File size: 40,914 Bytes
cf573f9 daa6be0 cf573f9 a8df6b3 9337e18 a8df6b3 845705b 9337e18 2254920 9337e18 a8df6b3 9337e18 cf573f9 daa6be0 cf573f9 22e4ecb 2254920 cf573f9 2254920 cf573f9 a8df6b3 9f54287 cf573f9 a8df6b3 2254920 cf573f9 a8df6b3 a8b71f1 ecd1f04 a8b71f1 2254920 ecd1f04 2254920 a8b71f1 ecd1f04 2254920 a8b71f1 2254920 770000d cf573f9 9c2d624 cf573f9 9c2d624 a8df6b3 9f54287 a8df6b3 7247a32 a8df6b3 9f54287 a8df6b3 7247a32 a8df6b3 9f54287 a8df6b3 9f54287 a8df6b3 2254920 a8df6b3 845705b a8df6b3 845705b a8df6b3 bd3e774 a8df6b3 9f54287 a8df6b3 9f54287 2254920 a8df6b3 9f54287 a8df6b3 9c2d624 cf573f9 a8df6b3 9c2d624 cf573f9 daa6be0 a8df6b3 cf573f9 9337e18 daa6be0 9c2d624 daa6be0 9c2d624 daa6be0 9c2d624 daa6be0 cf573f9 9f54287 cf573f9 9c2d624 cf573f9 a8df6b3 bd3e774 a8df6b3 655137d a8df6b3 41d056a daa6be0 655137d daa6be0 9200a73 daa6be0 9200a73 daa6be0 9200a73 daa6be0 cf573f9 9f54287 cf573f9 9c2d624 cf573f9 a8df6b3 cf573f9 daa6be0 cf573f9 9f54287 cf573f9 9337e18 a8df6b3 12044b2 a8df6b3 9337e18 9f54287 9337e18 9c2d624 41d056a a8df6b3 41d056a 9337e18 41d056a 9f54287 41d056a 9337e18 9c2d624 cf573f9 a8df6b3 9f54287 a8df6b3 9f54287 cf573f9 9337e18 daa6be0 9337e18 daa6be0 9337e18 daa6be0 9337e18 daa6be0 cf573f9 9f54287 cf573f9 a8df6b3 cf573f9 a8df6b3 7247a32 a8df6b3 7247a32 a8df6b3 7247a32 a8df6b3 7247a32 a8df6b3 2254920 cf573f9 daa6be0 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737 738 739 740 741 742 743 744 745 746 747 748 749 750 751 752 753 754 755 756 757 758 759 760 761 762 763 764 765 766 767 768 769 770 771 772 773 774 775 776 777 778 779 |
<!DOCTYPE html>
<html lang="en">
<head>
<meta charset="UTF-8">
<meta name="viewport" content="width=device-width, initial-scale=1.0">
<meta name="description" content="DDR-Bench: A Deep Data Research Agent Benchmark for LLMs">
<title>DDR-Bench | Deep Data Research Benchmark</title>
<link rel="preconnect" href="https://fonts.googleapis.com">
<link rel="preconnect" href="https://fonts.gstatic.com" crossorigin>
<script src="https://cdn.plot.ly/plotly-2.27.0.min.js"></script>
<script src="https://cdn.jsdelivr.net/npm/marked@4.3.0/marked.min.js"></script>
<link rel="stylesheet"
href="https://cdnjs.cloudflare.com/ajax/libs/highlight.js/11.9.0/styles/atom-one-dark.min.css">
<script src="https://cdnjs.cloudflare.com/ajax/libs/highlight.js/11.9.0/highlight.min.js"></script>
<script src="https://unpkg.com/sql-formatter@4.0.2/dist/sql-formatter.min.js"></script>
<script src="data.js" defer></script>
<script src="entropy_data.js" defer></script>
<script src="trajectory_data.js" defer></script>
<script src="trajectory.js" defer></script>
<script src="benchmarking_data.js" defer></script>
<script src="benchmarking.js" defer></script>
<script src="charts.js" defer></script>
<link rel="stylesheet" href="styles.css?v=3">
<style>
/* Inline critical CSS for chart loading states */
.chart-loading {
display: flex;
align-items: center;
justify-content: center;
min-height: 300px;
color: var(--color-text-muted, #64748B);
font-size: 14px;
}
.chart-loading::after {
content: 'Loading chart...';
animation: pulse 1.5s ease-in-out infinite;
}
@keyframes pulse {
0%,
100% {
opacity: 0.4;
}
50% {
opacity: 1;
}
}
</style>
</head>
<body>
<header class="hero">
<div class="hero-content">
<!-- <h1><img src="assets/logo.png" alt="DDR-Bench Logo" class="title-logo">DDR-Bench</h1> -->
<img src="assets/social_preview.png" alt="DDR-Bench - Deep Data Research" class="hero-preview-img">
<h2>Hunt Instead of Wait: Evaluating Deep Data Research on Large Language Models</h2>
<p class="description">
We distinguish <em>investigatory intelligence</em> (autonomously setting goals and exploring) from
<em>executional intelligence</em> (completing assigned tasks), arguing that true agency requires the
former.
To evaluate this, we introduce <strong>Deep Data Research (DDR)</strong>, an open-ended task where LLMs
autonomously extract insights from databases, and <strong>DDR-Bench</strong>, a large-scale,
checklist-based benchmark enabling verifiable evaluation.
Results show that while frontier models display emerging agency, long-horizon exploration remains
challenging, with effective investigatory intelligence depending on intrinsic agentic strategies beyond
mere scaffolding or scaling.
</p>
<div class="meta-info">
<div class="meta-row authors">
<span class="meta-item">
<a href="https://thinkwee.top/about" target="_blank" rel="noopener noreferrer">Wei Liu</a>,
<a href="https://github.com/yupeijei1997" target="_blank" rel="noopener noreferrer">Peijie
Yu</a>,
<a href="https://www.kcl.ac.uk/people/michele-orini" target="_blank"
rel="noopener noreferrer">Michele Orini</a>,
<a href="https://yalidu.github.io/" target="_blank" rel="noopener noreferrer">Yali Du</a>,
<a href="https://sites.google.com/view/yulanhe/home" target="_blank"
rel="noopener noreferrer">Yulan He</a>
</span>
</div>
<div class="meta-row affiliations">
<a href="https://kclnlp.github.io/" target="_blank" rel="noopener noreferrer">
<img src="assets/kcl.svg" alt="King's College London" class="affiliation-logo kcl-logo">
</a>
<a href="https://www.tencent.com/en-us/" target="_blank" rel="noopener noreferrer">
<img src="assets/tencent.png" alt="Tencent" class="affiliation-logo">
</a>
<a href="https://www.turing.ac.uk/" target="_blank" rel="noopener noreferrer">
<img src="assets/alan.png" alt="The Alan Turing Institute" class="affiliation-logo">
</a>
</div>
<div class="meta-row links">
<a href="https://huggingface.co/collections/thinkwee/ddrbench" class="platform-btn dataset-btn"
target="_blank" rel="noopener noreferrer">
<svg viewBox="0 0 24 24" width="30" height="30" fill="none" stroke="currentColor"
stroke-width="2" stroke-linecap="round" stroke-linejoin="round">
<ellipse cx="12" cy="5" rx="9" ry="3" />
<path d="M3 5v14c0 1.66 4 3 9 3s9-1.34 9-3V5" />
<path d="M3 12c0 1.66 4 3 9 3s9-1.34 9-3" />
</svg>
Dataset
</a>
<a href="https://github.com/thinkwee/DDR_Bench" class="platform-btn github-btn" target="_blank"
rel="noopener noreferrer">
<svg viewBox="0 0 24 24" width="30" height="30" fill="currentColor">
<path
d="M12 2C6.477 2 2 6.477 2 12c0 4.42 2.865 8.17 6.839 9.49.5.092.682-.217.682-.482 0-.237-.008-.866-.013-1.7-2.782.603-3.369-1.34-3.369-1.34-.454-1.156-1.11-1.463-1.11-1.463-.908-.62.069-.608.069-.608 1.003.07 1.531 1.03 1.531 1.03.892 1.529 2.341 1.087 2.91.831.092-.646.35-1.086.636-1.336-2.22-.253-4.555-1.11-4.555-4.943 0-1.091.39-1.984 1.029-2.683-.103-.253-.446-1.27.098-2.647 0 0 .84-.269 2.75 1.025A9.578 9.578 0 0112 6.836c.85.004 1.705.114 2.504.336 1.909-1.294 2.747-1.025 2.747-1.025.546 1.377.203 2.394.1 2.647.64.699 1.028 1.592 1.028 2.683 0 3.842-2.339 4.687-4.566 4.935.359.309.678.919.678 1.852 0 1.336-.012 2.415-.012 2.743 0 .267.18.578.688.48C19.138 20.167 22 16.418 22 12c0-5.523-4.477-10-10-10z" />
</svg>
Code
</a>
<a href="https://huggingface.co/papers/2602.02039" class="platform-btn huggingface-btn"
target="_blank" rel="noopener noreferrer">
<img src="assets/hf-logo-pirate.svg" alt="HuggingFace" width="30" height="30"
class="platform-icon">
HuggingFace
</a>
<a href="https://arxiv.org/abs/2602.02039" class="platform-btn arxiv-btn" target="_blank"
rel="noopener noreferrer">
<img src="assets/arxiv-logomark-small.svg" alt="arXiv" width="30" height="30"
class="platform-icon">
arXiv
</a>
<a href="https://www.alphaxiv.org/abs/2602.02039" class="platform-btn alphaxiv-btn" target="_blank"
rel="noopener noreferrer">
<img src="assets/alphaxiv_logo.png" alt="AlphaXiv" width="30" height="30" class="platform-icon">
AlphaXiv
</a>
<a href="https://thinkwee.notion.site/ddrbench" class="platform-btn notion-btn" target="_blank"
rel="noopener noreferrer">
<svg viewBox="0 0 24 24" width="30" height="30" fill="currentColor">
<path
d="M19 3H5c-1.103 0-2 .897-2 2v14c0 1.103.897 2 2 2h14c1.103 0 2-.897 2-2V5c0-1.103-.897-2-2-2zM9 17H7.17V7H9l5.83 6.91V7H16.83v10H15L9.17 10.09V17z" />
</svg>
Notion Blog
</a>
</div>
</div>
</div>
</header>
<!-- Main Content - All sections visible -->
<main class="content">
<!-- 1. Framework Overview Section -->
<section id="framework" class="section visible framework-section">
<div class="section-header">
<h2>
<svg xmlns="http://www.w3.org/2000/svg" width="24" height="24" viewBox="0 0 24 24" fill="none"
stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round">
<rect width="18" height="18" x="3" y="3" rx="2" ry="2" />
<line x1="3" x2="21" y1="9" y2="9" />
<line x1="9" x2="9" y1="21" y2="9" />
</svg>
Framework Overview
</h2>
<p>Overview of DDR-Bench.</p>
</div>
<div class="framework-grid">
<div class="framework-card">
<div class="framework-img-wrapper">
<div class="skeleton-loader"></div>
<img src="assets/framework_task.png" alt="Task Formulation Framework" class="framework-img"
loading="lazy"
onload="this.classList.add('loaded'); this.previousElementSibling.style.display='none';">
</div>
<h3>Task Formulation</h3>
<p class="framework-description">A case of Claude Sonnet 4.5's trajectory and evaluation checklist
in the MIMIC scenario of DDR-Bench. Verified fact and supporting insights are
<u>underlined</u>. The agent is asked to perform multiple ReAct turns to explore the database
without predefined targets or queries, autonomously mine insights from the exploration.
</p>
</div>
<div class="framework-card">
<div class="framework-img-wrapper">
<div class="skeleton-loader"></div>
<img src="assets/framework_pipeline.png" alt="Evaluation Pipeline Framework"
class="framework-img" loading="lazy"
onload="this.classList.add('loaded'); this.previousElementSibling.style.display='none';">
</div>
<h3>Evaluation Pipeline</h3>
<p class="framework-description"><b>Left</b>: Compared with previous tasks, <i>DDR</i> maximises
exploration openness and agency, focusing on the direct evaluation of insight quality.
<b>Right</b>: Overview of the DDR-Bench. The checklist derived from the freeform parts of the
database is used to evaluate the agent generated insights from the exploration on the structured
parts of the database.
</p>
</div>
</div>
</section>
<!-- 1.5. Agent Trajectory Section -->
<section id="trajectory" class="section visible trajectory-section">
<div class="section-header">
<h2>
<svg xmlns="http://www.w3.org/2000/svg" width="24" height="24" viewBox="0 0 24 24" fill="none"
stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round">
<polyline points="22 12 18 12 15 21 9 3 6 12 2 12"></polyline>
</svg>
Agent Trajectory
</h2>
<p>Observe the autonomous decision-making process of the agent across different scenarios.</p>
</div>
<div class="dimension-toggle">
<button class="dim-btn active" data-traj-scenario="mimic">MIMIC</button>
<button class="dim-btn" data-traj-scenario="10k">10-K</button>
<button class="dim-btn" data-traj-scenario="globem">GLOBEM</button>
</div>
<p id="trajectory-scenario-description" class="trajectory-description">
Exploring clinical patterns and patient outcomes in a large-scale electronic health record (EHR)
database.
</p>
<div class="trajectory-container">
<div id="chat-window" class="chat-window">
<!-- Messages will be injected here via JS -->
<div class="loading-message">Loading trajectory data...</div>
</div>
<div class="scroll-hint" id="scroll-hint">
<svg xmlns="http://www.w3.org/2000/svg" width="16" height="16" viewBox="0 0 24 24" fill="none"
stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round">
<path d="M12 5v14M19 12l-7 7-7-7" />
</svg>
<span>Scroll to see more</span>
</div>
</div>
</section>
<!-- 1.75. Benchmarking Section -->
<section id="benchmarking" class="section visible benchmarking-section">
<div class="section-header">
<h2>
<svg xmlns="http://www.w3.org/2000/svg" width="24" height="24" viewBox="0 0 24 24" fill="none"
stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round">
<line x1="12" x2="12" y1="20" y2="10" />
<line x1="18" x2="18" y1="20" y2="4" />
<line x1="6" x2="6" y1="20" y2="16" />
</svg>
Benchmarking
</h2>
<p>Overall average accuracy across all scenarios and evaluation metrics.
<br>
<span class="model-badge proprietary">Purple = Proprietary</span>
<span class="model-badge opensource">Green = Open-source</span>
</p>
</div>
<div class="charts-grid single">
<div class="chart-card wide">
<div id="benchmarking-chart" class="chart-container-benchmarking"></div>
</div>
</div>
<p class="section-description">Claude 4.5 Sonnet achieves the highest overall average accuracy at 47.73%,
significantly outperforming other models. Among open-source models, DeepSeek-V3.2 leads with 38.80%,
followed closely by GLM-4.6 (37.52%) and Kimi K2 (36.42%). The results demonstrate a clear performance
gap between frontier proprietary models and open-source alternatives, though top open-source models
remain competitive with mid-tier proprietary offerings.</p>
</section>
<!-- 2. Experiment Results Section -->
<section id="results" class="section visible results-section">
<div class="section-header">
<h2>
<svg xmlns="http://www.w3.org/2000/svg" width="24" height="24" viewBox="0 0 24 24" fill="none"
stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round">
<path d="M3 3v18h18" />
<path d="m19 9-5 5-4-4-3 3" />
</svg>
Experiments
</h2>
<p>Main benchmark results and in-depth analysis of agent capabilities.</p>
</div>
<!-- Carousel Container -->
<div class="carousel-wrapper">
<button class="carousel-btn carousel-prev" aria-label="Previous">
<svg xmlns="http://www.w3.org/2000/svg" width="24" height="24" viewBox="0 0 24 24" fill="none"
stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round">
<path d="m15 18-6-6 6-6" />
</svg>
</button>
<div class="carousel-track" id="results-carousel">
<!-- 1. Overall -->
<div class="carousel-card">
<img src="assets/overall.png" alt="Overall Performance">
<h4>Overall Performance</h4>
<p class="card-caption">Systematic evaluation of mainstream LLMs across MIMIC, 10-K, and GLOBEM
datasets reveals persistent limitations in frontier models.</p>
</div>
<!-- 2. Qwen Family -->
<div class="carousel-card">
<img src="assets/qwenfamily.png" alt="Qwen Family Performance">
<h4>Training-time Factors Analysis</h4>
<p class="card-caption">Training-time factors study within the Qwen family. From left to right,
the three columns examine inference-time scaling performance across all scenarios for models
with different parameter scales, context optimisation methods, and model generations with
different training strategies.</p>
</div>
<!-- 3. Reasoning -->
<div class="carousel-card">
<img src="assets/reasoning.png" alt="Reasoning Budget">
<h4>Reasoning Budget</h4>
<p class="card-caption">Increasing the reasoning budget reduces interaction rounds but
illustrates
a
trade-off between reasoning depth and exploration efficiency.</p>
</div>
<!-- 4. Memory -->
<div class="carousel-card">
<img src="assets/memory.png" alt="Memory Mechanism">
<h4>Memory Mechanism</h4>
<p class="card-caption">Long-short-term memory can create unpredictable behavior, often
increasing
tool usage without consistently improving final accuracy.</p>
</div>
<!-- 5. Agency -->
<div class="carousel-card">
<img src="assets/agency.png" alt="Proactive vs Reactive">
<h4>Proactive vs Reactive</h4>
<p class="card-caption">Models perform significantly better with explicit queries (Reactive),
highlighting the difficulty of true proactive goal formulation.</p>
</div>
<!-- 6. Hallucination -->
<div class="carousel-card">
<img src="assets/hallucination.png" alt="Hallucination Analysis">
<h4>Hallucination Analysis</h4>
<p class="card-caption">Hallucination rates (%) across models in DDR-Bench, measured as the
proportion of insights containing factual but unfaithful information that are not derivable
from the provided inputs, which is low.</p>
</div>
<!-- 6.5 Hallucination-Accuracy Correlation -->
<div class="carousel-card">
<img src="assets/hallu_acc_corr.png" alt="Hallucination-Accuracy Correlation">
<h4>Hallucination-Accuracy Correlation</h4>
<p class="card-caption">Hallucination rates show almost no correlation with final accuracy,
indicating
robustness against metric inflation via memorization.</p>
</div>
<!-- 7. Trustworthiness -->
<div class="carousel-card">
<img src="assets/trustworthiness.png" alt="Trustworthiness">
<h4>Trustworthiness</h4>
<p class="card-caption">Verification of the LLM-as-a-Checker pipeline demonstrating high
alignment
with human expert judgments, and it is stable across multiple runs.</p>
</div>
</div>
<button class="carousel-btn carousel-next" aria-label="Next">
<svg xmlns="http://www.w3.org/2000/svg" width="24" height="24" viewBox="0 0 24 24" fill="none"
stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round">
<path d="m9 18 6-6-6-6" />
</svg>
</button>
</div>
<!-- Carousel Dots -->
<div class="carousel-dots" id="results-dots"></div>
</section>
<!-- 3. Scaling Analysis Section -->
<section id="scaling" class="section visible">
<div class="section-header">
<h2>
<svg xmlns="http://www.w3.org/2000/svg" width="24" height="24" viewBox="0 0 24 24" fill="none"
stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round">
<line x1="12" x2="12" y1="20" y2="10" />
<line x1="18" x2="18" y1="20" y2="4" />
<line x1="6" x2="6" y1="20" y2="16" />
</svg>
Scaling Analysis
</h2>
<p>Explore how model performance scales with interaction turns, token usage, and inference cost.</p>
</div>
<div class="dimension-toggle">
<button class="dim-btn active" data-dim="turn">Turns</button>
<button class="dim-btn" data-dim="token">Tokens</button>
<button class="dim-btn" data-dim="cost">Cost</button>
</div>
<div id="scaling-legend" class="shared-legend"></div>
<div class="charts-grid three-col">
<div class="chart-card">
<h3>MIMIC</h3>
<div id="scaling-mimic" class="chart-container"></div>
</div>
<div class="chart-card">
<h3>10-K</h3>
<div id="scaling-10k" class="chart-container"></div>
</div>
<div class="chart-card">
<h3>GLOBEM</h3>
<div id="scaling-globem" class="chart-container"></div>
</div>
</div>
<p class="section-description">LLMs extract more accurate insights from delaying commitment, and they
concentrate reasoning into a small number of highly valuable late-stage interactions. These targeted
interactions are built upon longer early exploration.</p>
</section>
<!-- 2. Ranking Comparison Section -->
<section id="ranking" class="section visible">
<div class="section-header">
<h2>
<svg xmlns="http://www.w3.org/2000/svg" width="24" height="24" viewBox="0 0 24 24" fill="none"
stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round">
<path d="M6 9H4.5a2.5 2.5 0 0 1 0-5H6" />
<path d="M18 9h1.5a2.5 2.5 0 0 0 0-5H18" />
<path d="M4 22h16" />
<path d="M10 14.66V17c0 .55-.47.98-.97 1.21C7.85 18.75 7 20.24 7 22" />
<path d="M14 14.66V17c0 .55.47.98.97 1.21C16.15 18.75 17 20.24 17 22" />
<path d="M18 2H6v7a6 6 0 0 0 12 0V2Z" />
</svg>
Novelty vs Accuracy
</h2>
<p>
Novelty (Bradley-Terry) vs Accuracy ranking
<br>
● = Novelty, ◇ = Accuracy.
<br>
<span class="model-badge proprietary">Purple = Proprietary</span>
<span class="model-badge opensource">Green = Open-source</span>
</p>
</div>
<div class="dimension-toggle">
<button class="dim-btn ranking-dim active" data-mode="novelty">Sort by Novelty</button>
<button class="dim-btn ranking-dim" data-mode="accuracy">Sort by Accuracy</button>
</div>
<div class="charts-grid three-col">
<div class="chart-card">
<h3>MIMIC</h3>
<div id="ranking-mimic" class="chart-container-tall"></div>
</div>
<div class="chart-card">
<h3>10-K</h3>
<div id="ranking-10k" class="chart-container-tall"></div>
</div>
<div class="chart-card">
<h3>GLOBEM</h3>
<div id="ranking-globem" class="chart-container-tall"></div>
</div>
</div>
<p class="section-description">The ranking induced by novel insight usefulness closely aligns with the
ranking based on checklist accuracy. Differences between the two rankings are small, especially among
the top-performing models.</p>
</section>
<!-- 3. Turn Distribution Section -->
<section id="turn" class="section visible">
<div class="section-header">
<h2>
<svg xmlns="http://www.w3.org/2000/svg" width="24" height="24" viewBox="0 0 24 24" fill="none"
stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round">
<path d="M21 12a9 9 0 1 1-9-9c2.52 0 4.93 1 6.74 2.74L21 8" />
<path d="M21 3v5h-5" />
</svg>
Turn Distribution
</h2>
<p>Analyze the distribution of interaction turns across different models and datasets.</p>
</div>
<div class="charts-grid three-col">
<div class="chart-card">
<h3>MIMIC</h3>
<div id="turn-mimic" class="chart-container-tall"></div>
</div>
<div class="chart-card">
<h3>10-K</h3>
<div id="turn-10k" class="chart-container-tall"></div>
</div>
<div class="chart-card">
<h3>GLOBEM</h3>
<div id="turn-globem" class="chart-container-tall"></div>
</div>
</div>
<p class="section-description">Stronger models tend to explore for more rounds without external prompting.
Knowledge-intensive databases such as 10-K and MIMIC induce more interaction rounds than signal-based
datasets such as GLOBEM, and the resulting distributions are also more uniform.</p>
</section>
<!-- 4. Entropy Analysis Section -->
<section id="entropy" class="section visible">
<div class="section-header">
<h2>
<svg xmlns="http://www.w3.org/2000/svg" width="24" height="24" viewBox="0 0 24 24" fill="none"
stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round">
<circle cx="7.5" cy="7.5" r="1.5" />
<circle cx="18.5" cy="5.5" r="1.5" />
<circle cx="11.5" cy="11.5" r="1.5" />
<circle cx="7.5" cy="16.5" r="1.5" />
<circle cx="17.5" cy="14.5" r="1.5" />
</svg>
Exploration Pattern
</h2>
<p>Scatter plot showing Access Entropy vs Coverage by model. Opacity represents accuracy. Higher entropy
= more uniform access; Higher coverage = more fields explored.</p>
</div>
<div class="dimension-toggle">
<button class="toggle-btn active" data-entropy-scenario="10k">10-K</button>
<button class="toggle-btn" data-entropy-scenario="mimic">MIMIC</button>
</div>
<div class="charts-grid three-col">
<div class="chart-card">
<h3 id="entropy-model-0-title">GPT-5.2</h3>
<div id="entropy-model-0" class="chart-container-tall"></div>
</div>
<div class="chart-card">
<h3 id="entropy-model-1-title">Claude-4.5-Sonnet</h3>
<div id="entropy-model-1" class="chart-container-tall"></div>
</div>
<div class="chart-card">
<h3 id="entropy-model-2-title">Gemini-3-Flash</h3>
<div id="entropy-model-2" class="chart-container-tall"></div>
</div>
<div class="chart-card">
<h3 id="entropy-model-3-title">GLM-4.6</h3>
<div id="entropy-model-3" class="chart-container-tall"></div>
</div>
<div class="chart-card">
<h3 id="entropy-model-4-title">Qwen3-Next-80B-A3B</h3>
<div id="entropy-model-4" class="chart-container-tall"></div>
</div>
<div class="chart-card">
<h3 id="entropy-model-5-title">DeepSeek-V3.2</h3>
<div id="entropy-model-5" class="chart-container-tall"></div>
</div>
</div>
<p class="section-description">Advanced LLMs tend to operate in a balanced exploration regime that combines
adequate coverage with focused access. Such a regime is consistently observed across different
scenarios.</p>
</section>
<!-- 5. Error Analysis Section -->
<section id="error" class="section visible">
<div class="section-header">
<h2>
<svg xmlns="http://www.w3.org/2000/svg" width="24" height="24" viewBox="0 0 24 24" fill="none"
stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round">
<path d="m21.73 18-8-14a2 2 0 0 0-3.48 0l-8 14A2 2 0 0 0 4 21h16a2 2 0 0 0 1.73-3Z" />
<line x1="12" x2="12" y1="9" y2="13" />
<line x1="12" x2="12.01" y1="17" y2="17" />
</svg>
Error Analysis
</h2>
<p>Breakdown of error types encountered during agent interactions, grouped by main categories.</p>
</div>
<div class="charts-grid single">
<div class="chart-card wide">
<div id="error-chart" class="chart-container-double"></div>
</div>
</div>
<p class="section-description">Our findings revealed that 58% of errors stemmed from insufficient
exploration, both in terms of breadth and depth. This imbalance in exploration often leads to suboptimal
results, regardless of the model’s overall capability.
Additionally, around 40% of the errors were attributed to other factors. For more powerful models,
over-reasoning was common, where the model made assumptions not fully supported by the data. In other
cases, models misinterpreted the insights, such as mistaking a downward trend for an upward one. Less
capable models, on the other hand, tended to make more fundamental errors, such as repeatedly debugging
or struggling with missing data, which could disrupt the overall coherence of the analysis.</p>
</section>
<!-- 6. Probing Results Section -->
<section id="probing" class="section visible">
<div class="section-header">
<h2>
<svg xmlns="http://www.w3.org/2000/svg" width="24" height="24" viewBox="0 0 24 24" fill="none"
stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round">
<circle cx="11" cy="11" r="8" />
<path d="m21 21-4.3-4.3" />
</svg>
Self-Termination
</h2>
<p>Analyze the willingness of models to terminate their own analysis.</p>
</div>
<div id="probing-legend" class="shared-legend"></div>
<div class="charts-grid three-col">
<div class="chart-card">
<h3>MIMIC</h3>
<div id="probing-mimic" class="chart-container-tall"></div>
</div>
<div class="chart-card">
<h3>GLOBEM</h3>
<div id="probing-globem" class="chart-container-tall"></div>
</div>
<div class="chart-card">
<h3>10-K</h3>
<div id="probing-10k" class="chart-container-tall"></div>
</div>
</div>
<p class="section-description"> Clear differences emerge across model generations. Qwen3 and Qwen3-Next
exhibit a consistently increasing probability, indicating growing confidence that a complete report can
be produced as more information is accumulated, whereas the Qwen2.5 series shows pronounced fluctuations
and remains uncertain about whether exploration can be terminated at the current step. Moreover,
Qwen3-Next maintains higher confidence with lower variance throughout, suggesting that it has more
confidence that exploration is progressing towards a more comprehensive and deeper report.</p>
</section>
</main>
<!-- Footer -->
<footer class="footer">
<p>DDR-Bench © 2026 | King's College London · Tencent · The Alan Turing Institute</p>
</footer>
<!-- Carousel Script -->
<script>
document.addEventListener('DOMContentLoaded', function () {
const track = document.getElementById('results-carousel');
const dotsContainer = document.getElementById('results-dots');
const prevBtn = document.querySelector('.carousel-prev');
const nextBtn = document.querySelector('.carousel-next');
if (!track) return;
const cards = Array.from(track.querySelectorAll('.carousel-card'));
const cardCount = cards.length;
let currentIndex = 0;
let hasStartedAutoPlay = false;
// Create dots
for (let i = 0; i < cardCount; i++) {
const dot = document.createElement('button');
dot.className = 'carousel-dot' + (i === 0 ? ' active' : '');
dot.setAttribute('aria-label', `Go to slide ${i + 1}`);
dot.addEventListener('click', () => goToSlide(i));
dotsContainer.appendChild(dot);
}
const dots = dotsContainer.querySelectorAll('.carousel-dot');
function updateCarousel() {
// Card is 66.666% width, calculate offset to center current card
const cardWidth = track.offsetWidth * 0.66666;
const gap = 32; // --space-lg
const offset = (track.offsetWidth - cardWidth) / 2 - currentIndex * (cardWidth + gap);
track.style.transform = `translateX(${offset}px)`;
// Update card states
cards.forEach((card, i) => {
card.classList.remove('active', 'side');
if (i === currentIndex) {
card.classList.add('active');
} else {
card.classList.add('side');
}
});
// Update dots
dots.forEach((dot, i) => {
dot.classList.toggle('active', i === currentIndex);
});
}
function goToSlide(index) {
// Infinite loop
if (index < 0) {
currentIndex = cardCount - 1;
} else if (index >= cardCount) {
currentIndex = 0;
} else {
currentIndex = index;
}
updateCarousel();
}
function nextSlide() { goToSlide(currentIndex + 1); }
function prevSlide() { goToSlide(currentIndex - 1); }
// Navigation buttons
prevBtn.addEventListener('click', prevSlide);
nextBtn.addEventListener('click', nextSlide);
// Keyboard navigation
document.addEventListener('keydown', (e) => {
if (e.key === 'ArrowLeft') prevSlide();
if (e.key === 'ArrowRight') nextSlide();
});
// Touch/swipe support
let touchStartX = 0;
track.addEventListener('touchstart', (e) => {
touchStartX = e.changedTouches[0].screenX;
}, { passive: true });
track.addEventListener('touchend', (e) => {
const diff = touchStartX - e.changedTouches[0].screenX;
if (Math.abs(diff) > 50) {
if (diff > 0) nextSlide();
else prevSlide();
}
}, { passive: true });
// Auto Play Logic
let autoPlayInterval;
const AUTO_PLAY_DELAY = 5000;
function startAutoPlay() {
stopAutoPlay(); // Clear existing to specify
autoPlayInterval = setInterval(nextSlide, AUTO_PLAY_DELAY);
}
function stopAutoPlay() {
if (autoPlayInterval) clearInterval(autoPlayInterval);
}
// Pause on hover
const carouselWrapper = document.querySelector('.carousel-wrapper');
if (carouselWrapper) {
carouselWrapper.addEventListener('mouseenter', stopAutoPlay);
carouselWrapper.addEventListener('mouseleave', () => {
if (hasStartedAutoPlay) startAutoPlay();
});
}
// Initialize carousel position (but don't auto-play yet)
updateCarousel();
// Use Intersection Observer to start auto-play only when visible
const resultsSection = document.getElementById('results');
if (resultsSection) {
const carouselObserver = new IntersectionObserver((entries) => {
entries.forEach(entry => {
if (entry.isIntersecting) {
if (!hasStartedAutoPlay) {
hasStartedAutoPlay = true;
}
startAutoPlay();
} else {
stopAutoPlay();
}
});
}, {
threshold: 0.3 // Start when 30% visible
});
carouselObserver.observe(resultsSection);
}
});
</script>
<!-- Scroll Hint Script -->
<script>
document.addEventListener('DOMContentLoaded', function () {
const chatWindow = document.getElementById('chat-window');
const scrollHint = document.getElementById('scroll-hint');
if (chatWindow && scrollHint) {
// Hide hint when user scrolls
chatWindow.addEventListener('scroll', function () {
if (chatWindow.scrollTop > 50) {
scrollHint.classList.add('hidden');
}
});
// Show hint initially after a short delay
setTimeout(() => {
scrollHint.style.opacity = '1';
}, 1000);
}
});
</script>
</body>
</html> |