Meta AVTR Adds Visual Reasoning, Lifts Qwen3.5 by 4.5 Points
TL;DR
- AVTR introduces 'visual imagination' as a second reasoning modality alongside textual chain-of-thought for vision-language models.
- A learned router picks between visual reflection and textual reasoning per input, instead of applying long text chains uniformly.
- The abstract reports a 4.5-point average gain over Qwen3.5 across 16 benchmarks with 33% lower inference latency.
Meta researchers propose AVTR, a vision-language setup that pairs textual chain-of-thought with a second track called visual imagination, and then uses a learned router to pick between them for each input. The paper reports a 4.5-point average gain over a Qwen3.5 baseline across 16 benchmarks, with inference latency cut by 33%.
The team, led by Xiaoqian Shen, frames the design against two limits of the current paradigm: text-only reasoning can miss on "vision-centric questions, where accurate answers depend on interpreting task-relevant visual evidence," and applying long textual chains uniformly "often produc[es] verbose and unnecessary traces when little logical inference is required." Visual reflection, the abstract argues, lets the model "revisit its visual observations and distill the salient evidence through generative visual reflection."
The abstract names neither the 16 benchmarks nor per-task numbers, reporting only the averaged lift. The authors add that AVTR "scales effectively into a smaller size backbone, outperforming the evaluated 2B-scale baselines."
Originally reported by huggingface.co
Read the original article →Original headline: Meta's AVTR Adds 'Visual Imagination' Reasoning Mode, Lifts Qwen3.5 +4.5 Points Across 16 Benchmarks