Meta's AVTR adds visual imagination to multimodal reasoning, +4.5 pts across 16 benchmarks

liuzhuang1234 · x · 2026-11-21

Meta (with KAUST and Princeton) proposes AVTR, Adaptive Multimodal Reasoning with Visual Imagination. Textual reasoning hits two limits: perceptual bottlenecks (text can't capture visual evidence) and uniform reasoning depth on easy questions. AVTR adaptively picks between textual reasoning and visual imagination — generating intermediate visual reflections in latent space — with RL-distilled mode selection. Across 16 benchmarks it improves average score by 4.5 points over Qwen3.5 baselines while cutting inference latency by 32%, scaling effectively to smaller models and beating GPT-5 / Gemini 2.5 Pro on high-resolution image understanding and spatial reasoning.

Related event: Meta's AVTR Teaches Multimodal Models Visual Imagination(2 posts)→

Original post →

More from Multimodal

Multimodal channel →