LoopVL: 1B recurrent VLM trained on 0.14T tokens matches 36T-token models, shows 'Visual Aha Moment'

机器之心 · wechat · 2026-10-19

Researchers from Renmin University, Baidu, SJTU and Tsinghua propose LoopVL, extending recurrent computation to vision-language models, with paper, code and weights open-sourced.

Design: A 1B language backbone split into 16-layer L/H modules, running [L×3→H]×2 — 32 unique Transformer layers expand into 128 layer calls; images are encoded once while visual states keep updating across loops.

Efficiency: Trained on just 0.14T tokens (1/260 of Qwen3.5-2B/InternVL3.5-2B's 36T), it matches or beats them on MMMU-Pro, MathVision and more; under controlled same-budget comparisons it beats both deep and wide 4B baselines on 5 of 8 benchmarks.

Key finding — Visual Aha Moment: At the boundary between two Model-Loops (layer call 63→64), attention Gini jumps from 0.405 to 0.858 — attention shifts from diffuse to focused, with previously under-attended tokens rising into the top 25%. Even with identical shared parameters, the model redistributes visual attention across loops. Intervention experiments (blocking, freezing, retraining a read-only variant) confirm visual state updates are a functional part of the loop.

Limits: Loops can't be freely extended at inference — changing the schedule from H2L3 to H3L3/H4L3 drops MMStar from 63.47 to 58.20/24.00. Parameter reuse also doesn't mean cheaper inference (128 vs 32 layer calls).

Original post →

LoopVL: 1B recurrent VLM trained on 0.14T tokens matches 36T-token models, shows 'Visual Aha Moment'

More from Multimodal