Hugging Face's logo

Spaces:
Pleuron
/
ToolTree
Running

App Files Community
ToolTree / index.html
Pleuron's picture
Pleuron
Upload folder using huggingface_hub
bf0d7ad verified
Raw
History Blame Contribute Delete
18.9 kB
<!DOCTYPE html>
<html lang="en">
<head>
<meta charset="utf-8">
<meta name="viewport" content="width=device-width, initial-scale=1">
<title>ToolTree: Efficient LLM Agent Tool Planning via Dual-Feedback Monte Carlo Tree Search and Bidirectional Pruning</title>
<meta name="description" content="ToolTree plans LLM agent tool use via dual-feedback Monte Carlo tree search with bidirectional pruning — state-of-the-art across 4 benchmarks with an average gain of ~10% over existing methods.">
<link rel="icon" href="data:,">
<link rel="stylesheet" href="static/style.css">
<script>document.documentElement.classList.add('js');</script>
</head>
<body>
<!-- ================= Sticky section nav (shown after scrolling past the hero) ================= -->
<nav class="topnav" id="topnav" aria-label="Page sections">
 <div class="topnav-inner">
 <a class="topnav-brand" href="#top">ToolTree</a>
 <div class="topnav-links">
 <a href="#overview">Overview</a>
 <a href="#efficiency">Efficiency</a>
 <a href="#method">Method</a>
 <a href="#case-study">Case Study</a>
 <a href="#results">Results</a>
 <a href="#ablation">Ablation</a>
 <a href="demo.html">Demo</a>
 <a href="#bibtex">BibTeX</a>
 </div>
 </div>
</nav>
<!-- ================= Hero ================= -->
<header class="hero" id="top">
 <div class="container">
 <h1><span class="hero-name">ToolTree</span>: Efficient LLM Agent Tool Planning via Dual-Feedback Monte Carlo Tree Search and Bidirectional Pruning</h1>
 <p class="authors">
 Shuo Yang, Caren Han, Yihao Ding, Shuhe Wang, Eduard Hovy
 </p>
 <p class="affiliation">The University of Melbourne &#183; The University of Western Australia</p>
 <p class="venue"><span class="venue-chip">ICLR 2026</span></p>
 <nav class="button-row" aria-label="Paper links">
 <a class="btn" href="https://openreview.net/forum?id=Ef5O9gNNLE" target="_blank" rel="noopener"><svg class="btn-ico" viewBox="0 0 24 24" aria-hidden="true" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round"><path d="M14 2H6a2 2 0 0 0-2 2v16a2 2 0 0 0 2 2h12a2 2 0 0 0 2-2V8z"/><polyline points="14 2 14 8 20 8"/></svg>Paper</a>
 <a class="btn" href="https://arxiv.org/abs/2603.12740" target="_blank" rel="noopener"><svg class="btn-ico" viewBox="0 0 24 24" aria-hidden="true" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round"><path d="M14 2H6a2 2 0 0 0-2 2v16a2 2 0 0 0 2 2h12a2 2 0 0 0 2-2V8z"/><polyline points="14 2 14 8 20 8"/></svg>arXiv</a>
 <a class="btn" href="https://github.com/SYang2000/ICLR_2026_ToolTree" target="_blank" rel="noopener"><svg class="btn-ico" viewBox="0 0 24 24" aria-hidden="true" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round"><polyline points="16 18 22 12 16 6"/><polyline points="8 6 2 12 8 18"/></svg>Code</a>
 <a class="btn" href="demo.html"><svg class="btn-ico" viewBox="0 0 24 24" aria-hidden="true" fill="currentColor" stroke="none"><polygon points="7 4 20 12 7 20 7 4"/></svg>Demo</a>
 <a class="btn btn-ghost" href="#bibtex"><svg class="btn-ico" viewBox="0 0 24 24" aria-hidden="true" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round"><path d="M4 19.5A2.5 2.5 0 0 1 6.5 17H20"/><path d="M6.5 2H20v20H6.5A2.5 2.5 0 0 1 4 19.5v-15A2.5 2.5 0 0 1 6.5 2z"/></svg>BibTeX</a>
 </nav>
 </div>
</header>
<main>
<!-- ================= Metric callouts ================= -->
<section class="metrics">
 <div class="container">
 <div class="metric-grid reveal stagger">
 <div class="metric-card">
 <div class="metric-value">~10%</div>
 <div class="metric-label">Average gain over existing methods</div>
 <div class="metric-sub">Across both closed-set and open-set tool planning scenarios</div>
 </div>
 <div class="metric-card">
 <div class="metric-value">4 / 4</div>
 <div class="metric-label">State-of-the-art on all four benchmarks</div>
 <div class="metric-sub">GTA &#183; m&amp;m &#183; ToolBench &#183; RestBench</div>
 </div>
 <div class="metric-card">
 <div class="metric-value">#1</div>
 <div class="metric-label">Highest efficiency (performance gain per second)</div>
 <div class="metric-sub">Compared with all baselines across step limits</div>
 </div>
 </div>
 </div>
</section>
<!-- ================= Overview ================= -->
<section id="overview">
 <div class="container">
 <h2>Overview</h2>
 <div class="abstract-card reveal">
 <p class="abstract-text">
 <strong>ToolTree</strong> is a novel Monte Carlo tree search-inspired planning paradigm for
 LLM agent tool planning. It explores possible tool usage trajectories using a
 <strong>dual-stage LLM evaluation</strong> and <strong>bidirectional pruning</strong> mechanism
 that enables the agent to make informed, adaptive decisions over extended tool-use sequences
 while pruning less promising branches before and after the tool execution.
 </p>
 </div>
<figure class="paper-figure reveal">
 <img src="static/paper/comparison.png" alt="Concept comparison on a street photo asking how many wheels are in total: greedy-based planning commits to a single tool chain and answers incorrectly; search-based planning explores more tool branches but still returns a wrong count; ToolTree prunes branches before (pre-pruning) and after (post-pruning) execution and answers correctly.">
 <figcaption>Comparison of ToolTree with greedy search and search-based tool planning. ToolTree chooses
 the optimal tool trajectory and answers correctly with bidirectional pruning.</figcaption>
 </figure>
 </div>
</section>
<!-- ================= Efficiency (Figure 3) ================= -->
<section class="fig3-band section-alt" id="efficiency" aria-label="Progressive efficiency analysis">
 <div class="container">
 <h2>Efficiency</h2>
 <figure class="paper-figure reveal">
 <img src="static/paper/efficiency.png" alt="Three line charts across step limits comparing ReAct, Best-first, ToT, LATS, and ToolTree: performance versus step limit, running time versus step limit, and efficiency (performance gain per second) versus step limit — ToolTree sits highest on the performance and efficiency panels at every step limit.">
 <figcaption><span class="fig-tag">Figure 3:</span> Progressive efficiency analysis across step limits.
 ToolTree achieves the highest efficiency (performance gain per second) compared with all baselines.</figcaption>
 </figure>
 </div>
</section>
<!-- ================= Method ================= -->
<section id="method">
 <div class="container">
 <h2>Method</h2>
 <figure class="paper-figure reveal">
 <img src="static/paper/architecture.png" alt="ToolTree pipeline: an input query enters a repeated loop over the tool tree — selection, pre-evaluation of a candidate tool against a threshold before execution, expansion, execution, post-evaluation of the observed output, and backward-propagation — after which the answer predictor produces the final answer.">
 <figcaption>Architecture overview of ToolTree. An input query is processed sequentially via iterative
 dual evaluation-guided Monte Carlo Tree Search, including selection, pre-evaluation, expansion,
 execution, post-evaluation and backward-propagation.</figcaption>
 </figure>
<figure class="paper-figure reveal">
 <a href="static/paper/case3_tree.svg" target="_blank" rel="noopener"><img src="static/paper/case3_tree.svg" width="1680" height="1500" alt="Search tree reconstructed from a logged 11-rollout run: from the user query on the left, candidate tool calls branch rightward with pre- and post-evaluation score badges; pre-pruned drafts are dashed with a cross, a failed call is post-pruned in red, duplicate drafts are shown as skipped, and the bold path leads to the final answer bar."></a>
 <figcaption>Search over tool-call trajectories on a real logged run (11 rollouts): each node is a
 candidate tool call with its pre-/post-evaluation scores, pruned and duplicate drafts are shown
 as encountered, and the highest-reward trajectory (bold) is selected as the final plan.
 Click to view full size.</figcaption>
 </figure>
<div class="mech-grid mech-grid-4 reveal stagger">
 <div class="mech-card">
 <span class="mech-num">1</span>
 <h3>Pre-Evaluation</h3>
 <p>
 A fast predictive signal that estimates the utility of a tool <em>before</em> execution,
 filtering schema- or slot-incompatible calls before expansion.
 </p>
 </div>
 <div class="mech-card">
 <span class="mech-num">2</span>
 <h3>Post-Evaluation</h3>
 <p>
 Assesses the actual contribution of a tool <em>after</em> execution based on observed
 outcomes, pruning unproductive branches using real feedback.
 </p>
 </div>
 <div class="mech-card">
 <span class="mech-num">3</span>
 <h3>Bidirectional Pruning</h3>
 <p>
 Combines pre- and post-evaluation to eliminate unpromising branches, concentrating
 computational budget on promising tool chains.
 </p>
 </div>
 <div class="mech-card">
 <span class="mech-num">4</span>
 <h3>Answer Predictor</h3>
 <p>
 Incorporates the tool trajectories with the highest reward found by the MCTS to produce
 the final prediction.
 </p>
 </div>
 </div>
 </div>
</section>
<!-- ================= Case study ================= -->
<section id="case-study" class="section-alt">
 <div class="container">
 <h2>Case Study</h2>
 <span class="band-label">Qualitative case studies from the paper</span>
 <figure class="paper-figure reveal">
 <div class="figure-stack">
 <img src="static/paper/case_study_medical.png" alt="Case study (a): a radiology image question about lung cancer risk — greedy search relies on generic image captioning and heuristic search on shallow heuristics, both missing the finding, while the MCTS-planned trajectory orchestrates domain-specific imaging tools and detects the risk.">
 <img src="static/paper/case_study_reasoning.png" alt="Case study (b): a multi-hop knowledge reasoning question about the building seen behind an amusement area — greedy search and heuristic search answer from broad scene cues and the most prominent text, while the MCTS-planned trajectory chains recognition and search tools to identify the building correctly.">
 </div>
 <figcaption>Two qualitative case studies showcasing ToolTree&#8217;s iterative tool orchestration on
 (a) a radiology image question and (b) a multi-hop knowledge reasoning task.</figcaption>
 </figure>
 </div>
</section>
<!-- ================= Results ================= -->
<section id="results">
 <div class="container">
 <h2>Results</h2>
 <p>
 ToolTree achieves state-of-the-art performance across 4 benchmarks spanning both closed-set and
 open-set tool planning scenarios, with an average gain of ~10% over existing methods.
 </p>
<div class="table-wrap reveal">
 <table class="results">
 <thead>
 <tr><th>Benchmark</th><th>Setting</th><th>Tasks &amp; Tools</th><th>Official source</th></tr>
 </thead>
 <tbody>
 <tr><td>GTA</td><td>Closed-set</td><td>229 real-world tasks, 14 executable tools</td><td><a href="https://github.com/open-compass/GTA" target="_blank" rel="noopener">open-compass/GTA</a> &#183; <a href="https://huggingface.co/datasets/Jize1/GTA" target="_blank" rel="noopener">HF dataset</a></td></tr>
 <tr><td>m&amp;m</td><td>Closed-set</td><td>882 human-verified multi-step multimodal tasks, 33 tools</td><td><a href="https://github.com/RAIVNLab/mnms" target="_blank" rel="noopener">RAIVNLab/mnms</a> &#183; <a href="https://huggingface.co/datasets/zixianma/mnms" target="_blank" rel="noopener">HF dataset</a></td></tr>
 <tr><td>ToolBench</td><td>Open-set</td><td>16,464 real-world REST APIs (RapidAPI)</td><td><a href="https://github.com/OpenBMB/ToolBench" target="_blank" rel="noopener">OpenBMB/ToolBench</a></td></tr>
 <tr><td>RestBench</td><td>Open-set</td><td>TMDB &amp; Spotify REST scenarios</td><td><a href="https://github.com/Yifan-Song793/RestGPT" target="_blank" rel="noopener">Yifan-Song793/RestGPT</a></td></tr>
 </tbody>
 </table>
 </div>
<p class="table-note reveal">
 This repository ships no benchmark data; each benchmark is downloaded from its official source.
 </p>
</div>
</section>
<!-- ================= Ablation ================= -->
<section id="ablation" class="section-alt">
 <div class="container">
 <h2>Pruning Ablation</h2>
<div class="ablation-block reveal">
 <figure class="paper-figure figure-medium">
 <img src="static/paper/ablation_pruning.png" alt="Two box plots comparing ToolTree against variants without pre-pruning, without post-pruning, and without both: the number of rollouts and the number of expanded nodes are lowest for full ToolTree and grow as pruning stages are disabled.">
 </figure>
 <p class="takeaway">Disabling pre-pruning, post-pruning, or both consistently increases the number of
 rollouts and expanded nodes, confirming that bidirectional pruning concentrates the computational
 budget on promising tool chains.</p>
 </div>
 </div>
</section>
<!-- ================= BibTeX ================= -->
<section id="bibtex">
 <div class="container">
 <h2>BibTeX</h2>
 <div class="bibtex-box reveal">
 <button class="copy-btn" id="copy-btn" type="button">Copy</button>
 <pre id="bibtex-text">@inproceedings{yang2026tooltree,
title={ToolTree: Efficient {LLM} Tool Planning via Dual-Feedback Monte Carlo Tree Search and Bidirectional Pruning},
author={Shuo Yang and Caren Han and Yihao Ding and Shuhe Wang and Eduard Hovy},
booktitle={The Fourteenth International Conference on Learning Representations},
year={2026},
url={https://openreview.net/forum?id=Ef5O9gNNLE}
}</pre>
 </div>
 </div>
</section>
</main>
<footer>
 <div class="container">
 &#169; 2026 the authors &#183; The University of Melbourne &#183; The University of Western Australia &#183; Code released under the MIT License.
 </div>
</footer>
<script>
/* Sticky section nav: slides in once the hero has scrolled out of view. */
(function () {
 var hero = document.getElementById('top');
 var nav = document.getElementById('topnav');
 if (!hero || !nav || !('IntersectionObserver' in window)) return;
 var io = new IntersectionObserver(function (entries) {
 nav.classList.toggle('visible', !entries[0].isIntersecting);
 }, { rootMargin: '-56px 0px 0px 0px' });
 io.observe(hero);
})();
/* Scroll-reveal: fade/lift sections in as they enter the viewport.
 Elements are only hidden when <html> carries the .js class (added in <head>),
 so content stays visible with JS disabled; reduced-motion is handled in CSS. */
(function () {
 var els = document.querySelectorAll('.reveal');
 var i;
 if (!('IntersectionObserver' in window)) {
 for (i = 0; i < els.length; i++) els[i].classList.add('revealed');
 return;
 }
 var io = new IntersectionObserver(function (entries) {
 entries.forEach(function (en) {
 if (en.isIntersecting) {
 en.target.classList.add('revealed');
 io.unobserve(en.target);
 }
 });
 }, { threshold: 0.08, rootMargin: '0px 0px -8% 0px' });
 for (i = 0; i < els.length; i++) io.observe(els[i]);
})();
/* Lightbox: click (or Enter/Space) on a paper figure opens it enlarged; ESC / click closes. */
(function () {
 var imgs = document.querySelectorAll('figure.paper-figure img, .case-figure img');
 if (!imgs.length) return;
 var overlay = null, big = null, lastFocus = null;
 function build() {
 overlay = document.createElement('div');
 overlay.className = 'lightbox';
 overlay.setAttribute('role', 'dialog');
 overlay.setAttribute('aria-modal', 'true');
 overlay.setAttribute('aria-label', 'Enlarged figure');
 big = document.createElement('img');
 var closeBtn = document.createElement('button');
 closeBtn.className = 'lightbox-close';
 closeBtn.type = 'button';
 closeBtn.setAttribute('aria-label', 'Close enlarged figure');
 closeBtn.innerHTML = '&#215;';
 overlay.appendChild(big);
 overlay.appendChild(closeBtn);
 overlay.addEventListener('click', close);
 document.body.appendChild(overlay);
 }
 function open(img) {
 if (!overlay) build();
 lastFocus = img;
 big.src = img.currentSrc || img.src;
 big.alt = img.alt || '';
 overlay.classList.add('open');
 document.body.classList.add('no-scroll');
 overlay.querySelector('.lightbox-close').focus();
 }
 function close() {
 if (!overlay || !overlay.classList.contains('open')) return;
 overlay.classList.remove('open');
 document.body.classList.remove('no-scroll');
 if (lastFocus && lastFocus.focus) lastFocus.focus();
 }
 document.addEventListener('keydown', function (e) {
 if (e.key === 'Escape' || e.key === 'Esc') close();
 });
 Array.prototype.forEach.call(imgs, function (img) {
 img.setAttribute('tabindex', '0');
 img.setAttribute('role', 'button');
 img.addEventListener('click', function () { open(img); });
 img.addEventListener('keydown', function (e) {
 if (e.key === 'Enter' || e.key === ' ' || e.key === 'Spacebar') {
 e.preventDefault();
 open(img);
 }
 });
 });
})();
/* BibTeX copy button. */
(function () {
 var btn = document.getElementById('copy-btn');
 var pre = document.getElementById('bibtex-text');
 if (!btn || !pre) return;
 function flash(ok) {
 btn.textContent = ok ? 'Copied!' : 'Copy failed — select the text manually';
 setTimeout(function () { btn.textContent = 'Copy'; }, 2000);
 }
 function fallbackCopy(text) {
 var ta = document.createElement('textarea');
 ta.value = text;
 ta.setAttribute('readonly', '');
 ta.style.position = 'fixed';
 ta.style.left = '-9999px';
 document.body.appendChild(ta);
 ta.select();
 var ok = false;
 try { ok = document.execCommand('copy'); } catch (e) { ok = false; }
 document.body.removeChild(ta);
 flash(ok);
 }
 btn.addEventListener('click', function () {
 var text = pre.textContent;
 if (navigator.clipboard && navigator.clipboard.writeText) {
 navigator.clipboard.writeText(text).then(
 function () { flash(true); },
 function () { fallbackCopy(text); }
 );
 } else {
 fallbackCopy(text);
 }
 });
})();
</script>
</body>
</html>