MOCHA: Multi-modal Objects-aware Cross-arcHitecture Alignment

A new distillation framework called MOCHA transfers multimodal knowledge from a frozen vision-language model into a lightweight vision-only detector, enabling few-shot personalized object detection without the computational burden of large models at inference time, according to a preprint posted on arXiv [1]. The framework, formally named Multi-modal Objects-aware Cross-arcHitecture Alignment, was detailed in a paper submitted to the arXiv preprint repository on September 17, 2025, and last revised on June 22, 2026 [1]. The corresponding author is Elena Camuffo [2]. The work addresses a persistent tension in personalized object detection: lightweight models lack the semantic depth to recognize user-specific objects from only a handful of examples, while large vision-language models, or VLMs, are too computationally expensive for real-time or on-device use [2]. MOCHA resolves this by using a frozen VLM teacher — such as LLaVa — to guide the training of a compact student detector like YOLO [4]. A translation module maps the student’s features into the teacher’s multimodal space, where a dual-objective loss enforces both local alignment and global relational consistency across regions [4]. The student never requires textual input or teacher modifications at inference, keeping the deployment footprint small [2]. Across four personalized detection benchmarks — PerSeg, POD, CORe50, and iCubWorld — MOCHA delivered an average improvement of +10.1 over a YOLOv8n baseline and outperformed the previous best competitor, AuXFT, by +4.9 [3]. The method also generalized to other student architectures, including YOLOv11n and RT-DETR-l [3]. The authors note that the resulting compact detectors are suited for edge deployment on mobile and robotic platforms [3]. The paper’s submission history shows five versions, with the file size growing from 9,534 KB in the initial upload to 9,649 KB in the most recent revision [1]. arXiv, which hosts the preprint, is an open-access repository of electronic preprints moderated but not peer-reviewed, and it surpassed two million articles by the end of 2021 [8]. The MOCHA paper appears under the computer-vision category and is accompanied by arXivLabs integrations, including Bibliographic Explorer and Connected Papers, which allow readers to navigate citation trees and explore related research [7].