Looped layers don't help multimodal embeddings unless recurrence is pretrained (+4.23 points)
_reachsumit · x · 2026-11-18
An arXiv study tests whether running the same layers again improves multimodal embeddings. Across 11 loop designs, 3 checkpoints, and all 78 MMEB-V2 datasets, adding a loop to a standard model loses 0.46 points vs its control; adding an intermediate generative stage flips it to +0.36. Only a model pretrained with recurrence gains (+4.23 points over its loop-removed control). The authors prove query-independent image states persist under forward-only information flow and show query-conditioned visual updates matter mainly for visual Q&A.
More from Multimodal
- Real-time 3D worlds with Gaussian splats: 'the barriers of imagination are dissolving' — pvncher · 2026-11-18
- APOB AI pitches Seedream 5.5 workflow for AI idol album covers and K-pop MV teasers — aftahi_ai · 2026-11-18
- Ternary quantization shrinks VoxCPM2 5.3x to 827MB, runs real-time on laptop CPU — oguzeldereli · 2026-11-18
- Building a Claude skill for anime-style emotional framing techniques — andrew_n_carr · 2026-11-18
- sprite32: open-weights 32x32 text-to-sprite diffusion model trained from scratch — DaneMarguglio · 2026-11-18
- Claude makes videos at near-zero token cost: user builds automated film production engine — RileyRalmuto · 2026-11-18