Huu-Long Pham

Model retrieval searches for AI models the way a search engine searches for documents: given a task, it predicts how well each candidate model will perform and returns a ranking.

1. Task
2. Predict each model's performance
3. Ranked models

Case studies

1. Image-generation LoRA

   Input

   An image in the desired style

   Finds

   LoRA adapters that reproduce that style

   Approach

   Predicts without running models

   A Transformer learns an image–LoRA compatibility score. It ranks the correct LoRA higher than an earlier image-classification-based method and searches thousands of LoRAs instantly.

   Related papers

   • Image Style Transfer Model Retrieval using Transformer Encoder and Learning to Rank 2026
   • Image-Generation AI Model Retrieval by Contrastive Learning-Based Style Distance Calculation. 2025
   • Which LoRA Should Be Merged Next? Retrieving an Additional LoRA from a Target Image 2026
   • Retrieval of LoRA Models based on Layer-Wise Weight Embedding without Metadata 2026

2. Image segmentation

   Input

   Images of the target objects

   Finds

   Models that segment those objects accurately

   Approach

   Predicts without running models

   Predicts each model's performance from image and model features, evaluated on object types unseen in training. The method is integrated into KNiT's GeXeL AI image-analysis service.

   Related papers

   • Instance Segmentation Model Retrieval using Performance Prediction and Learning to Rank 2025
   • Retrieving Instance Segmentation Models using Local Image Features via Transformer Encoder 2026

3. Pre-trained BERT

   Input

   A labeled text-classification dataset

   Finds

   The BERT model best suited to the task

   Approach

   Quick trial on task data

   Scores each model by k-nearest-neighbour accuracy on its text embeddings, so no candidate needs fine-tuning. It is most effective when the choice of model strongly affects performance.

   Related papers

   • Pre-trained BERT Model Retrieval: Inference-Based No-Learning Approach using k-Nearest Neighbour Algorithm 2025
   • Inference-Based No-Learning Approach on Pre-Trained BERT Model Retrieval. 2024

4. Large language models Ongoing

   Input

   A question

   Finds

   LLMs likely to answer it well

   Approach

   Predicts without running models

   Predicts answer quality from question and LLM features, trained and evaluated with TREC data.