Meta's AVTR adds visual imagination routing: +4.5 pts, 33% lower latency
meta · hf · 2026-11-19
Meta AI, KAUST, and Princeton propose AVTR (Adaptive Visual-Textual Reasoning), targeting two flaws of text-heavy VLM reasoning: weak perception on vision-centric questions and verbose traces when little inference is needed.
- Introduces generative "visual imagination" as a complementary modality, revisiting visual evidence via reflective visual reasoning
- Formulates modality selection as a learned routing problem, adaptively balancing visual reflection and textual reasoning per input
- Improves average score by 4.5 points over the Qwen3.5 baseline across 16 benchmarks while cutting inference latency by 33%
- Transfers effectively to a smaller backbone, beating evaluated 2B-scale baselines
More from Multimodal
- WaveView Injects Warped Views Into Noise Phase, Beating NVS Baselines at 2.1x Speed — KIT-Karlsruhe · 2026-11-19
- GPT-6.2 Astra demo: a Corvette escapes a MacBook, rebuilt desk in Blender for the chase — paw_lean · 2026-11-19
- HeyGen creates instant voice + avatar from one filtered photo in 5 minutes — toolstelegraph · 2026-11-19
- Multiple Google employees tease "Gemini", hinting at an imminent new model release — haider1 · 2026-11-19
- MOSS-Image SFT releases 4-step distilled LoRA using eDMD + LADD for high-detail T2I and I2I edits — TimeTruth2490 · 2026-11-19
- MOSS 4-Step Distill LoRA Cuts GPU Time 14x With Better Detail — TimeTruth2490 · 2026-11-19