Skip to content

Commit f37a167

Browse files
jiayizxjiayizx
committed
update website
1 parent 882c469 commit f37a167

File tree

4 files changed

+75
-44
lines changed

4 files changed

+75
-44
lines changed

‎app/(dashboard)/layout.tsx‎

Lines changed: 31 additions & 12 deletions
Original file line numberDiff line numberDiff line change
@@ -41,6 +41,12 @@ function UserMenu() {
4141
Paper
4242
</Link>
4343
</Button>
44+
<Button asChild className="rounded-full">
45+
<Link href="https://www.verbalized-sampling.blog/" className="flex items-center gap-2" target="_blank" rel="noopener noreferrer">
46+
<BookOpen size={16} />
47+
Blog
48+
</Link>
49+
</Button>
4450
<Button asChild className="rounded-full">
4551
<Link href="https://github.com/CHATS-lab/verbalized-sampling" className="flex items-center gap-2">
4652
<Github size={16} />
@@ -55,18 +61,31 @@ function UserMenu() {
5561
X Thread
5662
</Link>
5763
</Button>
58-
<Button asChild className="rounded-full">
59-
<Link href="https://www.verbalized-sampling.com/#blog" className="flex items-center gap-2" target="_blank" rel="noopener noreferrer">
60-
<BookOpen size={16} />
61-
Blog
62-
</Link>
63-
</Button>
64-
<Button asChild className="rounded-full">
65-
<Link href="https://github.com/CHATS-lab/verbalized-sampling?tab=readme-ov-file#-interactive-notebooks" className="flex items-center gap-2">
66-
<NotebookPen size={16} />
67-
Notebook
68-
</Link>
69-
</Button>
64+
<DropdownMenu>
65+
<DropdownMenuTrigger asChild>
66+
<Button className="rounded-full flex items-center gap-2">
67+
<NotebookPen size={16} />
68+
Notebooks
69+
</Button>
70+
</DropdownMenuTrigger>
71+
<DropdownMenuContent align="end">
72+
<DropdownMenuItem asChild>
73+
<Link href="https://colab.research.google.com/drive/1UDk4W5w6gF0dQ9Tpu0sPQethEht51GXL#offline=true&sandboxMode=true" target="_blank" rel="noopener noreferrer" className="flex items-center gap-2">
74+
Direct vs. Verbalized Sampling
75+
</Link>
76+
</DropdownMenuItem>
77+
<DropdownMenuItem asChild>
78+
<Link href="https://colab.research.google.com/drive/1J18VJRnrCjIb6sTivY-znb8C3JsLQCIz#offline=true&sandboxMode=true" target="_blank" rel="noopener noreferrer" className="flex items-center gap-2">
79+
Image Generation with VS
80+
</Link>
81+
</DropdownMenuItem>
82+
<DropdownMenuItem asChild>
83+
<Link href="https://colab.research.google.com/drive/1eC0nIUVC1kyANxxzhNib44qmPphdWy9o#offline=true&sandboxMode=true" target="_blank" rel="noopener noreferrer" className="flex items-center gap-2">
84+
Complete Framework Tutorial
85+
</Link>
86+
</DropdownMenuItem>
87+
</DropdownMenuContent>
88+
</DropdownMenu>
7089
</>
7190
);
7291
}

‎app/(dashboard)/page.tsx‎

Lines changed: 44 additions & 32 deletions
Original file line numberDiff line numberDiff line change
@@ -547,10 +547,10 @@ const RollingText = () => {
547547
useEffect(() => {
548548
const interval = setInterval(() => {
549549
setCurrentIndex((prev) => (prev + 1) % texts.length);
550-
}, 3000); // Change every 3 seconds
550+
}, 200); // Change every 0.2 second
551551

552552
return () => clearInterval(interval);
553-
}, []);
553+
}, [texts.length]);
554554

555555
return (
556556
<span className="block overflow-hidden relative" style={{ height: '48px' }}>
@@ -656,8 +656,8 @@ export default function HomePage() {
656656
<div className="lg:grid lg:grid-cols-2 lg:gap-12 lg:items-center">
657657
<div>
658658
<h2 className="text-3xl font-bold text-gray-700 tracking-tight sm:text-4xl mb-6">
659-
Data-Level Cause of<br />
660-
Mode Collapse: <span className="text-orange-500">Typicality Bias</span>
659+
Why Mode Collapse Happens?<br />
660+
<span className="text-orange-500">Typicality Bias</span>
661661
</h2>
662662
<div className="space-y-4 text-lg text-gray-600">
663663
<p>
@@ -701,8 +701,8 @@ export default function HomePage() {
701701
</div>
702702
<div>
703703
<h2 className="text-3xl font-bold text-gray-700 tracking-tight sm:text-4xl mb-6">
704-
Three Types of
705-
<span className="block text-orange-500">Prompting Methods</span>
704+
How to Mitigate Mode Collapse?<br />
705+
<span className="block text-orange-500">Verbalized Sampling</span>
706706
</h2>
707707
<div className="space-y-4 text-lg text-gray-600">
708708
<p>
@@ -731,8 +731,9 @@ export default function HomePage() {
731731
<div className="max-w-7xl mx-auto px-4 sm:px-6 lg:px-8">
732732
<div className="max-w-5xl mx-auto text-center">
733733
<h2 className="text-3xl font-bold text-gray-700 tracking-tight sm:text-4xl mb-6">
734-
Verbalized Sampling Works On
735-
<RollingText />
734+
Where VS Works:<br />
735+
<span className="block text-orange-500">Creative Writing, Social Simulation, ..., and Your Task!</span>
736+
{/* <RollingText /> */}
736737
</h2>
737738
<div className="text-center mb-8">
738739
<img
@@ -756,25 +757,6 @@ export default function HomePage() {
756757
state names in the pretraining data. The verbalized probability distribution generated by VS, when averaged over 10 trials, closely aligns with this reference pretraining distribution (KL=0.12).
757758
In contrast, direct prompting collapses into a few modes, repeatedly outputting states like California and Texas.
758759
</p>
759-
<div className="lg:grid lg:grid-cols-2 lg:gap-12 lg:items-center">
760-
<div className="space-y-4 text-lg text-gray-600 text-left">
761-
<p>
762-
We observe an <strong>emergent trend</strong> where larger models benefit more from VS. Figure 5 shows the diversity gain over the direct prompting which suffers from mode collapse.
763-
Across all VS variants, larger models (GPT-4.1, Gemini-2.5-Pro) achieve diversity gains 1.5 to 2 times greater than smaller models (GPT-4.1-Mini, Gemini-2.5-Flash).
764-
</p>
765-
</div>
766-
<div className="mt-8 lg:mt-0">
767-
<img
768-
src="/images/emergent_trend.png"
769-
alt="Emergent Trend: Larger Models Benefit More from VS"
770-
className="w-full rounded-lg shadow-lg"
771-
/>
772-
<p className="text-base text-gray-500 mt-3 text-center italic">
773-
<strong>Figure 5:</strong> Emergent trend where larger models benefit more from VS. We show differences in
774-
diversity (e) and quality (f) over Direct across small and large models.
775-
</p>
776-
</div>
777-
</div>
778760
</div>
779761
</div>
780762
</div>
@@ -785,8 +767,38 @@ export default function HomePage() {
785767
<div className="lg:grid lg:grid-cols-2 lg:gap-12 lg:items-center">
786768
<div>
787769
<h2 className="text-3xl font-bold text-gray-700 tracking-tight sm:text-4xl mb-6">
788-
Pro-tip for
789-
<span className="block text-orange-500">Maximizing Diversity</span>
770+
What Else We Discovered:
771+
<span className="block text-orange-500">Emergent Trends</span>
772+
</h2>
773+
<div className="space-y-4 text-lg text-gray-600">
774+
<p>
775+
We observe an <strong>emergent trend</strong> where larger models benefit more from VS. Figure 5 shows the diversity gain over the direct prompting which suffers from mode collapse.
776+
Across all VS variants, larger models (GPT-4.1, Gemini-2.5-Pro) achieve diversity gains 1.5 to 2 times greater than smaller models (GPT-4.1-Mini, Gemini-2.5-Flash).
777+
</p>
778+
</div>
779+
</div>
780+
<div className="mt-8 lg:mt-0">
781+
<img
782+
src="/images/emergent_trend.png"
783+
alt="Emergent Trend: Larger Models Benefit More from VS"
784+
className="w-full rounded-lg shadow-lg"
785+
/>
786+
<p className="text-base text-gray-500 mt-3 text-center italic">
787+
<strong>Figure 5:</strong> Emergent trend where larger models benefit more from VS. We show differences in
788+
diversity (e) and quality (f) over Direct across small and large models.
789+
</p>
790+
</div>
791+
</div>
792+
</div>
793+
</section>
794+
795+
<section className="py-12 bg-gray-50">
796+
<div className="max-w-7xl mx-auto px-4 sm:px-6 lg:px-8">
797+
<div className="lg:grid lg:grid-cols-2 lg:gap-12 lg:items-center">
798+
<div>
799+
<h2 className="text-3xl font-bold text-gray-700 tracking-tight sm:text-4xl mb-6">
800+
How to Maximize Diversity:
801+
<span className="block text-orange-500">Probability Tuning</span>
790802
</h2>
791803
<div className="space-y-4 text-lg text-gray-600">
792804
<p>
@@ -812,12 +824,12 @@ export default function HomePage() {
812824
</section>
813825

814826
<section className="py-12 bg-gray-50">
815-
<div className="max-w-7xl mx-auto px-4 sm:px-6 lg:px-8">
827+
<div className="max-w-7xl mx-auto px-4 sm:px-6 lg:px-8">
816828
<div className="lg:grid lg:grid-cols-12 lg:gap-8">
817829
<div className="sm:text-center md:max-w-2xl md:mx-auto lg:col-span-7 lg:text-left">
818830
<h2 className="text-3xl font-bold text-gray-700 tracking-tight sm:text-4xl mb-4">
819-
Try the Magic Prompt:
820-
<span className="block text-orange-500">Unlock LLM Diversity</span>
831+
Try It Yourself:
832+
<span className="block text-orange-500">The Magic Prompt</span>
821833
</h2>
822834
<p className="mt-3 text-base text-gray-500 sm:mt-5 sm:text-xl lg:text-lg xl:text-xl">
823835
Verbalized Sampling provides a training-free, model-agnostic approach to mitigating mode collapse by prompting the model to generate response distributions with verbalized probability estimates.

‎public/images/qualitative_examples.png‎

35.7 KB
Loading

‎public/images/qualitative_examples_0.png‎

495 KB
Loading

0 commit comments

Comments
 (0)