{"id":973,"date":"2026-02-27T10:05:47","date_gmt":"2026-02-27T03:05:47","guid":{"rendered":"https:\/\/trivita.ai\/?p=973"},"modified":"2026-04-06T09:45:44","modified_gmt":"2026-04-06T02:45:44","slug":"trivita-ai-iclr-2026","status":"publish","type":"post","link":"https:\/\/wp-dev.trivita.ai\/en\/trivita-ai-iclr-2026\/","title":{"rendered":"Trivita AI achieves a breakthrough in Efficient AI with two lead-author papers at ICLR 2026\u00a0"},"content":{"rendered":"<ul class=\"wp-block-list\">\n\n\n\n\n\n<\/ul>\n\n\n\n\n\n\n\n\n\n<ul class=\"wp-block-list\">\n\n\n\n<\/ul>\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n<ul class=\"wp-block-list\">\n\n\n\n\n\n<\/ul>\n\n\n\n\n\n\n\n\n\n<ul class=\"wp-block-list\">\n\n\n\n<\/ul>\n\n\n\n\n\n<p class=\"wp-block-paragraph\">Trivita AI is proud to announce that two breakthrough research papers in Efficient AI have been accepted at the International Conference on Learning Representations (ICLR) 2026, with L\u00ea \u0110\u1ee9c Minh as the lead author.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">This achievement marks a significant milestone, highlighting Trivita AI\u2019s capability in optimizing model performance under constrained computational and resource conditions.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">These acceptances mark an important milestone in Trivita AI\u2019s research trajectory, reinforcing our commitment to developing scientifically grounded,&nbsp;scalable&nbsp;and deployable AI solutions for real-world applications, particularly in high-stakes domains such as healthcare.&nbsp;<\/p>\n\n\n\n<h2 class=\"wp-block-heading\"><strong>1. Revisit Visual Prompt Tuning: The Expressiveness of Prompt Experts<\/strong>&nbsp;<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>Paper link:<\/strong><a href=\"https:\/\/arxiv.org\/pdf\/2501.18936\" target=\"_blank\" rel=\"noreferrer noopener\">&nbsp;https:\/\/arxiv.org\/pdf\/2501.18936<\/a>&nbsp;<\/p>\n\n\n\n<figure class=\"wp-block-image size-full\"><img loading=\"lazy\" decoding=\"async\" width=\"800\" height=\"600\" src=\"http:\/\/124.197.20.221:8080\/wp-content\/uploads\/2026\/02\/Trivita-AI-ICLR-2026.webp\" alt=\"Trivita AI - ICLR 2026\" class=\"wp-image-977\" srcset=\"https:\/\/wp-dev.trivita.ai\/wp-content\/uploads\/2026\/02\/Trivita-AI-ICLR-2026.webp 800w, https:\/\/wp-dev.trivita.ai\/wp-content\/uploads\/2026\/02\/Trivita-AI-ICLR-2026-300x225.webp 300w, https:\/\/wp-dev.trivita.ai\/wp-content\/uploads\/2026\/02\/Trivita-AI-ICLR-2026-768x576.webp 768w, https:\/\/wp-dev.trivita.ai\/wp-content\/uploads\/2026\/02\/Trivita-AI-ICLR-2026-16x12.webp 16w\" sizes=\"auto, (max-width: 800px) 100vw, 800px\" \/><\/figure>\n\n\n\n<h3 class=\"wp-block-heading\"><strong>Rethinking Parameter-Efficient Adaptation for Foundation Vision Models<\/strong>&nbsp;<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">Large-scale foundation vision models have&nbsp;demonstrated&nbsp;remarkable generalization capabilities across diverse computer vision tasks. However, fully fine-tuning such models&nbsp;remains&nbsp;computationally prohibitive, especially in production environments or data-constrained domains such as medical imaging.&nbsp;<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">Visual Prompt Tuning (VPT) has&nbsp;emerged&nbsp;as a parameter-efficient alternative by introducing learnable prompt tokens that guide downstream adaptation. Despite strong empirical performance, recent theoretical connections between prompt-based methods and Mixture of Experts reveal a critical limitation: conventional prompt experts behave as constant functions of the input, restricting model expressiveness.&nbsp;<\/p>\n\n\n\n<h3 class=\"wp-block-heading\"><strong>Our Contribution: Visual Adaptive Prompt Tuning (VAPT)<\/strong>&nbsp;<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">To overcome this limitation, we propose&nbsp;<strong>Visual Adaptive Prompt Tuning (VAPT)<\/strong>, a novel framework incorporating input-adaptive prompt experts.&nbsp;<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">Key contributions include:&nbsp;<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>Enhanced expressiveness&nbsp;via input-conditioned prompt experts&nbsp;<\/li>\n\n\n\n<li>Only&nbsp;0.6%&nbsp;additional&nbsp;computational overhead&nbsp;compared to VPT&nbsp;<\/li>\n\n\n\n<li>Fewer trainable parameters&nbsp;than VPT&nbsp;<\/li>\n\n\n\n<li>Theoretical proof of&nbsp;optimal&nbsp;sample efficiency&nbsp;for prompt estimation<\/li>\n<\/ul>\n\n\n\n<h3 class=\"wp-block-heading\"><strong>Empirical Impact<\/strong>&nbsp;<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">In low-data regimes using only 1% of training data, VAPT improves performance by up to 50% over VPT. Importantly, VAPT also surpasses fully fine-tuned baselines on standard benchmarks while&nbsp;remaining&nbsp;significantly more&nbsp;parameter-efficient.&nbsp;<\/p>\n\n\n\n<h3 class=\"wp-block-heading\"><strong>Strategic Impact<\/strong>&nbsp;<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">VAPT directly strengthens Trivita AI\u2019s product pipeline in:&nbsp;<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>Data-scarce medical AI systems&nbsp;<\/li>\n\n\n\n<li>Edge deployment scenarios requiring computational efficiency&nbsp;<\/li>\n\n\n\n<li>Scalable customization of foundation models for enterprise clients<\/li>\n<\/ul>\n\n\n\n<p class=\"wp-block-paragraph\">By combining theoretical rigor with deployment efficiency, VAPT advances our mission to deliver high-performance AI under realistic operational constraints.&nbsp;<\/p>\n\n\n\n<h2 class=\"wp-block-heading\"><strong>2. One-Prompt Strikes Back: Sparse Mixture of Experts for Prompt-based Continual Learning<\/strong>&nbsp;<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>Paper link:<\/strong><a href=\"https:\/\/arxiv.org\/pdf\/2509.24483\" target=\"_blank\" rel=\"noreferrer noopener\">&nbsp;https:\/\/arxiv.org\/pdf\/2509.24483<\/a>&nbsp;<\/p>\n\n\n\n<figure class=\"wp-block-image size-full\"><img loading=\"lazy\" decoding=\"async\" width=\"936\" height=\"702\" src=\"http:\/\/124.197.20.221:8080\/wp-content\/uploads\/2026\/02\/image.png\" alt=\"\" class=\"wp-image-981\" srcset=\"https:\/\/wp-dev.trivita.ai\/wp-content\/uploads\/2026\/02\/image.png 936w, https:\/\/wp-dev.trivita.ai\/wp-content\/uploads\/2026\/02\/image-300x225.png 300w, https:\/\/wp-dev.trivita.ai\/wp-content\/uploads\/2026\/02\/image-768x576.png 768w, https:\/\/wp-dev.trivita.ai\/wp-content\/uploads\/2026\/02\/image-16x12.png 16w\" sizes=\"auto, (max-width: 936px) 100vw, 936px\" \/><\/figure>\n\n\n\n<h3 class=\"wp-block-heading\"><strong>Towards Scalable Continual Learning<\/strong>&nbsp;<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">Continual Learning aims to enable neural networks to acquire new knowledge without catastrophic forgetting.&nbsp;This is particularly critical in medical and privacy-sensitive environments, where data arrives sequentially and cannot be centrally stored.&nbsp;<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">Prompt-based approaches have recently shown&nbsp;state-of-the-art&nbsp;performance in continual learning while&nbsp;maintaining&nbsp;high memory efficiency. However, existing strategies typically&nbsp;allocate&nbsp;task-specific prompt subsets, resulting in parameter growth and increased computational overhead.&nbsp;<\/p>\n\n\n\n<h3 class=\"wp-block-heading\"><strong>Our Contribution: Sparse Mixture of Prompt Experts (SMoPE)<\/strong>&nbsp;<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">We introduce&nbsp;<strong>SMoPE<\/strong>, a Sparse Mixture of Experts framework that&nbsp;maintains&nbsp;a single shared prompt structured as multiple experts.&nbsp;<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">Core innovations include:&nbsp;<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>A unified shared prompt architecture&nbsp;<\/li>\n\n\n\n<li>Sparse activation of input-relevant experts&nbsp;<\/li>\n\n\n\n<li>Substantial parameter reduction&nbsp;<\/li>\n\n\n\n<li>Up to 50% computational cost reduction<\/li>\n<\/ul>\n\n\n\n<p class=\"wp-block-paragraph\">SMoPE&nbsp;significantly outperforms task-specific prompt methods despite using only a single shared prompt and matches or exceeds&nbsp;state-of-the-art&nbsp;continual learning performance.&nbsp;<\/p>\n\n\n\n<h3 class=\"wp-block-heading\"><strong>Strategic Impact<\/strong>&nbsp;<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">SMoPE&nbsp;enhances Trivita AI\u2019s capability in:&nbsp;<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>Lifelong medical AI systems&nbsp;<\/li>\n\n\n\n<li>Federated and privacy-preserving deployments&nbsp;<\/li>\n\n\n\n<li>Enterprise AI platforms requiring continuous model updates&nbsp;<\/li>\n<\/ul>\n\n\n\n<p class=\"wp-block-paragraph\">This work&nbsp;establishes&nbsp;a principled pathway for scalable,&nbsp;adaptive&nbsp;and resource-efficient continual learning systems.&nbsp;<\/p>\n\n\n\n<h3 class=\"wp-block-heading\"><strong>Advancing Research-Driven AI Innovation<\/strong>&nbsp;<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">The acceptance of these works at ICLR reflects Trivita AI\u2019s dual commitment:&nbsp;<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>Advancing the theoretical frontiers of prompt-based learning and Mixture of Experts&nbsp;<\/li>\n\n\n\n<li>Translating research breakthroughs into deployable AI systems with measurable impact&nbsp;<\/li>\n<\/ul>\n\n\n\n<p class=\"wp-block-paragraph\">Through rigorous research and strategic technology transfer, Trivita AI continues to position itself at the intersection of academic excellence and industrial innovation.&nbsp;<\/p>","protected":false},"excerpt":{"rendered":"<p>Trivita AI is proud to announce that two breakthrough research papers in Efficient AI have been accepted at the International Conference on Learning Representations (ICLR) 2026, with L\u00ea \u0110\u1ee9c Minh&#8230;<\/p>","protected":false},"author":1,"featured_media":979,"comment_status":"open","ping_status":"open","sticky":true,"template":"","format":"standard","meta":{"_acf_changed":false,"inline_featured_image":false,"_kad_post_transparent":"","_kad_post_title":"","_kad_post_layout":"","_kad_post_sidebar_id":"","_kad_post_content_style":"","_kad_post_vertical_padding":"","_kad_post_feature":"","_kad_post_feature_position":"","_kad_post_header":false,"_kad_post_footer":false,"_kad_post_classname":"","footnotes":""},"categories":[5],"tags":[],"class_list":["post-973","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-su-kien"],"acf":[],"_links":{"self":[{"href":"https:\/\/wp-dev.trivita.ai\/en\/wp-json\/wp\/v2\/posts\/973","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/wp-dev.trivita.ai\/en\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/wp-dev.trivita.ai\/en\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/wp-dev.trivita.ai\/en\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/wp-dev.trivita.ai\/en\/wp-json\/wp\/v2\/comments?post=973"}],"version-history":[{"count":5,"href":"https:\/\/wp-dev.trivita.ai\/en\/wp-json\/wp\/v2\/posts\/973\/revisions"}],"predecessor-version":[{"id":1099,"href":"https:\/\/wp-dev.trivita.ai\/en\/wp-json\/wp\/v2\/posts\/973\/revisions\/1099"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/wp-dev.trivita.ai\/en\/wp-json\/wp\/v2\/media\/979"}],"wp:attachment":[{"href":"https:\/\/wp-dev.trivita.ai\/en\/wp-json\/wp\/v2\/media?parent=973"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/wp-dev.trivita.ai\/en\/wp-json\/wp\/v2\/categories?post=973"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/wp-dev.trivita.ai\/en\/wp-json\/wp\/v2\/tags?post=973"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}