Trivita AI achieves a breakthrough in Efficient AI with two lead-author papers at ICLR 2026 

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 Lê Đức Minh as the lead author.

This achievement marks a significant milestone, highlighting Trivita AI’s capability in optimizing model performance under constrained computational and resource conditions.

These acceptances mark an important milestone in Trivita AI’s research trajectory, reinforcing our commitment to developing scientifically grounded, scalable and deployable AI solutions for real-world applications, particularly in high-stakes domains such as healthcare. 

1. Revisit Visual Prompt Tuning: The Expressiveness of Prompt Experts 

Paper link: https://arxiv.org/pdf/2501.18936 

Rethinking Parameter-Efficient Adaptation for Foundation Vision Models 

Large-scale foundation vision models have demonstrated remarkable generalization capabilities across diverse computer vision tasks. However, fully fine-tuning such models remains computationally prohibitive, especially in production environments or data-constrained domains such as medical imaging. 

Visual Prompt Tuning (VPT) has emerged 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. 

Our Contribution: Visual Adaptive Prompt Tuning (VAPT) 

To overcome this limitation, we propose Visual Adaptive Prompt Tuning (VAPT), a novel framework incorporating input-adaptive prompt experts. 

Key contributions include: 

• Enhanced expressiveness via input-conditioned prompt experts 
• Only 0.6% additional computational overhead compared to VPT 
• Fewer trainable parameters than VPT 
• Theoretical proof of optimal sample efficiency for prompt estimation

Empirical Impact 

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 remaining significantly more parameter-efficient. 

Strategic Impact 

VAPT directly strengthens Trivita AI’s product pipeline in: 

• Data-scarce medical AI systems 
• Edge deployment scenarios requiring computational efficiency 
• Scalable customization of foundation models for enterprise clients

By combining theoretical rigor with deployment efficiency, VAPT advances our mission to deliver high-performance AI under realistic operational constraints. 

2. One-Prompt Strikes Back: Sparse Mixture of Experts for Prompt-based Continual Learning 

Paper link: https://arxiv.org/pdf/2509.24483 

Towards Scalable Continual Learning 

Continual Learning aims to enable neural networks to acquire new knowledge without catastrophic forgetting. This is particularly critical in medical and privacy-sensitive environments, where data arrives sequentially and cannot be centrally stored. 

Prompt-based approaches have recently shown state-of-the-art performance in continual learning while maintaining high memory efficiency. However, existing strategies typically allocate task-specific prompt subsets, resulting in parameter growth and increased computational overhead. 

Our Contribution: Sparse Mixture of Prompt Experts (SMoPE) 

We introduce SMoPE, a Sparse Mixture of Experts framework that maintains a single shared prompt structured as multiple experts. 

Core innovations include: 

• A unified shared prompt architecture 
• Sparse activation of input-relevant experts 
• Substantial parameter reduction 
• Up to 50% computational cost reduction

SMoPE significantly outperforms task-specific prompt methods despite using only a single shared prompt and matches or exceeds state-of-the-art continual learning performance. 

Strategic Impact 

SMoPE enhances Trivita AI’s capability in: 

• Lifelong medical AI systems 
• Federated and privacy-preserving deployments 
• Enterprise AI platforms requiring continuous model updates 

This work establishes a principled pathway for scalable, adaptive and resource-efficient continual learning systems. 

Advancing Research-Driven AI Innovation 

The acceptance of these works at ICLR reflects Trivita AI’s dual commitment: 

• Advancing the theoretical frontiers of prompt-based learning and Mixture of Experts 
• Translating research breakthroughs into deployable AI systems with measurable impact 

Through rigorous research and strategic technology transfer, Trivita AI continues to position itself at the intersection of academic excellence and industrial innovation.