New paper probes multimodal MoE expert structure to make model adaptation much faster
georgiagkioxari · x · 2026-10-06
Researchers including Georgios Gkioxari published a new blog post and paper examining what internal structure experts in multimodal MoEs develop during pre-training.
The work explores whether this learned structure can be put to practical use to make model adaptation significantly faster. Details and results are in the accompanying paper.
More from Research
- AgentIR retriever reads agent thinking tokens, lifts BrowseComp-Plus to 67% from 35% — CShorten30 · 2026-10-06
- OneCanvas turns a single image into 3D spatial reasoning, hitting new SOTA — Roger_M_Taylor · 2026-10-06
- Hugging Face researcher: encrypted reasoning is the worst thing to happen to LLM evals — _lewtun · 2026-10-06
- WM-VLM: world model generates visual intermediate states for spatial reasoning — ZhitingHu · 2026-10-06
- ProAR turns autoregressive video models into goal-oriented prospective reasoners — muhao_chen · 2026-10-06
- 90+ Claude Opus 5.5 agents find two room-temperature magnetic semiconductor candidates in 3 days — Dr_Singularity · 2026-10-06