Avi Cohen

Daniel Shalam, Emanuel Ben Baruch, Avi Ben Cohen et al.

Multimodal large language models can emit localized predictions, bounding boxes for objects and temporal windows for video and audio events, but they hallucinate these regions prolifically. The model's own token log-probabilities are nearly uninformative: they conflate grounding quality with input ambiguity, and coordinate tokens become near-deterministic once the model commits. We propose Multi-Token Localized Attention (MTLA): a training-free, post-hoc score that measures how strongly a prediction's tokens attend to the region they claim. Prior attention-based detectors, which sum attention over the entire input modality and read a single response token, are weaker special cases; we show that summing only within the claimed region and aggregating across all prediction tokens recovers a stronger grounding signal. The same recipe applies almost trivially to other modalities and tasks: object detection in images and temporal localization in video and audio. Across multiple MLLM families and three modalities, MTLA improves hallucination AUROC by +7 to +38 over the best prior training-free baseline. Used as a confidence score for re-ranking, it nearly doubles the zero-shot COCO detection AP of an open-source 8B generalist (from 20.4 to 37.0), narrowing the gap to supervised detectors without any task-specific training.

Ty Nguyen, Ian D. Miller, Avi Cohen et al.

Scalable training data generation is a critical problem in deep learning. We propose PennSyn2Real - a photo-realistic synthetic dataset consisting of more than 100,000 4K images of more than 20 types of micro aerial vehicles (MAVs). The dataset can be used to generate arbitrary numbers of training images for high-level computer vision tasks such as MAV detection and classification. Our data generation framework bootstraps chroma-keying, a mature cinematography technique with a motion tracking system, providing artifact-free and curated annotated images where object orientations and lighting are controlled. This framework is easy to set up and can be applied to a broad range of objects, reducing the gap between synthetic and real-world data. We show that synthetic data generated using this framework can be directly used to train CNN models for common object recognition tasks such as detection and segmentation. We demonstrate competitive performance in comparison with training using only real images. Furthermore, bootstrapping the generated synthetic data in few-shot learning can significantly improve the overall performance, reducing the number of required training data samples to achieve the desired accuracy.