Skip to content

Commit a1e69a4

Browse files
committed
sync text width narrower
1 parent 8a721ad commit a1e69a4

File tree

1 file changed

+61
-63
lines changed

1 file changed

+61
-63
lines changed

app/(dashboard)/page.tsx

Lines changed: 61 additions & 63 deletions
Original file line numberDiff line numberDiff line change
@@ -612,19 +612,56 @@ export default function HomePage() {
612612
</div>
613613
</section>
614614

615+
616+
{/* Introduction */}
617+
<section className="py-12 bg-white">
618+
<div className="max-w-4xl mx-auto px-8 sm:px-12 lg:px-16">
619+
<div className="text-center">
620+
<div className="text-center mb-8">
621+
<img
622+
src="/images/intro_teaser.jpg"
623+
alt="Verbalized Sampling Introduction"
624+
className="w-full max-w-4xl mx-auto rounded-lg shadow-lg"
625+
/>
626+
<div className="text-sm text-gray-500 mt-2">
627+
<strong>Figure 1:</strong> Overview of Verbalized Sampling (VS) for unlocking LLM diversity.
628+
</div>
629+
</div>
630+
631+
<h2 className="text-xl font-bold text-gray-700 tracking-tight sm:text-3xl mb-8">
632+
Abstract
633+
</h2>
634+
635+
<div className="max-w-4xl mx-auto">
636+
<p className="text-base text-gray-600 leading-relaxed text-left">
637+
Post-training alignment often reduces LLM diversity, leading to a phenomenon known as <em>mode collapse</em>.
638+
Unlike prior work that attributes this effect to algorithmic limitations, we identify a fundamental, pervasive data-level driver: <em>typicality bias</em> in preference data,
639+
whereby annotators systematically favor familiar text as a result of well-established findings in cognitive psychology.
640+
We formalize this bias theoretically, verify it empirically on preference datasets, and show that it plays a central role in mode collapse.
641+
</p>
642+
<p className="text-base text-gray-600 leading-relaxed text-left mt-6">
643+
Motivated by this analysis, we introduce <strong>Verbalized Sampling (VS)</strong>, a simple, training-free prompting method to circumvent mode collapse. VS prompts the model to verbalize a probability distribution over a set of responses (e.g., "Generate 5 jokes about coffee and their corresponding probabilities").
644+
Comprehensive experiments show that VS significantly improves performance across creative writing (poems, stories, jokes), dialogue simulation, open-ended QA, and synthetic data generation, without sacrificing factual accuracy and safety. For instance, in creative writing, VS increases diversity by 1.6-2.1× over direct prompting. We further observe an emergent trend that more capable models benefit more from VS.
645+
In sum, our work provides a new data-centric perspective on mode collapse and a practical inference-time remedy that helps unlock pre-trained generative diversity.
646+
</p>
647+
</div>
648+
</div>
649+
</div>
650+
</section>
651+
615652
{/* Verbalized Sampling: Title & Description left, install/code right */}
616-
<section className="py-8 bg-white">
653+
<section className="py-12 bg-gray-50">
617654
<div className="max-w-6xl mx-auto px-8 sm:px-12 lg:px-16">
618655
<div className="lg:grid lg:grid-cols-2 lg:gap-12 lg:items-center">
619656
<div>
620657
<h2 className="text-xl font-bold text-gray-700 tracking-tight sm:text-3xl mb-6">
621658
Make Your LLMs More Creative!<br />
622-
<span className="text-orange-500"> with Verbalized Sampling</span>
659+
<span className="text-orange-500"> With Verbalized Sampling</span>
623660
</h2>
624661
<div className="space-y-4 text-base text-gray-600">
625662
<p>
626-
Run verbalized sampling and unlock diverse LLM generations in seconds.
627-
Just install and use—the open-source package is free for research and ready for your next project!
663+
Run Verbalized Sampling and unlock diverse LLM generations in seconds.
664+
Just install and use our open-source package!
628665
</p>
629666
<p>
630667
Check our{' '}
@@ -651,47 +688,8 @@ export default function HomePage() {
651688
</div>
652689
</div>
653690
</section>
654-
655-
{/* Introduction */}
656-
<section className="py-12 bg-gray-50">
657-
<div className="max-w-6xl mx-auto px-8 sm:px-12 lg:px-16">
658-
<div className="text-center">
659-
<img
660-
src="/images/intro_teaser.jpg"
661-
alt="Verbalized Sampling Introduction"
662-
className="w-full max-w-4xl mx-auto rounded-lg shadow-lg"
663-
/>
664-
<div className="text-sm text-gray-500 mt-2">
665-
<strong>Figure 1:</strong> Overview of Verbalized Sampling (VS) for unlocking LLM diversity.
666-
</div>
667-
</div>
668-
</div>
669-
</section>
670-
671-
<section className="py-12 bg-white">
672-
<div className="max-w-6xl mx-auto px-8 sm:px-12 lg:px-16">
673-
<div className="text-center">
674-
<h2 className="text-xl font-bold text-gray-700 tracking-tight sm:text-3xl mb-8">
675-
Abstract
676-
</h2>
677-
<div className="max-w-6xl mx-auto">
678-
<p className="text-base text-gray-600 leading-relaxed text-left">
679-
Post-training alignment often reduces LLM diversity, leading to a phenomenon known as <em>mode collapse</em>.
680-
Unlike prior work that attributes this effect to algorithmic limitations, we identify a fundamental, pervasive data-level driver: <em>typicality bias</em> in preference data,
681-
whereby annotators systematically favor familiar text as a result of well-established findings in cognitive psychology.
682-
We formalize this bias theoretically, verify it empirically on preference datasets, and show that it plays a central role in mode collapse.
683-
</p>
684-
<p className="text-base text-gray-600 leading-relaxed text-left mt-6">
685-
Motivated by this analysis, we introduce <strong>Verbalized Sampling (VS)</strong>, a simple, training-free prompting method to circumvent mode collapse. VS prompts the model to verbalize a probability distribution over a set of responses (e.g., "Generate 5 jokes about coffee and their corresponding probabilities").
686-
Comprehensive experiments show that VS significantly improves performance across creative writing (poems, stories, jokes), dialogue simulation, open-ended QA, and synthetic data generation, without sacrificing factual accuracy and safety. For instance, in creative writing, VS increases diversity by 1.6-2.1× over direct prompting. We further observe an emergent trend that more capable models benefit more from VS.
687-
In sum, our work provides a new data-centric perspective on mode collapse and a practical inference-time remedy that helps unlock pre-trained generative diversity.
688-
</p>
689-
</div>
690-
</div>
691-
</div>
692-
</section>
693691

694-
<section className="py-12 bg-gray-50">
692+
<section className="py-12 bg-white">
695693
<div className="max-w-6xl mx-auto px-8 sm:px-12 lg:px-16">
696694
<div className="lg:grid lg:grid-cols-2 lg:gap-12 lg:items-center">
697695
<div>
@@ -726,7 +724,7 @@ export default function HomePage() {
726724
</div>
727725
</section>
728726

729-
<section className="py-12 bg-white">
727+
<section className="py-12 bg-gray-50">
730728
<div className="max-w-6xl mx-auto px-8 sm:px-12 lg:px-16">
731729
<div className="lg:grid lg:grid-cols-2 lg:gap-12 lg:items-center">
732730
<div className="mt-8 lg:mt-0">
@@ -767,7 +765,7 @@ export default function HomePage() {
767765
</div>
768766
</section>
769767

770-
<section className="py-12 bg-gray-50">
768+
<section className="py-12 bg-white">
771769
<div className="max-w-6xl mx-auto px-8 sm:px-12 lg:px-16">
772770
<div className="max-w-6xl mx-auto text-center">
773771
<h2 className="text-xl font-bold text-gray-700 tracking-tight sm:text-3xl mb-6">
@@ -785,24 +783,24 @@ export default function HomePage() {
785783
<strong>Figure 4:</strong> Qualitative and quantitative examples of Verbalized Sampling on creative writing, dialogue simulation, and enumerative open-ended QA.
786784
</p>
787785
</div>
788-
<div className="space-y-4 text-base text-gray-600 text-left">
789-
<p>
790-
Our comprehensive experiments on multiple tasks demonstrate that Verbalized Sampling significantly improves the diversity-quality trade-off across tasks and model families,
791-
without compromising factual accuracy and safety.
792-
</p>
793-
<p>
794-
As shown in Figure 4, for <strong>story writing</strong>, VS improves the output diversity.
795-
For <strong>dialogue simulation</strong>, VS simulates the donation amount distribution much closer to the human distribution, and generates more realistic persuasion behaviors.
796-
On the task of <strong>enumerative open-ended QA</strong>, we ask the model to "generate US states". We first query a pretraining corpus (RedPajama) to establish a "reference" distribution of US
797-
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).
798-
In contrast, direct prompting collapses into a few modes, repeatedly outputting states like California and Texas.
799-
</p>
800-
</div>
786+
<div className="space-y-4 text-base text-gray-600 text-left max-w-5xl mx-auto xl:max-w-3xl">
787+
<p>
788+
Our comprehensive experiments on multiple tasks demonstrate that Verbalized Sampling significantly improves the diversity-quality trade-off across tasks and model families,
789+
without compromising factual accuracy and safety.
790+
</p>
791+
<p>
792+
As shown in Figure 4, for <strong>story writing</strong>, VS improves the output diversity.
793+
For <strong>dialogue simulation</strong>, VS simulates the donation amount distribution much closer to the human distribution, and generates more realistic persuasion behaviors.
794+
On the task of <strong>enumerative open-ended QA</strong>, we ask the model to "generate US states". We first query a pretraining corpus (RedPajama) to establish a "reference" distribution of US
795+
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).
796+
In contrast, direct prompting collapses into a few modes, repeatedly outputting states like California and Texas.
797+
</p>
798+
</div>
801799
</div>
802800
</div>
803801
</section>
804802

805-
<section className="py-12 bg-white">
803+
<section className="py-12 bg-gray-50">
806804
<div className="max-w-6xl mx-auto px-8 sm:px-12 lg:px-16">
807805
<div className="lg:grid lg:grid-cols-2 lg:gap-12 lg:items-center">
808806
<div>
@@ -832,7 +830,7 @@ export default function HomePage() {
832830
</div>
833831
</section>
834832

835-
<section className="py-12 bg-gray-50">
833+
<section className="py-12 bg-white">
836834
<div className="max-w-6xl mx-auto px-8 sm:px-12 lg:px-16">
837835
<div className="lg:grid lg:grid-cols-2 lg:gap-12 lg:items-center">
838836
<div>
@@ -863,7 +861,7 @@ export default function HomePage() {
863861
</div>
864862
</section>
865863

866-
<section className="py-12 bg-white w-full">
864+
<section className="py-12 bg-gray-50 w-full">
867865
<div className="max-w-6xl mx-auto px-8 sm:px-12 lg:px-16">
868866
<div className="lg:grid lg:grid-cols-12 lg:gap-4">
869867
<div className="sm:text-center md:max-w-2xl md:mx-auto lg:col-span-7 lg:text-left">
@@ -885,7 +883,7 @@ export default function HomePage() {
885883
</section>
886884

887885
{/* BibTeX Citation */}
888-
<section className="py-12 bg-gray-50">
886+
<section className="py-12 bg-white">
889887
<div className="max-w-6xl mx-auto px-8 sm:px-12 lg:px-16">
890888
<div className="text-center">
891889
<h2 className="text-xl font-bold text-gray-700 sm:text-3xl mb-8">

0 commit comments

Comments
 (0)