Apple ML Research

Discourse-aware translation can drastically improve sign language gloss accuracy, enhancing spatial consistency and entity tracking beyond traditional methods.

HarmTrace reveals that fine-grained target identification can drastically improve harmful meme detection accuracy, closing a critical gap in multimodal AI understanding.

Personalization in incremental video search can boost ranking performance by over 8% for ambiguous queries, revealing the critical role of user intent in real-time interactions.

MT-EditFlow bridges the gap between local planning and global success in multi-turn image editing, achieving a significant performance boost over leading models.

Forget training: PinPoint, a training-free point selector, closes the performance gap between zero-shot VLMs and fine-tuned specialists in referring image segmentation by intelligently choosing interior points for prompting SAM.

Text-to-video generation gets a 1.58x speed boost with CalibAtt, a training-free method that exploits consistent sparsity patterns in attention layers.

RL fine-tuning can make vision-language models *less* reliable reasoners, as gains in benchmark accuracy come at the cost of faithfulness to the underlying visual grounding and chain-of-thought.