Update index.html
Browse files- index.html +54 -323
index.html
CHANGED
|
@@ -2,24 +2,24 @@
|
|
| 2 |
<html>
|
| 3 |
<head>
|
| 4 |
<meta charset="utf-8">
|
| 5 |
-
<meta name="description"
|
| 6 |
-
|
| 7 |
-
<meta name="keywords" content="Nerfies, D-NeRF, NeRF">
|
| 8 |
<meta name="viewport" content="width=device-width, initial-scale=1">
|
| 9 |
-
<title>
|
| 10 |
|
| 11 |
-
|
| 12 |
-
|
| 13 |
|
|
|
|
| 14 |
<link rel="stylesheet" href="./static/css/bulma.min.css">
|
| 15 |
<link rel="stylesheet" href="./static/css/bulma-carousel.min.css">
|
| 16 |
<link rel="stylesheet" href="./static/css/bulma-slider.min.css">
|
| 17 |
<link rel="stylesheet" href="./static/css/fontawesome.all.min.css">
|
| 18 |
-
<link rel="stylesheet"
|
| 19 |
-
href="https://cdn.jsdelivr.net/gh/jpswalsh/academicons@1/css/academicons.min.css">
|
| 20 |
<link rel="stylesheet" href="./static/css/index.css">
|
| 21 |
<link rel="icon" href="./static/images/favicon.svg">
|
| 22 |
|
|
|
|
| 23 |
<script src="https://ajax.googleapis.com/ajax/libs/jquery/3.5.1/jquery.min.js"></script>
|
| 24 |
<script defer src="./static/js/fontawesome.all.min.js"></script>
|
| 25 |
<script src="./static/js/bulma-carousel.min.js"></script>
|
|
@@ -28,403 +28,134 @@
|
|
| 28 |
</head>
|
| 29 |
<body>
|
| 30 |
|
|
|
|
| 31 |
<section class="hero">
|
| 32 |
<div class="hero-body">
|
| 33 |
<div class="container is-max-desktop">
|
| 34 |
<div class="columns is-centered">
|
| 35 |
<div class="column has-text-centered">
|
| 36 |
-
<h1 class="title is-1 publication-title">
|
| 37 |
<div class="is-size-5 publication-authors">
|
| 38 |
-
<span class="author-block">
|
| 39 |
-
|
| 40 |
-
<span class="author-block">
|
| 41 |
-
|
| 42 |
-
<span class="author-block">
|
| 43 |
-
|
| 44 |
-
</span>
|
| 45 |
-
<span class="author-block">
|
| 46 |
-
<a href="http://sofienbouaziz.com" target="_blank">Sofien Bouaziz</a><sup>2</sup>,
|
| 47 |
-
</span>
|
| 48 |
-
<span class="author-block">
|
| 49 |
-
<a href="https://www.danbgoldman.com" target="_blank">Dan B Goldman</a><sup>2</sup>,
|
| 50 |
-
</span>
|
| 51 |
-
<span class="author-block">
|
| 52 |
-
<a href="https://homes.cs.washington.edu/~seitz/" target="_blank">Steven M. Seitz</a><sup>1,2</sup>,
|
| 53 |
-
</span>
|
| 54 |
-
<span class="author-block">
|
| 55 |
-
<a href="http://www.ricardomartinbrualla.com" target="_blank">Ricardo Martin-Brualla</a><sup>2</sup>
|
| 56 |
-
</span>
|
| 57 |
</div>
|
| 58 |
|
| 59 |
<div class="is-size-5 publication-authors">
|
| 60 |
-
<span class="author-block"><sup>1</sup>University of
|
| 61 |
-
<span class="author-block"><sup>2</sup>
|
|
|
|
| 62 |
</div>
|
| 63 |
|
|
|
|
| 64 |
<div class="column has-text-centered">
|
| 65 |
<div class="publication-links">
|
| 66 |
-
<!-- PDF
|
| 67 |
<span class="link-block">
|
| 68 |
-
<a href="
|
| 69 |
-
|
| 70 |
-
<span class="icon">
|
| 71 |
-
<i class="fas fa-file-pdf"></i>
|
| 72 |
-
</span>
|
| 73 |
<span>Paper</span>
|
| 74 |
</a>
|
| 75 |
</span>
|
|
|
|
| 76 |
<span class="link-block">
|
| 77 |
-
<a href="https://arxiv.org/abs/
|
| 78 |
-
|
| 79 |
-
<span class="icon">
|
| 80 |
-
<i class="ai ai-arxiv"></i>
|
| 81 |
-
</span>
|
| 82 |
<span>arXiv</span>
|
| 83 |
</a>
|
| 84 |
</span>
|
| 85 |
-
<!-- Video
|
| 86 |
<span class="link-block">
|
| 87 |
-
<a href="https://
|
| 88 |
-
|
| 89 |
-
<span class="icon">
|
| 90 |
-
<i class="fab fa-youtube"></i>
|
| 91 |
-
</span>
|
| 92 |
<span>Video</span>
|
| 93 |
</a>
|
| 94 |
</span>
|
| 95 |
-
<!-- Code
|
| 96 |
<span class="link-block">
|
| 97 |
-
<a href="https://github.com/
|
| 98 |
-
|
| 99 |
-
<span class="icon">
|
| 100 |
-
<i class="fab fa-github"></i>
|
| 101 |
-
</span>
|
| 102 |
<span>Code</span>
|
| 103 |
-
|
| 104 |
</span>
|
| 105 |
-
<!-- Dataset Link. -->
|
| 106 |
-
<span class="link-block">
|
| 107 |
-
<a href="https://github.com/google/nerfies/releases/tag/0.1" target="_blank"
|
| 108 |
-
class="external-link button is-normal is-rounded is-dark">
|
| 109 |
-
<span class="icon">
|
| 110 |
-
<i class="far fa-images"></i>
|
| 111 |
-
</span>
|
| 112 |
-
<span>Data</span>
|
| 113 |
-
</a>
|
| 114 |
</div>
|
| 115 |
-
|
| 116 |
</div>
|
|
|
|
| 117 |
</div>
|
| 118 |
</div>
|
| 119 |
</div>
|
| 120 |
</div>
|
| 121 |
</section>
|
| 122 |
|
|
|
|
| 123 |
<section class="hero teaser">
|
| 124 |
<div class="container is-max-desktop">
|
| 125 |
<div class="hero-body">
|
| 126 |
<video id="teaser" autoplay muted loop playsinline height="100%">
|
| 127 |
-
<source src="./static/videos/
|
| 128 |
-
type="video/mp4">
|
| 129 |
</video>
|
| 130 |
<h2 class="subtitle has-text-centered">
|
| 131 |
-
<span class="dnerf">
|
| 132 |
-
free-viewpoint
|
| 133 |
-
portraits.
|
| 134 |
</h2>
|
| 135 |
</div>
|
| 136 |
</div>
|
| 137 |
</section>
|
| 138 |
|
| 139 |
-
|
| 140 |
-
<section class="hero is-light is-small">
|
| 141 |
-
<div class="hero-body">
|
| 142 |
-
<div class="container">
|
| 143 |
-
<div id="results-carousel" class="carousel results-carousel">
|
| 144 |
-
<div class="item item-steve">
|
| 145 |
-
<video poster="" id="steve" autoplay controls muted loop playsinline height="100%">
|
| 146 |
-
<source src="./static/videos/steve.mp4"
|
| 147 |
-
type="video/mp4">
|
| 148 |
-
</video>
|
| 149 |
-
</div>
|
| 150 |
-
<div class="item item-chair-tp">
|
| 151 |
-
<video poster="" id="chair-tp" autoplay controls muted loop playsinline height="100%">
|
| 152 |
-
<source src="./static/videos/chair-tp.mp4"
|
| 153 |
-
type="video/mp4">
|
| 154 |
-
</video>
|
| 155 |
-
</div>
|
| 156 |
-
<div class="item item-shiba">
|
| 157 |
-
<video poster="" id="shiba" autoplay controls muted loop playsinline height="100%">
|
| 158 |
-
<source src="./static/videos/shiba.mp4"
|
| 159 |
-
type="video/mp4">
|
| 160 |
-
</video>
|
| 161 |
-
</div>
|
| 162 |
-
<div class="item item-fullbody">
|
| 163 |
-
<video poster="" id="fullbody" autoplay controls muted loop playsinline height="100%">
|
| 164 |
-
<source src="./static/videos/fullbody.mp4"
|
| 165 |
-
type="video/mp4">
|
| 166 |
-
</video>
|
| 167 |
-
</div>
|
| 168 |
-
<div class="item item-blueshirt">
|
| 169 |
-
<video poster="" id="blueshirt" autoplay controls muted loop playsinline height="100%">
|
| 170 |
-
<source src="./static/videos/blueshirt.mp4"
|
| 171 |
-
type="video/mp4">
|
| 172 |
-
</video>
|
| 173 |
-
</div>
|
| 174 |
-
<div class="item item-mask">
|
| 175 |
-
<video poster="" id="mask" autoplay controls muted loop playsinline height="100%">
|
| 176 |
-
<source src="./static/videos/mask.mp4"
|
| 177 |
-
type="video/mp4">
|
| 178 |
-
</video>
|
| 179 |
-
</div>
|
| 180 |
-
<div class="item item-coffee">
|
| 181 |
-
<video poster="" id="coffee" autoplay controls muted loop playsinline height="100%">
|
| 182 |
-
<source src="./static/videos/coffee.mp4"
|
| 183 |
-
type="video/mp4">
|
| 184 |
-
</video>
|
| 185 |
-
</div>
|
| 186 |
-
<div class="item item-toby">
|
| 187 |
-
<video poster="" id="toby" autoplay controls muted loop playsinline height="100%">
|
| 188 |
-
<source src="./static/videos/toby2.mp4"
|
| 189 |
-
type="video/mp4">
|
| 190 |
-
</video>
|
| 191 |
-
</div>
|
| 192 |
-
</div>
|
| 193 |
-
</div>
|
| 194 |
-
</div>
|
| 195 |
-
</section>
|
| 196 |
-
|
| 197 |
-
|
| 198 |
<section class="section">
|
| 199 |
<div class="container is-max-desktop">
|
| 200 |
-
<!-- Abstract. -->
|
| 201 |
<div class="columns is-centered has-text-centered">
|
| 202 |
<div class="column is-four-fifths">
|
| 203 |
<h2 class="title is-3">Abstract</h2>
|
| 204 |
<div class="content has-text-justified">
|
| 205 |
-
<p>
|
| 206 |
-
We present the first method capable of photorealistically reconstructing a non-rigidly
|
| 207 |
-
deforming scene using photos/videos captured casually from mobile phones.
|
| 208 |
-
</p>
|
| 209 |
-
<p>
|
| 210 |
-
Our approach augments neural radiance fields
|
| 211 |
-
(NeRF) by optimizing an
|
| 212 |
-
additional continuous volumetric deformation field that warps each observed point into a
|
| 213 |
-
canonical 5D NeRF.
|
| 214 |
-
We observe that these NeRF-like deformation fields are prone to local minima, and
|
| 215 |
-
propose a coarse-to-fine optimization method for coordinate-based models that allows for
|
| 216 |
-
more robust optimization.
|
| 217 |
-
By adapting principles from geometry processing and physical simulation to NeRF-like
|
| 218 |
-
models, we propose an elastic regularization of the deformation field that further
|
| 219 |
-
improves robustness.
|
| 220 |
-
</p>
|
| 221 |
-
<p>
|
| 222 |
-
We show that <span class="dnerf">Nerfies</span> can turn casually captured selfie
|
| 223 |
-
photos/videos into deformable NeRF
|
| 224 |
-
models that allow for photorealistic renderings of the subject from arbitrary
|
| 225 |
-
viewpoints, which we dub <i>"nerfies"</i>. We evaluate our method by collecting data
|
| 226 |
-
using a
|
| 227 |
-
rig with two mobile phones that take time-synchronized photos, yielding train/validation
|
| 228 |
-
images of the same pose at different viewpoints. We show that our method faithfully
|
| 229 |
-
reconstructs non-rigidly deforming scenes and reproduces unseen views with high
|
| 230 |
-
fidelity.
|
| 231 |
-
</p>
|
| 232 |
</div>
|
| 233 |
</div>
|
| 234 |
</div>
|
| 235 |
-
<!--/ Abstract. -->
|
| 236 |
|
| 237 |
-
<!-- Paper video
|
| 238 |
<div class="columns is-centered has-text-centered">
|
| 239 |
<div class="column is-four-fifths">
|
| 240 |
<h2 class="title is-3">Video</h2>
|
| 241 |
<div class="publication-video">
|
| 242 |
-
<iframe src="https://www.youtube.com/embed/
|
| 243 |
-
frameborder="0" allow="autoplay; encrypted-media" allowfullscreen></iframe>
|
| 244 |
-
</div>
|
| 245 |
-
</div>
|
| 246 |
-
</div>
|
| 247 |
-
<!--/ Paper video. -->
|
| 248 |
-
</div>
|
| 249 |
-
</section>
|
| 250 |
-
|
| 251 |
-
|
| 252 |
-
<section class="section">
|
| 253 |
-
<div class="container is-max-desktop">
|
| 254 |
-
|
| 255 |
-
<div class="columns is-centered">
|
| 256 |
-
|
| 257 |
-
<!-- Visual Effects. -->
|
| 258 |
-
<div class="column">
|
| 259 |
-
<div class="content">
|
| 260 |
-
<h2 class="title is-3">Visual Effects</h2>
|
| 261 |
-
<p>
|
| 262 |
-
Using <i>nerfies</i> you can create fun visual effects. This Dolly zoom effect
|
| 263 |
-
would be impossible without nerfies since it would require going through a wall.
|
| 264 |
-
</p>
|
| 265 |
-
<video id="dollyzoom" autoplay controls muted loop playsinline height="100%">
|
| 266 |
-
<source src="./static/videos/dollyzoom-stacked.mp4"
|
| 267 |
-
type="video/mp4">
|
| 268 |
-
</video>
|
| 269 |
-
</div>
|
| 270 |
-
</div>
|
| 271 |
-
<!--/ Visual Effects. -->
|
| 272 |
-
|
| 273 |
-
<!-- Matting. -->
|
| 274 |
-
<div class="column">
|
| 275 |
-
<h2 class="title is-3">Matting</h2>
|
| 276 |
-
<div class="columns is-centered">
|
| 277 |
-
<div class="column content">
|
| 278 |
-
<p>
|
| 279 |
-
As a byproduct of our method, we can also solve the matting problem by ignoring
|
| 280 |
-
samples that fall outside of a bounding box during rendering.
|
| 281 |
-
</p>
|
| 282 |
-
<video id="matting-video" controls playsinline height="100%">
|
| 283 |
-
<source src="./static/videos/matting.mp4"
|
| 284 |
-
type="video/mp4">
|
| 285 |
-
</video>
|
| 286 |
-
</div>
|
| 287 |
-
|
| 288 |
-
</div>
|
| 289 |
-
</div>
|
| 290 |
-
</div>
|
| 291 |
-
<!--/ Matting. -->
|
| 292 |
-
|
| 293 |
-
<!-- Animation. -->
|
| 294 |
-
<div class="columns is-centered">
|
| 295 |
-
<div class="column is-full-width">
|
| 296 |
-
<h2 class="title is-3">Animation</h2>
|
| 297 |
-
|
| 298 |
-
<!-- Interpolating. -->
|
| 299 |
-
<h3 class="title is-4">Interpolating states</h3>
|
| 300 |
-
<div class="content has-text-justified">
|
| 301 |
-
<p>
|
| 302 |
-
We can also animate the scene by interpolating the deformation latent codes of two input
|
| 303 |
-
frames. Use the slider here to linearly interpolate between the left frame and the right
|
| 304 |
-
frame.
|
| 305 |
-
</p>
|
| 306 |
-
</div>
|
| 307 |
-
<div class="columns is-vcentered interpolation-panel">
|
| 308 |
-
<div class="column is-3 has-text-centered">
|
| 309 |
-
<img src="./static/images/interpolate_start.jpg"
|
| 310 |
-
class="interpolation-image"
|
| 311 |
-
alt="Interpolate start reference image."/>
|
| 312 |
-
<p>Start Frame</p>
|
| 313 |
-
</div>
|
| 314 |
-
<div class="column interpolation-video-column">
|
| 315 |
-
<div id="interpolation-image-wrapper">
|
| 316 |
-
Loading...
|
| 317 |
-
</div>
|
| 318 |
-
<input class="slider is-fullwidth is-large is-info"
|
| 319 |
-
id="interpolation-slider"
|
| 320 |
-
step="1" min="0" max="100" value="0" type="range">
|
| 321 |
-
</div>
|
| 322 |
-
<div class="column is-3 has-text-centered">
|
| 323 |
-
<img src="./static/images/interpolate_end.jpg"
|
| 324 |
-
class="interpolation-image"
|
| 325 |
-
alt="Interpolation end reference image."/>
|
| 326 |
-
<p class="is-bold">End Frame</p>
|
| 327 |
-
</div>
|
| 328 |
-
</div>
|
| 329 |
-
<br/>
|
| 330 |
-
<!--/ Interpolating. -->
|
| 331 |
-
|
| 332 |
-
<!-- Re-rendering. -->
|
| 333 |
-
<h3 class="title is-4">Re-rendering the input video</h3>
|
| 334 |
-
<div class="content has-text-justified">
|
| 335 |
-
<p>
|
| 336 |
-
Using <span class="dnerf">Nerfies</span>, you can re-render a video from a novel
|
| 337 |
-
viewpoint such as a stabilized camera by playing back the training deformations.
|
| 338 |
-
</p>
|
| 339 |
-
</div>
|
| 340 |
-
<div class="content has-text-centered">
|
| 341 |
-
<video id="replay-video"
|
| 342 |
-
controls
|
| 343 |
-
muted
|
| 344 |
-
preload
|
| 345 |
-
playsinline
|
| 346 |
-
width="75%">
|
| 347 |
-
<source src="./static/videos/replay.mp4"
|
| 348 |
-
type="video/mp4">
|
| 349 |
-
</video>
|
| 350 |
-
</div>
|
| 351 |
-
<!--/ Re-rendering. -->
|
| 352 |
-
|
| 353 |
-
</div>
|
| 354 |
-
</div>
|
| 355 |
-
<!--/ Animation. -->
|
| 356 |
-
|
| 357 |
-
|
| 358 |
-
<!-- Concurrent Work. -->
|
| 359 |
-
<div class="columns is-centered">
|
| 360 |
-
<div class="column is-full-width">
|
| 361 |
-
<h2 class="title is-3">Related Links</h2>
|
| 362 |
-
|
| 363 |
-
<div class="content has-text-justified">
|
| 364 |
-
<p>
|
| 365 |
-
There's a lot of excellent work that was introduced around the same time as ours.
|
| 366 |
-
</p>
|
| 367 |
-
<p>
|
| 368 |
-
<a href="https://arxiv.org/abs/2104.09125" target="_blank">Progressive Encoding for Neural Optimization</a> introduces an idea similar to our windowed position encoding for coarse-to-fine optimization.
|
| 369 |
-
</p>
|
| 370 |
-
<p>
|
| 371 |
-
<a href="https://www.albertpumarola.com/research/D-NeRF/index.html" target="_blank">D-NeRF</a> and <a href="https://gvv.mpi-inf.mpg.de/projects/nonrigid_nerf/" target="_blank">NR-NeRF</a>
|
| 372 |
-
both use deformation fields to model non-rigid scenes.
|
| 373 |
-
</p>
|
| 374 |
-
<p>
|
| 375 |
-
Some works model videos with a NeRF by directly modulating the density, such as <a href="https://video-nerf.github.io/" target="_blank">Video-NeRF</a>, <a href="https://www.cs.cornell.edu/~zl548/NSFF/" target="_blank">NSFF</a>, and <a href="https://neural-3d-video.github.io/" target="_blank">DyNeRF</a>
|
| 376 |
-
</p>
|
| 377 |
-
<p>
|
| 378 |
-
There are probably many more by the time you are reading this. Check out <a href="https://dellaert.github.io/NeRF/" target="_blank">Frank Dellart's survey on recent NeRF papers</a>, and <a href="https://github.com/yenchenlin/awesome-NeRF" target="_blank">Yen-Chen Lin's curated list of NeRF papers</a>.
|
| 379 |
-
</p>
|
| 380 |
</div>
|
| 381 |
</div>
|
| 382 |
</div>
|
| 383 |
-
<!--/ Concurrent Work. -->
|
| 384 |
-
|
| 385 |
</div>
|
| 386 |
</section>
|
| 387 |
|
|
|
|
|
|
|
| 388 |
|
| 389 |
<section class="section" id="BibTeX">
|
| 390 |
<div class="container is-max-desktop content">
|
| 391 |
<h2 class="title">BibTeX</h2>
|
| 392 |
-
<pre><code>@article{
|
| 393 |
-
author = {
|
| 394 |
-
title = {
|
| 395 |
-
journal = {
|
| 396 |
-
year = {
|
|
|
|
| 397 |
}</code></pre>
|
| 398 |
</div>
|
| 399 |
</section>
|
| 400 |
|
| 401 |
-
|
| 402 |
<footer class="footer">
|
| 403 |
<div class="container">
|
| 404 |
<div class="content has-text-centered">
|
| 405 |
-
<a class="icon-link" target="_blank"
|
| 406 |
-
|
| 407 |
-
<i class="fas fa-file-pdf"></i>
|
| 408 |
-
</a>
|
| 409 |
-
<a class="icon-link" href="https://github.com/keunhong" target="_blank" class="external-link" disabled>
|
| 410 |
-
<i class="fab fa-github"></i>
|
| 411 |
-
</a>
|
| 412 |
</div>
|
| 413 |
<div class="columns is-centered">
|
| 414 |
<div class="column is-8">
|
| 415 |
<div class="content">
|
| 416 |
-
<p>
|
| 417 |
-
This website is licensed under a <a rel="license" target="_blank"
|
| 418 |
-
href="http://creativecommons.org/licenses/by-sa/4.0/">Creative
|
| 419 |
-
Commons Attribution-ShareAlike 4.0 International License</a>.
|
| 420 |
-
</p>
|
| 421 |
-
<p>
|
| 422 |
-
This means you are free to borrow the <a target="_blank"
|
| 423 |
-
href="https://github.com/nerfies/nerfies.github.io">source code</a> of this website,
|
| 424 |
-
we just ask that you link back to this page in the footer.
|
| 425 |
-
Please remember to remove the analytics code included in the header of the website which
|
| 426 |
-
you do not want on your website.
|
| 427 |
-
</p>
|
| 428 |
</div>
|
| 429 |
</div>
|
| 430 |
</div>
|
|
|
|
| 2 |
<html>
|
| 3 |
<head>
|
| 4 |
<meta charset="utf-8">
|
| 5 |
+
<meta name="description" content="Beyond ‘Aha!’: Toward Systematic Meta‑Abilities Alignment in Large Reasoning Models introduces a three‑stage recipe that explicitly teaches deduction, induction, and abduction to large language models, delivering state‑of‑the‑art reasoning performance.">
|
| 6 |
+
<meta name="keywords" content="Meta‑Abilities, Deduction, Induction, Abduction, Reinforcement Learning, Large Reasoning Models, LLM, Chain‑of‑Thought">
|
|
|
|
| 7 |
<meta name="viewport" content="width=device-width, initial-scale=1">
|
| 8 |
+
<title>Beyond ‘Aha!’: Meta‑Ability Alignment for Large Reasoning Models</title>
|
| 9 |
|
| 10 |
+
<!-- Fonts & Icons -->
|
| 11 |
+
<link href="https://fonts.googleapis.com/css?family=Google+Sans|Noto+Sans|Castoro" rel="stylesheet">
|
| 12 |
|
| 13 |
+
<!-- Bulma & Theme Assets (keep original Nerfies paths / copy to your static folder) -->
|
| 14 |
<link rel="stylesheet" href="./static/css/bulma.min.css">
|
| 15 |
<link rel="stylesheet" href="./static/css/bulma-carousel.min.css">
|
| 16 |
<link rel="stylesheet" href="./static/css/bulma-slider.min.css">
|
| 17 |
<link rel="stylesheet" href="./static/css/fontawesome.all.min.css">
|
| 18 |
+
<link rel="stylesheet" href="https://cdn.jsdelivr.net/gh/jpswalsh/academicons@1/css/academicons.min.css">
|
|
|
|
| 19 |
<link rel="stylesheet" href="./static/css/index.css">
|
| 20 |
<link rel="icon" href="./static/images/favicon.svg">
|
| 21 |
|
| 22 |
+
<!-- JS -->
|
| 23 |
<script src="https://ajax.googleapis.com/ajax/libs/jquery/3.5.1/jquery.min.js"></script>
|
| 24 |
<script defer src="./static/js/fontawesome.all.min.js"></script>
|
| 25 |
<script src="./static/js/bulma-carousel.min.js"></script>
|
|
|
|
| 28 |
</head>
|
| 29 |
<body>
|
| 30 |
|
| 31 |
+
<!-- ────────────────────────────── HERO ───────────────────────────── → -->
|
| 32 |
<section class="hero">
|
| 33 |
<div class="hero-body">
|
| 34 |
<div class="container is-max-desktop">
|
| 35 |
<div class="columns is-centered">
|
| 36 |
<div class="column has-text-centered">
|
| 37 |
+
<h1 class="title is-1 publication-title">Beyond ‘Aha!’: Toward Systematic Meta‑Abilities Alignment in Large Reasoning Models</h1>
|
| 38 |
<div class="is-size-5 publication-authors">
|
| 39 |
+
<span class="author-block"><a href="#" target="_blank">Zhiyuan Hu</a><sup>1*</sup>,</span>
|
| 40 |
+
<span class="author-block"><a href="#" target="_blank">Yibo Wang</a><sup>2</sup>,</span>
|
| 41 |
+
<span class="author-block"><a href="#" target="_blank">Hanze Dong</a><sup>3</sup>,</span>
|
| 42 |
+
<span class="author-block"><a href="#" target="_blank">Yuhui Xu</a><sup>3</sup>,</span>
|
| 43 |
+
<span class="author-block"><a href="#" target="_blank">Amrita Saha</a><sup>3</sup>,</span>
|
| 44 |
+
<span class="author-block"><a href="#" target="_blank">Caiming Xiong</a><sup>3</sup>,</span>
|
| 45 |
+
<span class="author-block"><a href="#" target="_blank">Bryan Hooi</a><sup>1†</sup>,</span>
|
| 46 |
+
<span class="author-block"><a href="#" target="_blank">Junnan Li</a><sup>3†</sup></span>
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 47 |
</div>
|
| 48 |
|
| 49 |
<div class="is-size-5 publication-authors">
|
| 50 |
+
<span class="author-block"><sup>1</sup>National University of Singapore,</span>
|
| 51 |
+
<span class="author-block"><sup>2</sup>Tsinghua University,</span>
|
| 52 |
+
<span class="author-block"><sup>3</sup>Salesforce AI Research</span>
|
| 53 |
</div>
|
| 54 |
|
| 55 |
+
<!-- Links → -->
|
| 56 |
<div class="column has-text-centered">
|
| 57 |
<div class="publication-links">
|
| 58 |
+
<!-- PDF -->
|
| 59 |
<span class="link-block">
|
| 60 |
+
<a href="./static/pdfs/meta_ability_alignment.pdf" target="_blank" class="external-link button is-normal is-rounded is-dark">
|
| 61 |
+
<span class="icon"><i class="fas fa-file-pdf"></i></span>
|
|
|
|
|
|
|
|
|
|
| 62 |
<span>Paper</span>
|
| 63 |
</a>
|
| 64 |
</span>
|
| 65 |
+
<!-- arXiv -->
|
| 66 |
<span class="link-block">
|
| 67 |
+
<a href="https://arxiv.org/abs/2505.00000" target="_blank" class="external-link button is-normal is-rounded is-dark">
|
| 68 |
+
<span class="icon"><i class="ai ai-arxiv"></i></span>
|
|
|
|
|
|
|
|
|
|
| 69 |
<span>arXiv</span>
|
| 70 |
</a>
|
| 71 |
</span>
|
| 72 |
+
<!-- Video (optional) -->
|
| 73 |
<span class="link-block">
|
| 74 |
+
<a href="https://youtu.be/your_video" target="_blank" class="external-link button is-normal is-rounded is-dark">
|
| 75 |
+
<span class="icon"><i class="fab fa-youtube"></i></span>
|
|
|
|
|
|
|
|
|
|
| 76 |
<span>Video</span>
|
| 77 |
</a>
|
| 78 |
</span>
|
| 79 |
+
<!-- Code -->
|
| 80 |
<span class="link-block">
|
| 81 |
+
<a href="https://github.com/zhiyuanhubj/Meta-Ability-Alignment" target="_blank" class="external-link button is-normal is-rounded is-dark">
|
| 82 |
+
<span class="icon"><i class="fab fa-github"></i></span>
|
|
|
|
|
|
|
|
|
|
| 83 |
<span>Code</span>
|
| 84 |
+
</a>
|
| 85 |
</span>
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 86 |
</div>
|
|
|
|
| 87 |
</div>
|
| 88 |
+
<!--/ Links -->
|
| 89 |
</div>
|
| 90 |
</div>
|
| 91 |
</div>
|
| 92 |
</div>
|
| 93 |
</section>
|
| 94 |
|
| 95 |
+
<!-- ────────────────────────────── TEASER ──────────────────────────── → -->
|
| 96 |
<section class="hero teaser">
|
| 97 |
<div class="container is-max-desktop">
|
| 98 |
<div class="hero-body">
|
| 99 |
<video id="teaser" autoplay muted loop playsinline height="100%">
|
| 100 |
+
<source src="./static/videos/meta_teaser.mp4" type="video/mp4">
|
|
|
|
| 101 |
</video>
|
| 102 |
<h2 class="subtitle has-text-centered">
|
| 103 |
+
<span class="dnerf">Meta‑Ability Alignment</span> systematically teaches deduction, induction, and abduction to large language models, turning unpredictable “aha” moments into reliable reasoning capabilities.
|
|
|
|
|
|
|
| 104 |
</h2>
|
| 105 |
</div>
|
| 106 |
</div>
|
| 107 |
</section>
|
| 108 |
|
| 109 |
+
<!-- ────────────────────────────── ABSTRACT ─────────────────────────── → -->
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 110 |
<section class="section">
|
| 111 |
<div class="container is-max-desktop">
|
|
|
|
| 112 |
<div class="columns is-centered has-text-centered">
|
| 113 |
<div class="column is-four-fifths">
|
| 114 |
<h2 class="title is-3">Abstract</h2>
|
| 115 |
<div class="content has-text-justified">
|
| 116 |
+
<p>Large reasoning models (LRMs) possess a latent capacity for long chain‑of‑thought reasoning, but emergent “aha” behaviors are unpredictable and hard to control. We introduce an explicit <em>Meta‑Ability Alignment</em> strategy that separately trains deduction, induction, and abduction specialists on self‑verifiable tasks, then merges them in parameter space and continues domain‑specific RL. This three‑stage recipe boosts performance by >10% over instruction‑tuned baselines and lifts the attainable ceiling after downstream RL, yielding consistent gains across math, coding, and science benchmarks.</p>
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 117 |
</div>
|
| 118 |
</div>
|
| 119 |
</div>
|
|
|
|
| 120 |
|
| 121 |
+
<!-- Paper video (optional) → -->
|
| 122 |
<div class="columns is-centered has-text-centered">
|
| 123 |
<div class="column is-four-fifths">
|
| 124 |
<h2 class="title is-3">Video</h2>
|
| 125 |
<div class="publication-video">
|
| 126 |
+
<iframe src="https://www.youtube.com/embed/your_video?rel=0&showinfo=0" frameborder="0" allow="autoplay; encrypted-media" allowfullscreen></iframe>
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 127 |
</div>
|
| 128 |
</div>
|
| 129 |
</div>
|
|
|
|
|
|
|
| 130 |
</div>
|
| 131 |
</section>
|
| 132 |
|
| 133 |
+
<!-- ────────────────────────────── PIPELINE IMAGE / CAROUSEL (OPTIONAL) ───────────────────────────── → -->
|
| 134 |
+
<!-- You can replace this section with evaluation figures or a carousel of task demonstrations. → -->
|
| 135 |
|
| 136 |
<section class="section" id="BibTeX">
|
| 137 |
<div class="container is-max-desktop content">
|
| 138 |
<h2 class="title">BibTeX</h2>
|
| 139 |
+
<pre><code>@article{hu2025metaability,
|
| 140 |
+
author = {Hu, Zhiyuan and Wang, Yibo and Dong, Hanze and Xu, Yuhui and Saha, Amrita and Xiong, Caiming and Hooi, Bryan and Li, Junnan},
|
| 141 |
+
title = {Beyond \`Aha!\': Toward Systematic Meta‑Abilities Alignment in Large Reasoning Models},
|
| 142 |
+
journal = {NeurIPS},
|
| 143 |
+
year = {2025},
|
| 144 |
+
url = {https://arxiv.org/abs/2505.00000}
|
| 145 |
}</code></pre>
|
| 146 |
</div>
|
| 147 |
</section>
|
| 148 |
|
|
|
|
| 149 |
<footer class="footer">
|
| 150 |
<div class="container">
|
| 151 |
<div class="content has-text-centered">
|
| 152 |
+
<a class="icon-link" target="_blank" href="./static/pdfs/meta_ability_alignment.pdf"><i class="fas fa-file-pdf"></i></a>
|
| 153 |
+
<a class="icon-link" href="https://github.com/zhiyuanhubj/Meta-Ability-Alignment" target="_blank"><i class="fab fa-github"></i></a>
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 154 |
</div>
|
| 155 |
<div class="columns is-centered">
|
| 156 |
<div class="column is-8">
|
| 157 |
<div class="content">
|
| 158 |
+
<p>This website template is adapted from the <a target="_blank" href="https://nerfies.github.io">Nerfies</a> project page and is licensed under a <a rel="license" target="_blank" href="http://creativecommons.org/licenses/by-sa/4.0/">CC BY‑SA 4.0</a> License. Please attribute appropriately.</p>
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 159 |
</div>
|
| 160 |
</div>
|
| 161 |
</div>
|