Amazon Science

A single verification prompt is insufficient; learning a reliability score from multiple, complementary perspectives unlocks robust self-verification for VLMs.

Splitting same-outcome groups by execution quality can cut rollout requirements by over half while enhancing policy performance in complex VLA tasks.

MemTree3D slashes the computational cost of 3D question answering, boosting performance while avoiding the inefficiencies of traditional visual search methods.

No single perspective is sufficient for art interpretation; MMArt reveals that each perspective uniquely enhances understanding and retrieval of visual art.

A-PACK reveals that deferring audio pruning can lead to a 78% reduction in prefill costs while boosting performance in omni-modal LLMs.

Surface keypoint trajectories enable seamless coordination between human motion and multiple articulated objects, outperforming traditional methods in complex interaction scenarios.

A novel auxiliary loss, FlowMirror, significantly boosts image generation quality by harnessing the predictive power of unsupervised conditioning embeddings.

Multimodal unlearning could revolutionize how we handle sensitive data in AI, enabling targeted removal without sacrificing model performance.

Achieving over 42% recall in semantic video communication could redefine how we transmit meaning in bandwidth-limited networks.

Current MLLMs struggle with Bangla form comprehension, missing key granular details that could hinder their real-world application in low-resource languages.

TrajLoc achieves unprecedented trajectory adherence and visual fidelity in multi-object motion control, outperforming existing methods by isolating object trajectories with Gaussian heatmaps.

Zero-shot generation of 360 panoramas is now possible without the costly fine-tuning or optimization typically required, unlocking new creative potentials in image synthesis.

Synthesizing 48,000 interaction trajectories without human input enables a humanoid robot to learn complex loco-manipulation tasks effectively.

ForceBand transforms human muscle signals into precise force data, enabling robots to learn manipulation tasks with unprecedented accuracy.

GLACIER achieves high predictive accuracy while drastically cutting down the computational costs typically associated with multimodal molecular property prediction.

Massive multimodal models like Qwen and CLIP excel at information retrieval, but their sheer size makes them impractical – this workshop tackles the efficiency gap.

Despite matching or exceeding human expert performance on generating potential diagnoses, current MLLMs struggle to synthesize multimodal clinical evidence for final diagnosis, revealing a critical gap in their clinical reasoning abilities.

Object hallucination in MLLMs can be significantly reduced by simply masking salient visual features during contrastive decoding.