File size: 35,548 Bytes
d00556c 193ee10 d00556c e9628be d00556c e9628be d00556c e9628be d00556c e9628be d00556c 785f414 d00556c 785f414 d00556c 8d50cf2 d00556c e9628be d00556c 193ee10 |
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 |
<!DOCTYPE html>
<html>
<head>
<meta charset="utf-8">
<meta name="description"
content="Deformable Neural Radiance Fields creates free-viewpoint portraits (nerfies) from casually captured videos.">
<meta name="keywords" content="Nerfies, D-NeRF, NeRF">
<meta name="viewport" content="width=device-width, initial-scale=1">
<title>CarBoN: Calibrated Best-of-N Sampling Improves Test-time Reasoning</title>
<!-- Global site tag (gtag.js) - Google Analytics -->
<script async src="https://www.googletagmanager.com/gtag/js?id=G-PYVRSFMDRL"></script>
<script>
window.dataLayer = window.dataLayer || [];
function gtag() {
dataLayer.push(arguments);
}
gtag('js', new Date());
gtag('config', 'G-PYVRSFMDRL');
</script>
<link href="https://fonts.googleapis.com/css?family=Google+Sans|Noto+Sans|Castoro"
rel="stylesheet">
<link rel="stylesheet" href="./static/css/bulma.min.css">
<link rel="stylesheet" href="./static/css/bulma-carousel.min.css">
<link rel="stylesheet" href="./static/css/bulma-slider.min.css">
<link rel="stylesheet" href="./static/css/fontawesome.all.min.css">
<link rel="stylesheet"
href="https://cdn.jsdelivr.net/gh/jpswalsh/academicons@1/css/academicons.min.css">
<link rel="stylesheet" href="./static/css/index.css">
<link rel="icon" href="./static/images/favicon.svg">
<script src="https://ajax.googleapis.com/ajax/libs/jquery/3.5.1/jquery.min.js"></script>
<script defer src="./static/js/fontawesome.all.min.js"></script>
<script src="./static/js/bulma-carousel.min.js"></script>
<script src="./static/js/bulma-slider.min.js"></script>
<script src="./static/js/index.js"></script>
<!-- MathJax for mathematical notation -->
<script src="https://polyfill.io/v3/polyfill.min.js?features=es6"></script>
<script id="MathJax-script" async src="https://cdn.jsdelivr.net/npm/mathjax@3/es5/tex-mml-chtml.js"></script>
</head>
<body>
<section class="hero">
<div class="hero-body">
<div class="container is-max-desktop">
<div class="columns is-centered">
<div class="column has-text-centered">
<h1 class="title is-1 publication-title">CarBoN: Calibrated Best-of-N Sampling<br>Improves Test-time Reasoning</h1>
<div class="is-size-5 publication-authors">
<span class="author-block">
<a href="https://sites.google.com/view/yungchentang" target="_blank" rel="noopener noreferrer">Yung-Chen Tang</a><sup>1,2</sup>,
</span>
<span class="author-block">
<a href="https://sites.google.com/site/pinyuchenpage/home" target="_blank" rel="noopener noreferrer">Pin-Yu Chen</a><sup>3</sup>,
</span>
<span class="author-block">
<a href="https://people.epfl.ch/andrea.cavallaro?lang=en" target="_blank" rel="noopener noreferrer">Andrea Cavallaro</a><sup>1,2</sup>
</span>
</div>
<div class="is-size-5 publication-authors">
<span class="author-block"><sup>1</sup>EPFL,</span>
<span class="author-block"><sup>2</sup>Idiap Research Institute,</span>
<span class="author-block"><sup>3</sup>IBM Research</span>
</div>
<div class="column has-text-centered">
<div class="publication-links">
<!-- arXiv Link. -->
<span class="link-block">
<a href="https://arxiv.org/abs/2510.15674" target="_blank" rel="noopener noreferrer"
class="external-link button is-normal is-rounded is-dark">
<span class="icon">
<i class="ai ai-arxiv"></i>
</span>
<span>arXiv</span>
</a>
</span>
<!-- Code Link. -->
<span class="link-block">
<a href="#"
class="external-link button is-normal is-rounded is-dark">
<span class="icon">
<i class="fab fa-github"></i>
</span>
<span>Code</span>
</a>
</span>
</div>
</div>
</div>
</div>
</div>
</div>
</section>
<section class="section">
<div class="container is-max-desktop">
<!-- Abstract. -->
<div class="columns is-centered has-text-centered">
<div class="column is-four-fifths">
<h2 class="title is-3">Abstract</h2>
<div class="content has-text-justified">
<p>
Allocating more computation during inference time (test-time scaling) improves language model performance, especially for reasoning tasks.
However, popular methods like Best-of-N sampling often show diminishing returns as N increases.
To address this inefficiency, we introduce a general <strong>test-time calibration framework</strong> that adaptively modifies the model toward high-reward reasoning paths, with theoretical guarantees of improving the lower bound of expected reward under finite sampling, all without large language model (LLM) retraining.
Within this framework, we propose <strong>CarBoN</strong> (Calibrated Best-of-N), a two-phase method that first explores the solution space and then learns a calibration of the logits via an input-specific temperature T and additive shift vector δ, guiding generation toward more reliable reasoning.
</p>
<p>
Experiments on MATH-500 and AIME-2024 show that CarBoN improves efficiency, with up to 4× fewer rollouts to reach the same accuracy, while often achieving higher accuracy under fixed budgets.
We also analyze the complementary roles of T and δ in balancing output diversity and correctness, and demonstrate that the framework also generalizes to step-level sampling strategies such as beam search.
</p>
</div>
</div>
</div>
<!--/ Abstract. -->
</div>
</section>
<section class="section">
<div class="container is-max-desktop">
<!-- Motivation. -->
<div class="columns is-centered">
<div class="column is-full-width">
<h2 class="title is-3">Why calibration for test-time scaling (TTS)? A motivating example of reward-based binary search.</h2>
<div class="content has-text-justified">
<p>
Let the task be finding a target in [0,10<sup>4</sup>].
Calibration means that before each search step the model can query <i>n</i> candidate points for reward, where <i>n</i> denotes the number of reward queries per step.
The reward is defined as the inverse distance to the target plus noise.
The baseline binary search, corresponding to naive TTS (<i>n</i>=0), requires 13.3 steps on average. Increasing <i>n</i> significantly accelerates convergence: for example, with <i>n</i>=16 the search depth is reduced by up to 74% (see Figure 1, left).
Figure 1 (right) shows an example run where calibration quickly converges to the target, while vanilla binary search continues oscillating.
This example highlights that reward feedback for calibration reshapes the sampling distribution and motivates its use for TTS.
</p>
</div>
<!-- Figure 1 -->
<div class="content has-text-centered">
<img src="./static/images/figure_1.png" alt="Reward-guided calibration accelerates binary search" style="width: 100%; height: auto; margin: 20px 0; display: block;">
<p class="has-text-justified" style="margin: 0;">
<strong>Figure 1. Reward-guided calibration accelerates binary search.</strong>
Left: Increasing per-step noisy reward (inverse-distance signal + noise) lowers average search steps versus vanilla.
Right: Example showing reward guidance converges early; vanilla keeps oscillating.
</p>
</div>
<!--/ Figure 1 -->
</div>
</div>
<!--/ Motivation. -->
</div>
</section>
<section class="section">
<div class="container is-max-desktop">
<!-- Framework Overview. -->
<div class="columns is-centered">
<div class="column is-full-width">
<h2 class="title is-3">Overview of Test-time Calibration Framework</h2>
<div class="content has-text-justified">
<p>
Building on this principle, our framework reuses sampled completions that are normally discarded in parallel sampling methods to extract reward signals and perform calibration.
Within this framework, we introduce <strong>CarBoN</strong> (Calibrated Best-of-N). Without modifying the original LLM, our framework allocates part of the budget to exploration and calibration, then focuses the remaining budget on high-scoring regions using logit calibration.
Reusing high-scoring answer selected by reward model enhances answer quality and query efficiency, under the same inference budget.
</p>
</div>
<!-- Figure 2 -->
<div class="content has-text-centered">
<img src="./static/images/figure_2.png" alt="Test-time calibration framework and MATH-500 results" style="width: 100%; height: auto; margin: 20px 0; display: block;">
<p class="has-text-justified" style="margin: 0;">
<strong>Figure 2. (a) Test-time calibration framework.</strong> With a rollout budget N = N<sub>1</sub> + N<sub>2</sub>, the model first explores by generating and scoring N<sub>1</sub> candidate responses. The model then learns calibration parameters (δ, T) from high-scoring responses, using them to adjust the logits for the remaining N<sub>2</sub> generations. The final answer is selected from all N candidates.
<strong>(b) MATH-500 Results.</strong> CarBoN improves weighted Best-of-N accuracy across four models. For all models, calibrated accuracy at N=64 (orange dash line) matches or exceeds uncalibrated accuracy at N=256, corresponding to up to a 4× reduction in rollout budgets. Notably, with Qwen2.5-Math-1.5B-Instruct at N=64, CarBoN surpasses GPT-4o (red dashed line), while uncalibrated Best-of-N with N=256 does not.
</p>
</div>
<!--/ Figure 2 -->
</div>
</div>
<!--/ Framework Overview. -->
</div>
</section>
<section class="section">
<div class="container is-max-desktop">
<!-- Theoretical Guarantees. -->
<div class="columns is-centered">
<div class="column is-full-width">
<h2 class="title is-3">Theoretical Analysis</h2>
<div class="content has-text-justified">
<p>
We provide theoretical foundations for our calibration framework through three formal results.
<strong>Lemma 1</strong> establishes the existence of joint calibration parameters \((\delta^*, T^*)\) that provably improve the model's predictive distribution when the base model is not perfectly calibrated.
<strong>Theorem 1</strong> proves that applying such calibration strictly increases the expected reward under finite Best-of-\(N\) sampling, by showing that calibration achieves first-order stochastic dominance—making high-reward outputs more likely.
<strong>Corollary 1</strong> demonstrates that discarding exploration samples and using only calibrated samples is strictly suboptimal.
Together, these results rigorously justify why test-time calibration works and why both phases are essential.
</p>
</div>
<!-- Lemma 1 -->
<details style="margin-bottom: 1rem; border: 1px solid #dbdbdb; border-radius: 6px; padding: 1rem;">
<summary style="cursor: pointer; font-weight: bold; font-size: 1.1rem; color: #3273dc;">
Lemma 1: Existence of an Improving Joint Solution \((\delta, T)\)
</summary>
<div style="margin-top: 1rem; padding-left: 1rem; border-left: 3px solid #3273dc;">
<p>
Let the joint loss function be \(\mathcal{L}(\delta, T) = \mathbb{E}_{y \sim \mathcal{D}_\text{calib}(x)} \left[ -\log p_\theta(y \mid x; \delta, T) \right]\).
Let \(\bar{p}_{\theta}\) be the model's average predictive distribution and \(\bar{p}_{\text{target}}\) be the empirical average one-hot distribution, both averaged over all generation steps in the calibration set \(\mathcal{D}_\text{calib}(x)\).
Suppose the base model is not perfectly calibrated in the sense that at least one of the following conditions holds: (1) \(\bar{p}_{\theta} \neq \bar{p}_{\text{target}}\), or (2) the average logit of ground-truth tokens does not equal the average expected logit.
</p>
<p>
Then there exists a joint solution \((\delta, T) \in \mathbb{R}^D \times (0, \infty)\), where \((\delta, T) \neq (\mathbf{0}, 1)\), such that the loss is strictly reduced:
\[\mathcal{L}(\delta, T) < \mathcal{L}(\mathbf{0}, 1)\]
</p>
</div>
</details>
<!-- Theorem 1 -->
<details style="margin-bottom: 1rem; border: 1px solid #dbdbdb; border-radius: 6px; padding: 1rem;">
<summary style="cursor: pointer; font-weight: bold; font-size: 1.1rem; color: #3273dc;">
Theorem 1: Joint Calibration \((\delta, T)\) Improves Expected Reward from Best-of-\(N\) Sampling
</summary>
<div style="margin-top: 1rem; padding-left: 1rem; border-left: 3px solid #3273dc;">
<p>
Let \(p_{\theta}(y \mid x; \delta, T)\) be the model's probability distribution over outputs \(y \in \mathcal{Y}\), parameterized by a calibration vector \(\delta\) and a temperature \(T\). Let the base model be configured with parameters \((\mathbf{0}, T_{base})\) for some \(T_{base} > 0\).
Let \(R(x,y)\) be a reward function, and assume there exists a unique output \(y^* \in \mathcal{Y}\) with a strictly maximum reward, i.e., \(r^* = R(x,y^*) > \max_{y \neq y^*} R(x,y) = r_{\text{other\_max}}\).
</p>
<p>
We consider cases where joint calibration with parameters \((\delta^*, T^*)\) improves upon the base model by increasing the probability of the unique optimal output:
\[p_{\theta}(y^* \mid x; \delta^*, T^*) > p_{\theta}(y^* \mid x; \mathbf{0}, T_{base})\]
</p>
<p>
Then, for any \(n \geq 1\) within the remaining inference budget after calibration, the lower bound on the expected best-of-\(N\) reward under the jointly calibrated model is strictly greater than that of the base model.
Specifically, let \(R_{LB}(p) = r^* - (1-p)^n (r^* - r_{\text{other\_max}})\) be a valid lower bound for the expected best-of-\(N\) reward, where \(p\) is the probability of sampling \(y^*\). The improvement in this lower bound is strictly positive:
\[\Delta_{R_{LB}}(x,n) = R_{LB}(p_{\theta}(y^* \mid x; \delta^*, T^*)) - R_{LB}(p_{\theta}(y^* \mid x; \mathbf{0}, T_{base})) > 0\]
</p>
</div>
</details>
<!-- Corollary 1 -->
<details style="margin-bottom: 1rem; border: 1px solid #dbdbdb; border-radius: 6px; padding: 1rem;">
<summary style="cursor: pointer; font-weight: bold; font-size: 1.1rem; color: #3273dc;">
Corollary 1: Sub-optimality of Exploitation Alone
</summary>
<div style="margin-top: 1rem; padding-left: 1rem; border-left: 3px solid #3273dc;">
<p>
The final candidate is selected by maximizing \(R(x, y)\) over a set of candidates \(\mathcal{Y}\). Since \(\mathcal{Y}_{\text{exploit}}\) is a subset of the union \(\mathcal{Y} = \mathcal{Y}_{\text{explore}} \cup \mathcal{Y}_{\text{exploit}}\), the strategy of only selecting from \(\mathcal{Y}_{\text{exploit}}\) is sub-optimal compared to selecting from the union.
</p>
<p>
This is because the maximum reward achievable from the union is greater than or equal to the maximum reward achievable from the exploitation set alone:
\[\max_{y \in \mathcal{Y}_{\text{explore}} \cup \mathcal{Y}_{\text{exploit}}} R(x, y) \geq \max_{y \in \mathcal{Y}_{\text{exploit}}} R(x, y)\]
</p>
<p>
Therefore, including exploration samples in the final candidate pool is essential for maximizing expected reward.
</p>
</div>
</details>
</div>
</div>
<!--/ Theoretical Guarantees. -->
</div>
</section>
<section class="section">
<div class="container is-max-desktop">
<!-- Empirical Results. -->
<div class="columns is-centered">
<div class="column is-full-width">
<h2 class="title is-3">Empirical Results</h2>
<div class="content has-text-justified">
<p>
We evaluate CarBoN on MATH-500 and AIME-2024 benchmarks across multiple language models. The results demonstrate that CarBoN consistently improves upon standard Best-of-\(N\) sampling, with calibrated accuracy at \(N=64\) matching or exceeding uncalibrated results at \(N=256\) (\(4\times\) reduction in rollout budget).
</p>
</div>
<!-- Table 1: MATH-500 Results -->
<h3 class="title is-4" style="margin-top: 2rem;">MATH-500 Results</h3>
<div style="max-width: 800px; margin: 0 auto 2rem auto;">
<p style="font-size: 0.95rem; margin-bottom: 0.5rem;">
<strong>Table 1.</strong> Accuracy (%) of four models on MATH-500, comparing Weighted Best-of-<span style='font-style:italic;'>N</span> methods before and after calibration. CarBoN enables further improvements beyond the plateau of standard Best-of-<span style='font-style:italic;'>N</span>, with calibrated accuracy at <span style='font-style:italic;'>N</span>=64 exceeding the uncalibrated results at <span style='font-style:italic;'>N</span>=256. Bold indicates better accuracy for each <span style='font-style:italic;'>N</span>.
</p>
<div style="overflow-x: auto;">
<table class="table is-bordered is-striped is-narrow is-hoverable" style="margin: 0 auto; width: 100%; min-width: 600px;">
<thead>
<tr>
<th style="text-align: center; width: 180px;">Model</th>
<th style="text-align: center; width: 110px;">Method</th>
<th style="text-align: center; width: 70px;">N=8</th>
<th style="text-align: center; width: 70px;">N=16</th>
<th style="text-align: center; width: 70px;">N=32</th>
<th style="text-align: center; width: 70px;">N=64</th>
<th style="text-align: center; width: 70px;">N=128</th>
<th style="text-align: center; width: 70px;">N=256</th>
</tr>
</thead>
<tbody>
<tr>
<td rowspan="2" style="vertical-align: middle; text-align: left; background-color: white;">Llama-3.2-1B-Instruct</td>
<td style="text-align: left; background-color: white;">Best-of-<span style='font-style:italic;'>N</span></td>
<td style="text-align: center; background-color: white;">42.0</td>
<td style="text-align: center; background-color: white;">44.6</td>
<td style="text-align: center; background-color: white;">47.8</td>
<td style="text-align: center; background-color: white;">48.6</td>
<td style="text-align: center; background-color: white;">50.6</td>
<td style="text-align: center; background-color: white;">50.8</td>
</tr>
<tr>
<td style="text-align: left; background-color: white;">CarBoN</td>
<td style="text-align: center; background-color: white;"><strong>43.0</strong></td>
<td style="text-align: center; background-color: white;"><strong>45.6</strong></td>
<td style="text-align: center; background-color: white;"><strong>48.4</strong></td>
<td style="text-align: center; background-color: white;"><strong>51.0</strong></td>
<td style="text-align: center; background-color: white;"><strong>51.8</strong></td>
<td style="text-align: center; background-color: white;"><strong>51.8</strong></td>
</tr>
<tr>
<td rowspan="2" style="vertical-align: middle; text-align: left; background-color: #f9f9f9;">Llama-3.2-3B-Instruct</td>
<td style="text-align: left; background-color: #f9f9f9;">Best-of-<span style='font-style:italic;'>N</span></td>
<td style="text-align: center; background-color: #f9f9f9;">56.8</td>
<td style="text-align: center; background-color: #f9f9f9;">58.2</td>
<td style="text-align: center; background-color: #f9f9f9;">59.6</td>
<td style="text-align: center; background-color: #f9f9f9;">61.6</td>
<td style="text-align: center; background-color: #f9f9f9;">61.8</td>
<td style="text-align: center; background-color: #f9f9f9;">62.2</td>
</tr>
<tr>
<td style="text-align: left; background-color: #f9f9f9;">CarBoN</td>
<td style="text-align: center; background-color: #f9f9f9;"><strong>57.6</strong></td>
<td style="text-align: center; background-color: #f9f9f9;"><strong>59.0</strong></td>
<td style="text-align: center; background-color: #f9f9f9;"><strong>60.8</strong></td>
<td style="text-align: center; background-color: #f9f9f9;"><strong>62.2</strong></td>
<td style="text-align: center; background-color: #f9f9f9;"><strong>63.2</strong></td>
<td style="text-align: center; background-color: #f9f9f9;"><strong>63.4</strong></td>
</tr>
<tr>
<td rowspan="2" style="vertical-align: middle; text-align: left; background-color: white;">Qwen2.5-1.5B-Instruct</td>
<td style="text-align: left; background-color: white;">Best-of-<span style='font-style:italic;'>N</span></td>
<td style="text-align: center; background-color: white;"><strong>56.4</strong></td>
<td style="text-align: center; background-color: white;">57.6</td>
<td style="text-align: center; background-color: white;">61.4</td>
<td style="text-align: center; background-color: white;">62.0</td>
<td style="text-align: center; background-color: white;">62.6</td>
<td style="text-align: center; background-color: white;">62.2</td>
</tr>
<tr>
<td style="text-align: left; background-color: white;">CarBoN</td>
<td style="text-align: center; background-color: white;">55.0</td>
<td style="text-align: center; background-color: white;"><strong>60.0</strong></td>
<td style="text-align: center; background-color: white;"><strong>61.8</strong></td>
<td style="text-align: center; background-color: white;"><strong>62.4</strong></td>
<td style="text-align: center; background-color: white;"><strong>64.0</strong></td>
<td style="text-align: center; background-color: white;"><strong>64.4</strong></td>
</tr>
<tr>
<td rowspan="2" style="vertical-align: middle; text-align: left; background-color: #f9f9f9;">Qwen2.5-Math-1.5B-Instruct</td>
<td style="text-align: left; background-color: #f9f9f9;">Best-of-<span style='font-style:italic;'>N</span></td>
<td style="text-align: center; background-color: #f9f9f9;">73.6</td>
<td style="text-align: center; background-color: #f9f9f9;">75.4</td>
<td style="text-align: center; background-color: #f9f9f9;"><strong>76.4</strong></td>
<td style="text-align: center; background-color: #f9f9f9;">75.6</td>
<td style="text-align: center; background-color: #f9f9f9;">76.4</td>
<td style="text-align: center; background-color: #f9f9f9;">76.8</td>
</tr>
<tr>
<td style="text-align: left; background-color: #f9f9f9;">CarBoN</td>
<td style="text-align: center; background-color: #f9f9f9;"><strong>74.2</strong></td>
<td style="text-align: center; background-color: #f9f9f9;"><strong>76.0</strong></td>
<td style="text-align: center; background-color: #f9f9f9;"><strong>76.4</strong></td>
<td style="text-align: center; background-color: #f9f9f9;"><strong>77.2</strong></td>
<td style="text-align: center; background-color: #f9f9f9;"><strong>77.2</strong></td>
<td style="text-align: center; background-color: #f9f9f9;"><strong>77.8</strong></td>
</tr>
</tbody>
</table>
</div>
</div>
<!-- Table 2: AIME-2024 Results -->
<h3 class="title is-4" style="margin-top: 3rem;">AIME-2024 Results</h3>
<div style="max-width: 800px; margin: 0 auto 2rem auto;">
<p style="font-size: 0.95rem; margin-bottom: 0.5rem;">
<strong>Table 2.</strong> Correct answers (out of 30) on the AIME-2024 benchmark for two math-specialized models, comparing Best-of-<span style='font-style:italic;'>N</span> and CarBoN across different rollout budgets. CarBoN enables further improvements beyond the plateau of standard Best-of-<span style='font-style:italic;'>N</span>. Bold numbers indicate the higher number of correct answers for each <span style='font-style:italic;'>N</span>.
</p>
<div style="overflow-x: auto;">
<table class="table is-bordered is-striped is-narrow is-hoverable" style="margin: 0 auto; width: 100%; min-width: 600px;">
<thead>
<tr>
<th style="text-align: center; width: 180px;">Model</th>
<th style="text-align: center; width: 110px;">Method</th>
<th style="text-align: center; width: 70px;">N=16</th>
<th style="text-align: center; width: 70px;">N=32</th>
<th style="text-align: center; width: 70px;">N=64</th>
<th style="text-align: center; width: 70px;">N=128</th>
<th style="text-align: center; width: 70px;">N=256</th>
</tr>
</thead>
<tbody>
<tr>
<td rowspan="2" style="vertical-align: middle; text-align: left; background-color: white;">Qwen2.5-Math-1.5B-Instruct</td>
<td style="text-align: left; background-color: white;">Best-of-<span style='font-style:italic;'>N</span></td>
<td style="text-align: center; background-color: white;">4/30</td>
<td style="text-align: center; background-color: white;">5/30</td>
<td style="text-align: center; background-color: white;">6/30</td>
<td style="text-align: center; background-color: white;">6/30</td>
<td style="text-align: center; background-color: white;">6/30</td>
</tr>
<tr>
<td style="text-align: left; background-color: white;">CarBoN</td>
<td style="text-align: center; background-color: white;">4/30</td>
<td style="text-align: center; background-color: white;">5/30</td>
<td style="text-align: center; background-color: white;">6/30</td>
<td style="text-align: center; background-color: white;"><strong>7/30</strong></td>
<td style="text-align: center; background-color: white;"><strong>7/30</strong></td>
</tr>
<tr>
<td rowspan="2" style="vertical-align: middle; text-align: left; background-color: #f9f9f9;">Qwen2.5-Math-7B-Instruct</td>
<td style="text-align: left; background-color: #f9f9f9;">Best-of-<span style='font-style:italic;'>N</span></td>
<td style="text-align: center; background-color: #f9f9f9;">5/30</td>
<td style="text-align: center; background-color: #f9f9f9;">5/30</td>
<td style="text-align: center; background-color: #f9f9f9;">6/30</td>
<td style="text-align: center; background-color: #f9f9f9;">6/30</td>
<td style="text-align: center; background-color: #f9f9f9;">6/30</td>
</tr>
<tr>
<td style="text-align: left; background-color: #f9f9f9;">CarBoN</td>
<td style="text-align: center; background-color: #f9f9f9;">5/30</td>
<td style="text-align: center; background-color: #f9f9f9;"><strong>6/30</strong></td>
<td style="text-align: center; background-color: #f9f9f9">6/30</td>
<td style="text-align: center; background-color: #f9f9f9">6/30</td>
<td style="text-align: center; background-color: #f9f9f9;"><strong>7/30</strong></td>
</tr>
</tbody>
</table>
</div>
</div>
<!-- Generalizing Test-time Calibration Beyond Best-of-N -->
<h3 class="title is-4" style="margin-top: 3rem;">Generalizing Test-time Calibration Beyond Best-of-<span style='font-style:italic;'>N</span></h3>
<div class="content has-text-justified" style="margin-bottom: 1rem;">
<p>
Test-time calibration is not limited to Best-of-<span style='font-style:italic;'>N</span> sampling. We show that calibrating step-level strategies such as beam search can further improve sample efficiency and accuracy. As shown below, calibrated beam search achieves similar or better performance with fewer candidates, demonstrating the generality of our approach.
</p>
</div>
<div style="max-width: 800px; margin: 0 auto 2rem auto;">
<p style="font-size: 0.95rem; margin-bottom: 0.5rem;">
<strong>Table 3.</strong> Accuracy (%) of standard and calibrated beam search on the MATH-500 benchmark. Calibrated beam search generally improves test-time reasoning performance, especially for larger <span style='font-style:italic;'>N</span>.
</p>
<div style="overflow-x: auto;">
<table class="table is-bordered is-striped is-narrow is-hoverable" style="margin: 0 auto; width: 100%; min-width: 600px;">
<thead>
<tr>
<th style="text-align: center; width: 180px;">Model</th>
<th style="text-align: center; width: 170px;">Method</th>
<th style="text-align: center; width: 70px;">N=8</th>
<th style="text-align: center; width: 70px;">N=16</th>
<th style="text-align: center; width: 70px;">N=32</th>
<th style="text-align: center; width: 70px;">N=64</th>
</tr>
</thead>
<tbody>
<tr>
<td rowspan="2" style="vertical-align: middle; text-align: left; background-color: white;">Llama-3.2-1B-Instruct</td>
<td style="text-align: left; background-color: white;">Beam Search</td>
<td style="text-align: center; background-color: white;">56.0</td>
<td style="text-align: center; background-color: white;">58.4</td>
<td style="text-align: center; background-color: white;">60.4</td>
<td style="text-align: center; background-color: white;">62.2</td>
</tr>
<tr>
<td style="text-align: left; background-color: white;">Calibrated Beam Search</td>
<td style="text-align: center; background-color: white;"><strong>57.2</strong></td>
<td style="text-align: center; background-color: white;"><strong>60.0</strong></td>
<td style="text-align: center; background-color: white;"><strong>62.2</strong></td>
<td style="text-align: center; background-color: white;"><strong>64.2</strong></td>
</tr>
<tr>
<td rowspan="2" style="vertical-align: middle; text-align: left; background-color: #f9f9f9;">Qwen2.5-Math-1.5B-Instruct</td>
<td style="text-align: left; background-color: #f9f9f9;">Beam Search</td>
<td style="text-align: center; background-color: #f9f9f9;"><strong>79.0</strong></td>
<td style="text-align: center; background-color: #f9f9f9;">79.2</td>
<td style="text-align: center; background-color: #f9f9f9;">80.2</td>
<td style="text-align: center; background-color: #f9f9f9;">81.4</td>
</tr>
<tr>
<td style="text-align: left; background-color: #f9f9f9;">Calibrated Beam Search</td>
<td style="text-align: center; background-color: #f9f9f9;">78.6</td>
<td style="text-align: center; background-color: #f9f9f9;"><strong>79.6</strong></td>
<td style="text-align: center; background-color: #f9f9f9;"><strong>81.2</strong></td>
<td style="text-align: center; background-color: #f9f9f9;"><strong>82.8</strong></td>
</tr>
</tbody>
</table>
</div>
</div>
<div class="content has-text-justified" style="margin-top: 2rem;">
<p>
For more detailed ablation studies and analysis, please refer to our paper.
</p>
</div>
</div>
</div>
<!--/ Empirical Results. -->
</div>
</section>
<section class="section" id="BibTeX">
<div class="container is-max-desktop content">
<h2 class="title">BibTeX</h2>
<pre><code>@misc{tang2025carboncalibratedbestofnsampling,
title={CarBoN: Calibrated Best-of-N Sampling Improves Test-time Reasoning},
author={Yung-Chen Tang and Pin-Yu Chen and Andrea Cavallaro},
year={2025},
eprint={2510.15674},
archivePrefix={arXiv},
primaryClass={cs.LG},
url={https://arxiv.org/abs/2510.15674},
}
</code></pre>
</div>
</section>
<footer class="footer">
<div class="container">
<div class="columns is-centered">
<div class="column is-8">
<div class="content">
<p>
This website is licensed under a <a rel="license"
href="http://creativecommons.org/licenses/by-sa/4.0/">Creative
Commons Attribution-ShareAlike 4.0 International License</a>.
</p>
<p>
This means you are free to borrow the <a
href="https://github.com/nerfies/nerfies.github.io">source code</a> of this website,
we just ask that you link back to this page in the footer.
Please remember to remove the analytics code included in the header of the website which
you do not want on your website.
</p>
</div>
</div>
</div>
</div>
</footer>
</body>
</html>
|