Falcon Perception Update: RL Fixes High-Density Perception

HildeKuehne · x · 2026-08-22

Falcon Perception updated with GRPO post-training weights, significantly improving referring expression abilities. The update addresses pitfalls of maximum-likelihood training (e.g., repetition, low recall) and unlocks state-of-the-art performance in high-density scenes (up to 500 objects). It eliminates the need for post-processing heuristics like NMS.

Original post →

More from Multimodal

Multimodal channel →