post-training, recursive self-improvement, iterative self-improvement, self-improving AI, LLM adaptation, agent harness"> | |
| <meta name="citation_title" content="ImproveAnyTask: An Autonomous Post-Training Harness for Iterative Model Self-Improvement"> | |
| <meta name="citation_publication_date" content="2026"> | |
| <meta name="citation_arxiv_id" content="2610.06347"> | |
| <meta name="citation_pdf_url" content="https://arxiv.org/pdf/2610.06347"> | |
| <title>ImproveAnyTask: Autonomous Post-Training for Iterative Model Self-Improvement</title> | |
| <style> | |
| :root{--ink:#172033;--muted:#5d687c;--line:#d8deea;--blue:#1f67c1;--blue2:#eaf3ff;--paper:#fff;--wash:#f6f8fb} | |
| *{box-sizing:border-box} html{scroll-behavior:smooth} | |
| body{margin:0;background:var(--paper);color:var(--ink);font-family:Inter,ui-sans-serif,-apple-system,BlinkMacSystemFont,"Segoe UI",sans-serif;line-height:1.65} | |
| a{color:var(--blue);text-decoration:none} a:hover{text-decoration:underline} | |
| .wrap{width:min(1120px,calc(100% - 40px));margin:auto} | |
| nav{border-bottom:1px solid var(--line);background:#fff} | |
| nav .wrap{display:flex;align-items:center;justify-content:space-between;min-height:62px;gap:20px} | |
| .brand{font-weight:800;color:var(--ink);font-size:18px}.navlinks{display:flex;gap:20px;font-size:14px;font-weight:650} | |
| header{padding:58px 0 30px;text-align:center} | |
| .tag{display:inline-block;border:1px solid #b7cbea;background:var(--blue2);color:#174f93;border-radius:999px;padding:5px 12px;font-size:13px;font-weight:750} | |
| h1{font-family:Georgia,"Times New Roman",serif;font-size:clamp(36px,5.2vw,62px);line-height:1.08;letter-spacing:0;margin:20px auto 14px;max-width:1050px} | |
| .subtitle{font-family:Georgia,"Times New Roman",serif;font-size:clamp(19px,2.4vw,27px);color:#39465c;margin:0 auto 22px;max-width:920px} | |
| .authors{font-size:15px;max-width:980px;margin:0 auto;color:#354158}.authors strong{color:var(--ink)} | |
| .actions{display:flex;justify-content:center;gap:10px;flex-wrap:wrap;margin-top:25px} | |
| .button{display:inline-flex;align-items:center;justify-content:center;border:1px solid #afbdd0;border-radius:6px;padding:9px 16px;color:var(--ink);background:#fff;font-weight:750;font-size:14px} | |
| .button.primary{background:var(--blue);border-color:var(--blue);color:#fff}.button:hover{text-decoration:none;transform:translateY(-1px)} | |
| .hero-figure{margin:24px auto 0;border-top:1px solid var(--line);border-bottom:1px solid var(--line);padding:24px 0} | |
| .hero-figure img{display:block;width:100%;height:auto} | |
| .caption{font-size:13px;color:var(--muted);margin:10px 0 0;text-align:center} | |
| .metrics{display:grid;grid-template-columns:repeat(3,1fr);border-top:1px solid var(--line);border-bottom:1px solid var(--line);margin:34px 0} | |
| .metric{padding:22px 18px;text-align:center}.metric+ .metric{border-left:1px solid var(--line)} | |
| .metric b{display:block;font-size:30px;color:#124f99}.metric span{font-size:13px;color:var(--muted)} | |
| section{padding:34px 0} h2{font-family:Georgia,"Times New Roman",serif;font-size:31px;line-height:1.2;margin:0 0 12px} | |
| .lead{font-size:18px;color:#36435a;max-width:900px} | |
| .method{display:grid;grid-template-columns:repeat(3,1fr);gap:14px;margin-top:22px} | |
| .step{border:1px solid var(--line);border-radius:7px;padding:20px;background:#fff}.step .num{font-size:12px;font-weight:800;color:var(--blue);letter-spacing:.08em}.step h3{margin:6px 0 8px;font-size:18px}.step p{margin:0;color:var(--muted);font-size:14px} | |
| .band{background:var(--wash);border-top:1px solid var(--line);border-bottom:1px solid var(--line)} | |
| .links{display:grid;grid-template-columns:1.4fr 1fr;gap:34px;align-items:start} | |
| pre{white-space:pre-wrap;overflow-wrap:anywhere;background:#111827;color:#e5edf8;border-radius:7px;padding:18px;font-size:12px;line-height:1.55;margin:0} | |
| footer{border-top:1px solid var(--line);padding:24px 0 36px;color:var(--muted);font-size:13px} | |
| @media(max-width:760px){.navlinks{gap:10px}.navlinks a:nth-child(2){display:none}header{padding-top:38px}.metrics,.method,.links{grid-template-columns:1fr}.metric+.metric{border-left:0;border-top:1px solid var(--line)}.wrap{width:min(100% - 24px,1120px)}} | |
| </style> | |
| </head> | |
| <body> | |
| <nav><div class="wrap"><a class="brand" href="https://improveanytask.github.io/">ImproveAnyTask</a><div class="navlinks"><a href="#method">Method</a><a href="#results">Results</a><a href="#citation">Citation</a></div></div></nav> | |
| <header class="wrap"> | |
| <span class="tag">arXiv:2610.06347 · Autonomous Post-Training</span> | |
| <h1>ImproveAnyTask</h1> | |
| <p class="subtitle">An Autonomous Post-Training Harness for Iterative Model Self-Improvement</p> | |
| <p class="authors"><strong>Xingbo Yao*</strong>, <strong>Xiaoman Wang*</strong>, Zhengwu Lei, Tinghui Luo, YiLin Zhang, Yuefeng Wu, Yijie Xu, Tianfu Wang, Qingyuan Zhan, Ye Guo, Daoxin Zhang, Zhe Xu, Jian Liu†, Hui Xiong†<br><small>* Equal contribution · † Corresponding authors</small></p> | |
| <div class="actions"> | |
| <a class="button primary" href="https://arxiv.org/abs/2610.06347" target="_blank" rel="noopener">Paper</a> | |
| <a class="button" href="https://github.com/improveanytask/improveanytask.github.io" target="_blank" rel="noopener">GitHub</a> | |
| <a class="button" href="https://improveanytask.github.io/" target="_blank" rel="noopener">Project Page</a> | |
| <a class="button" href="https://arxiv.org/pdf/2610.06347" target="_blank" rel="noopener">PDF</a> | |
| </div> | |
| </header> | |
| <main> | |
| <div class="wrap hero-figure"> | |
| <img src="https://improveanytask.github.io/assets/teaser-paper-final.png" alt="ImproveAnyTask results across eleven benchmarks and five capability domains"> | |
| <p class="caption">ImproveAnyTask improves Base and Instruct models across knowledge, reasoning, instruction following, coding, and tool-use tasks.</p> | |
| </div> | |
| <section id="results" class="wrap"> | |
| <div class="metrics"> | |
| <div class="metric"><b>+18.29 pp</b><span>Mean gain on Qwen3.5-4B-Base</span></div> | |
| <div class="metric"><b>+11.97 pp</b><span>Mean gain on Qwen3.5-4B-Instruct</span></div> | |
| <div class="metric"><b>+41.96 pp</b><span>Maximum gain across 11 tasks</span></div> | |
| </div> | |
| <h2>Task-specific improvement with less manual iteration</h2> | |
| <p class="lead">Adapting a general-purpose LLM to a target task usually requires repeated human decisions about error analysis, data construction, and training. ImproveAnyTask autonomously organizes this process under a limited compute budget equivalent to eight NVIDIA H20 GPUs for 24 hours.</p> | |
| </section> | |
| <section id="method" class="band"><div class="wrap"> | |
| <h2>A connected optimization loop</h2> | |
| <p class="lead">The harness turns evaluation evidence into focused, research-backed, and executable model updates, then repeats the process as the model's error distribution changes.</p> | |
| <div class="method"> | |
| <article class="step"><div class="num">01 · DIAGNOSE</div><h3>Error Attribution</h3><p>Combines metric-level and case-level evidence to identify the highest-priority weakness for the current iteration.</p></article> | |
| <article class="step"><div class="num">02 · RESEARCH</div><h3>Update Direction</h3><p>Compares candidate strategies around the same weakness using reported gains, task fit, and reproduction difficulty.</p></article> | |
| <article class="step"><div class="num">03 · EXECUTE</div><h3>Model Updates</h3><p>Builds training data and configurations, checks execution at small scale, trains, evaluates, and retains reusable assets.</p></article> | |
| </div> | |
| </div></section> | |
| <section id="citation" class="wrap"> | |
| <div class="links"><div><h2>Paper and resources</h2><p class="lead">Read the full paper on arXiv, explore the public project page, or follow development on GitHub.</p><div class="actions" style="justify-content:flex-start"><a class="button primary" href="https://arxiv.org/abs/2610.06347">arXiv</a><a class="button" href="https://github.com/improveanytask/improveanytask.github.io">GitHub</a><a class="button" href="https://improveanytask.github.io/">Project</a></div></div><pre>@article{yao2026improveanytask, | |
| title={ImproveAnyTask: An Autonomous Post-Training Harness for Iterative Model Self-Improvement}, | |
| author={Yao, Xingbo and Wang, Xiaoman and Lei, Zhengwu and others}, | |
| journal={arXiv preprint arXiv:2610.06347}, | |
| year={2026} | |
| }</pre></div> | |
| </section> | |
| </main> | |
| <footer><div class="wrap">ImproveAnyTask · Autonomous post-training · Iterative and recursive model self-improvement</div></footer> | |
| </body> | |
| </html> |