Update index.html
Browse files- index.html +9 -5
index.html
CHANGED
|
@@ -97,7 +97,7 @@
|
|
| 97 |
<div class="authors">
|
| 98 |
<span>Jinkun Hou<sup>1*</sup>,</span><span>Zhuo Liu<sup>2*</sup>,</span><span>Huimin Ren<sup>3†</sup>,</span><span>Hongsheng Xin<sup>3</sup>,</span><span>Pan Zhou<sup>3</sup>,</span><span>Kun Zhan<sup>3</sup></span>
|
| 99 |
</div>
|
| 100 |
-
<div class="affiliations"><sup>1</sup>Peking University · <sup>2</sup>Beijing Institute of Technology · <sup>3</sup>Li Auto Inc.<br>*
|
| 101 |
<nav class="links" aria-label="Project links">
|
| 102 |
<a class="button" href="https://arxiv.org/pdf/2608.09123" target="_blank" rel="noopener">Paper PDF</a>
|
| 103 |
<a class="button secondary" href="https://arxiv.org/abs/2608.09123" target="_blank" rel="noopener">arXiv</a>
|
|
@@ -108,7 +108,8 @@
|
|
| 108 |
<!-- Export Figure 1 from your local PDF and replace this div with:
|
| 109 |
<img src="./static/images/overview.jpg" alt="RISE-RL guidance and exploration overview">
|
| 110 |
-->
|
| 111 |
-
<
|
|
|
|
| 112 |
<figcaption><strong>Guidance expands the reachable policy space.</strong> RISE-RL targets repeatedly missed rubric criteria early, then returns to autonomous on-policy exploration.</figcaption>
|
| 113 |
</figure>
|
| 114 |
</div>
|
|
@@ -137,7 +138,8 @@
|
|
| 137 |
<!-- Export Figure 2 from your local PDF and replace this div with:
|
| 138 |
<img src="./static/images/framework.jpg" alt="Full RISE-RL algorithm framework">
|
| 139 |
-->
|
| 140 |
-
|
|
|
|
| 141 |
<figcaption><strong>Figure 2.</strong> Natural-prompt GRPO is paired with an early selective-guidance branch. The auxiliary guidance loss is removed after the reward advantage saturates.</figcaption>
|
| 142 |
</figure>
|
| 143 |
<div class="explain-grid">
|
|
@@ -156,7 +158,8 @@
|
|
| 156 |
<!-- Export Table 1 from your local PDF and replace this div with:
|
| 157 |
<img src="./static/images/main-results.png" alt="Main results across four open-ended domains">
|
| 158 |
-->
|
| 159 |
-
<
|
|
|
|
| 160 |
<figcaption><strong>Table 1.</strong> Main results across four domains. Green values denote gains over the corresponding initial model; bold rows show RISE-RL.</figcaption>
|
| 161 |
</figure>
|
| 162 |
<div class="result-note">
|
|
@@ -174,7 +177,8 @@
|
|
| 174 |
<!-- Export Figure 3 from your local PDF and replace this div with:
|
| 175 |
<img src="./static/images/guidance-comparison.jpg" alt="Rubric-RL, persistent guidance, and RISE-RL comparison">
|
| 176 |
-->
|
| 177 |
-
|
|
|
|
| 178 |
<figcaption><strong>Figure 3.</strong> Rubric-RL, persistent guidance, and RISE-RL on Writing and Health. Dashed vertical lines indicate guidance removal.</figcaption>
|
| 179 |
</figure>
|
| 180 |
<div class="dynamics-copy">
|
|
|
|
| 97 |
<div class="authors">
|
| 98 |
<span>Jinkun Hou<sup>1*</sup>,</span><span>Zhuo Liu<sup>2*</sup>,</span><span>Huimin Ren<sup>3†</sup>,</span><span>Hongsheng Xin<sup>3</sup>,</span><span>Pan Zhou<sup>3</sup>,</span><span>Kun Zhan<sup>3</sup></span>
|
| 99 |
</div>
|
| 100 |
+
<div class="affiliations"><sup>1</sup>Peking University · <sup>2</sup>Beijing Institute of Technology · <sup>3</sup>Li Auto Inc.<br>* This work was done during an internship at Li Auto Inc. · † Corresponding author</div>
|
| 101 |
<nav class="links" aria-label="Project links">
|
| 102 |
<a class="button" href="https://arxiv.org/pdf/2608.09123" target="_blank" rel="noopener">Paper PDF</a>
|
| 103 |
<a class="button secondary" href="https://arxiv.org/abs/2608.09123" target="_blank" rel="noopener">arXiv</a>
|
|
|
|
| 108 |
<!-- Export Figure 1 from your local PDF and replace this div with:
|
| 109 |
<img src="./static/images/overview.jpg" alt="RISE-RL guidance and exploration overview">
|
| 110 |
-->
|
| 111 |
+
<img src="./static/images/overview.jpg"
|
| 112 |
+
alt="RISE-RL guidance and exploration overview">
|
| 113 |
<figcaption><strong>Guidance expands the reachable policy space.</strong> RISE-RL targets repeatedly missed rubric criteria early, then returns to autonomous on-policy exploration.</figcaption>
|
| 114 |
</figure>
|
| 115 |
</div>
|
|
|
|
| 138 |
<!-- Export Figure 2 from your local PDF and replace this div with:
|
| 139 |
<img src="./static/images/framework.jpg" alt="Full RISE-RL algorithm framework">
|
| 140 |
-->
|
| 141 |
+
<img src="./static/images/framework.jpg"
|
| 142 |
+
alt="Full RISE-RL algorithm framework">
|
| 143 |
<figcaption><strong>Figure 2.</strong> Natural-prompt GRPO is paired with an early selective-guidance branch. The auxiliary guidance loss is removed after the reward advantage saturates.</figcaption>
|
| 144 |
</figure>
|
| 145 |
<div class="explain-grid">
|
|
|
|
| 158 |
<!-- Export Table 1 from your local PDF and replace this div with:
|
| 159 |
<img src="./static/images/main-results.png" alt="Main results across four open-ended domains">
|
| 160 |
-->
|
| 161 |
+
<img src="./static/images/main-results.png"
|
| 162 |
+
alt="Main results across four open-ended domains">
|
| 163 |
<figcaption><strong>Table 1.</strong> Main results across four domains. Green values denote gains over the corresponding initial model; bold rows show RISE-RL.</figcaption>
|
| 164 |
</figure>
|
| 165 |
<div class="result-note">
|
|
|
|
| 177 |
<!-- Export Figure 3 from your local PDF and replace this div with:
|
| 178 |
<img src="./static/images/guidance-comparison.jpg" alt="Rubric-RL, persistent guidance, and RISE-RL comparison">
|
| 179 |
-->
|
| 180 |
+
<img src="./static/images/guidance-comparison.jpg"
|
| 181 |
+
alt="Rubric-RL, persistent guidance, and RISE-RL comparison">
|
| 182 |
<figcaption><strong>Figure 3.</strong> Rubric-RL, persistent guidance, and RISE-RL on Writing and Health. Dashed vertical lines indicate guidance removal.</figcaption>
|
| 183 |
</figure>
|
| 184 |
<div class="dynamics-copy">
|