LoopVL: 1B-parameter VLM reuses 32 layers across 128 calls, beats 4B baselines on 5/8 benchmarks
jiqizhixin · x · 2026-10-29
Researchers from Renmin University, Baidu, Shanghai Jiao Tong and Tsinghua introduce LoopVL, arguing that letting the same parameters 'take another look' beats scaling up.
- Bottleneck: repeating a module only helps if the model's image understanding actually changes; fixed visual features mean the same evidence is missed every pass.
- Mechanism: LoopVL updates a shared vision-language state via two nested loops; its 1B language backbone reuses 32 unique Transformer layers across 128 layer calls, shifting visual attention as internal reasoning develops.
- Results: under a shared 0.14T training-token budget, it beats 4B deep and wide baselines on 5 of 8 benchmarks — RealWorldQA 70.98 vs 66.54/66.01, MathVision 38.49 vs 35.xx.
More from Models
- OpenAI lets ChatGPT Pro and Business subscribers claim up to $500/month in API credits — OpenAIDevs · 2026-10-29
- Researcher: GPT6 Astra brings a step change in review quality, guided prompts help further — anshulkundaje · 2026-10-29
- Paper Instruments launches Ultramarine 1 Flash, beating frontier models at up to 60x lower cost — baseten · 2026-10-29
- Gemini 4 reportedly "proved" nearly half of the false theorems it was given — RishabJainK · 2026-10-29
- Sentdex: DeepSeek v4.1 Flash now his top local token generator ever — Sentdex · 2026-10-29
- Fable 5 praised for blending deep scientific knowledge with storytelling delight — repligate · 2026-10-29