Update index.html
Browse files- index.html +112 -139
index.html
CHANGED
@@ -1,166 +1,139 @@
|
|
1 |
<!DOCTYPE html>
|
2 |
-
<html>
|
3 |
<head>
|
4 |
-
<meta charset="utf-8"
|
5 |
-
<meta name="
|
6 |
-
<meta name="
|
7 |
-
<meta name="
|
8 |
-
<title>Beyond
|
9 |
|
10 |
-
|
11 |
-
<link href="
|
|
|
|
|
|
|
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>
|
26 |
-
<script src="./static/js/bulma-slider.min.js"></script>
|
27 |
-
<script src="./static/js/index.js"></script>
|
28 |
</head>
|
29 |
<body>
|
30 |
-
|
31 |
-
|
32 |
-
<
|
33 |
-
|
34 |
-
|
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 |
-
<
|
58 |
-
|
59 |
-
|
60 |
-
|
61 |
-
|
62 |
-
|
63 |
-
|
64 |
-
</span>
|
65 |
-
|
66 |
-
<span class="
|
67 |
-
|
68 |
-
|
69 |
-
|
70 |
-
|
71 |
-
</span>
|
72 |
-
|
73 |
-
|
74 |
-
|
75 |
-
|
76 |
-
|
77 |
-
|
78 |
-
|
79 |
-
|
80 |
-
|
81 |
-
|
82 |
-
|
83 |
-
|
84 |
-
|
85 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
86 |
</div>
|
87 |
</div>
|
88 |
-
<!--/ Links -->
|
89 |
</div>
|
90 |
</div>
|
91 |
</div>
|
92 |
-
</
|
93 |
-
</section>
|
94 |
|
95 |
-
<!--
|
96 |
-
<section class="
|
97 |
-
|
98 |
-
|
99 |
-
|
100 |
-
|
101 |
-
|
102 |
-
|
103 |
-
|
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 |
-
|
122 |
-
|
123 |
-
|
124 |
-
|
125 |
-
|
126 |
-
|
127 |
-
|
128 |
-
</
|
|
|
|
|
|
|
|
|
|
|
|
|
129 |
</div>
|
130 |
-
</
|
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 |
-
|
137 |
-
<
|
138 |
-
<
|
139 |
-
|
140 |
-
|
141 |
-
|
142 |
-
|
143 |
-
|
144 |
-
|
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 |
-
|
156 |
-
|
157 |
-
|
158 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
159 |
</div>
|
160 |
</div>
|
161 |
</div>
|
162 |
-
</
|
163 |
-
</footer>
|
164 |
-
|
165 |
</body>
|
166 |
</html>
|
|
|
1 |
<!DOCTYPE html>
|
2 |
+
<html lang="en">
|
3 |
<head>
|
4 |
+
<meta charset="utf-8" />
|
5 |
+
<meta name="viewport" content="width=device-width, initial-scale=1" />
|
6 |
+
<meta name="description" content="Beyond ‘Aha!’ — Systematic Meta‑Ability Alignment in Large Reasoning Models presents a three‑stage recipe that explicitly teaches deduction, induction, and abduction, achieving state‑of‑the‑art reasoning performance." />
|
7 |
+
<meta name="keywords" content="Meta‑Abilities, Deduction, Induction, Abduction, Reinforcement Learning, Large Reasoning Models" />
|
8 |
+
<title>Beyond “Aha!” — Meta‑Ability Alignment for Reasoning Models</title>
|
9 |
|
10 |
+
<link href="https://fonts.googleapis.com/css?family=Google+Sans|Noto+Sans|Castoro" rel="stylesheet" />
|
11 |
+
<link rel="stylesheet" href="./static/css/bulma.min.css" />
|
12 |
+
<link rel="stylesheet" href="./static/css/fontawesome.all.min.css" />
|
13 |
+
<link rel="stylesheet" href="./static/css/index.css" />
|
14 |
+
<link rel="icon" href="./static/images/favicon.svg" />
|
15 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
16 |
<script defer src="./static/js/fontawesome.all.min.js"></script>
|
|
|
|
|
|
|
17 |
</head>
|
18 |
<body>
|
19 |
+
<!-- Header / Title -->
|
20 |
+
<section class="hero">
|
21 |
+
<div class="hero-body">
|
22 |
+
<div class="container is-max-desktop">
|
23 |
+
<div class="columns is-centered">
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
24 |
<div class="column has-text-centered">
|
25 |
+
<h1 class="title is-1 publication-title">Beyond “Aha!”: Systematic Meta‑Ability Alignment in Large Reasoning Models</h1>
|
26 |
+
<div class="is-size-5 publication-authors">
|
27 |
+
<!-- Update author list as needed -->
|
28 |
+
<span class="author-block"><a href="#" target="_blank">Zhiyuan Hu</a><sup>1</sup>,</span>
|
29 |
+
<span class="author-block"><a href="#" target="_blank">Yibo Wang</a><sup>2</sup>,</span>
|
30 |
+
<span class="author-block"><a href="#" target="_blank">Hanze Dong</a><sup>3</sup>,</span>
|
31 |
+
<span class="author-block"><a href="#" target="_blank">Yuhui Xu</a><sup>3</sup>,</span>
|
32 |
+
<span class="author-block"><strong>Amrita Saha</strong><sup>3</sup>,</span>
|
33 |
+
<span class="author-block"><strong>Caiming Xiong</strong><sup>3</sup>,</span>
|
34 |
+
<span class="author-block"><strong>Bryan Hooi</strong><sup>1</sup>,</span>
|
35 |
+
<span class="author-block"><strong>Junnan Li</strong><sup>3</sup></span>
|
36 |
+
</div>
|
37 |
+
<div class="is-size-5 publication-authors">
|
38 |
+
<span class="author-block"><sup>1</sup>National University of Singapore,</span>
|
39 |
+
<span class="author-block"><sup>2</sup>Tsinghua University,</span>
|
40 |
+
<span class="author-block"><sup>3</sup>Salesforce AI Research</span>
|
41 |
+
</div>
|
42 |
+
|
43 |
+
<!-- Links -->
|
44 |
+
<div class="column has-text-centered">
|
45 |
+
<div class="publication-links">
|
46 |
+
<span class="link-block">
|
47 |
+
<a href="https://github.com/zhiyuanhubj/Meta-Ability-Alignment/blob/main/Paper.pdf" target="_blank" class="external-link button is-normal is-rounded is-dark">
|
48 |
+
<span class="icon"><i class="fas fa-file-pdf"></i></span>
|
49 |
+
<span>Paper</span>
|
50 |
+
</a>
|
51 |
+
</span>
|
52 |
+
<span class="link-block">
|
53 |
+
<a href="https://github.com/zhiyuanhubj/Meta-Ability-Alignment/blob/main/Paper.pdf" target="_blank" class="external-link button is-normal is-rounded is-dark">
|
54 |
+
<span class="icon"><i class="ai ai-arxiv"></i></span>
|
55 |
+
<span>arXiv</span>
|
56 |
+
</a>
|
57 |
+
</span>
|
58 |
+
<span class="link-block">
|
59 |
+
<a href="https://github.com/your‑repo" target="_blank" class="external-link button is-normal is-rounded is-dark">
|
60 |
+
<span class="icon"><i class="fab fa-github"></i></span>
|
61 |
+
<span>Code</span>
|
62 |
+
</a>
|
63 |
+
</span>
|
64 |
+
<span class="link-block">
|
65 |
+
<a href="https://github.com/your‑repo/releases" target="_blank" class="external-link button is-normal is-rounded is-dark">
|
66 |
+
<span class="icon"><i class="far fa-images"></i></span>
|
67 |
+
<span>Data</span>
|
68 |
+
</a>
|
69 |
+
</span>
|
70 |
+
</div>
|
71 |
</div>
|
72 |
</div>
|
|
|
73 |
</div>
|
74 |
</div>
|
75 |
</div>
|
76 |
+
</section>
|
|
|
77 |
|
78 |
+
<!-- Abstract -->
|
79 |
+
<section class="section">
|
80 |
+
<div class="container is-max-desktop">
|
81 |
+
<div class="columns is-centered has-text-centered">
|
82 |
+
<div class="column is-four-fifths">
|
83 |
+
<h2 class="title is-3">Abstract</h2>
|
84 |
+
<div class="content has-text-justified">
|
85 |
+
<p>Large reasoning models (LRMs) possess a latent capacity for long chain‑of‑thought reasoning, but the timing and consistency of emergent “aha” behaviors remain unpredictable. We explicitly align LRMs with three meta‑abilities—<strong>deduction, induction, and abduction</strong>—using automatically generated, self‑verifiable tasks. Our three‑stage pipeline (individual alignment, parameter‑space merging, and domain‑specific reinforcement learning) lifts performance ceilings by ≤10 % over instruction‑tuned baselines and delivers state‑of‑the‑art accuracy across math, coding, and science benchmarks.</p>
|
86 |
+
</div>
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
87 |
</div>
|
88 |
</div>
|
89 |
</div>
|
90 |
+
</section>
|
91 |
|
92 |
+
<!-- Results & Framework Figures -->
|
93 |
+
<section class="section is-light">
|
94 |
+
<div class="container is-max-desktop">
|
95 |
+
<h2 class="title is-3 has-text-centered">Key Results</h2>
|
96 |
+
<figure class="image">
|
97 |
+
<img src="./static/images/results_tables.png" alt="Performance tables showing consistent gains from meta‑ability alignment." />
|
98 |
+
<figcaption class="has-text-centered">Table 1 & 2: Meta‑ability alignment boosts reasoning performance at both 7B and 32B scales.</figcaption>
|
99 |
+
</figure>
|
100 |
+
<br />
|
101 |
+
<h2 class="title is-3 has-text-centered">Three‑Stage Training Framework</h2>
|
102 |
+
<figure class="image">
|
103 |
+
<img src="./static/images/framework.png" alt="Three‑stage meta‑ability alignment framework diagram." />
|
104 |
+
<figcaption class="has-text-centered">Stage A: Meta‑ability alignment ⟶ Stage B: Parameter‑space merging ⟶ Stage C: Domain‑specific RL.</figcaption>
|
105 |
+
</figure>
|
106 |
</div>
|
107 |
+
</section>
|
|
|
|
|
|
|
|
|
108 |
|
109 |
+
<!-- Related Links (optional) -->
|
110 |
+
<section class="section" id="BibTeX">
|
111 |
+
<div class="container is-max-desktop content">
|
112 |
+
<h2 class="title">BibTeX</h2>
|
113 |
+
<pre><code>@article{hu2025metaability,
|
114 |
+
author = {Hu, Zhiyuan and Wang, Yibo and Dong, Hanze and Xu, Yuhui and Saha, Amrita and Xiong, Caiming and Hooi, Bryan and Li, Junnan},
|
115 |
+
title = {Beyond “Aha!”: Systematic Meta‑Ability Alignment in Large Reasoning Models},
|
116 |
+
journal = {NeurIPS},
|
117 |
+
year = {2025}
|
118 |
}</code></pre>
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
119 |
</div>
|
120 |
+
</section>
|
121 |
+
|
122 |
+
<footer class="footer">
|
123 |
+
<div class="container">
|
124 |
+
<div class="content has-text-centered">
|
125 |
+
<a class="icon-link" target="_blank" href="https://github.com/zhiyuanhubj/Meta-Ability-Alignment/blob/main/Paper.pdf"><i class="fas fa-file-pdf"></i></a>
|
126 |
+
<a class="icon-link" target="_blank" href="https://github.com/your‑repo"><i class="fab fa-github"></i></a>
|
127 |
+
</div>
|
128 |
+
<div class="columns is-centered">
|
129 |
+
<div class="column is-8">
|
130 |
+
<div class="content">
|
131 |
+
<p>This website is licensed under a <a rel="license" target="_blank" href="http://creativecommons.org/licenses/by-sa/4.0/">Creative Commons Attribution‑ShareAlike 4.0 International License</a>.</p>
|
132 |
+
<p>You are free to reuse the <a target="_blank" href="https://github.com/nerfies/nerfies.github.io">source code</a>; please include a link back in the footer.</p>
|
133 |
+
</div>
|
134 |
</div>
|
135 |
</div>
|
136 |
</div>
|
137 |
+
</footer>
|
|
|
|
|
138 |
</body>
|
139 |
</html>
|