Training-Free MoE Merging Paper DUME Accepted to NeurIPS 2026
benfielding · x · 2026-09-29
The team behind DUME (Training-Free Dynamic Upcycling of Expert Language Models) announced the paper's acceptance to NeurIPS 2026.
DUME is a training-free LLM routing method that combines multiple already-trained domain expert models into a single Mixture-of-Experts model, eliminating the post-merge training step typically required. Authored by @oguzer90 and @ErosFani.
More from Research
- TimeEvo: failure-driven tool synthesis lifts time series agent accuracy on every task and backbone — Jie Yang · 2026-09-29
- LastOPD: latent on-policy distillation collapses late, last-layer-only signal gains 5.55 on MATH-500 — Jie Yang · 2026-09-29
- Protein design challenge ups ante: pH-dependent EGFR binders with mouse cross-reactivity — nc_frey · 2026-09-29
- Weaviate Podcast: Bradley Allen on 50 years of knowledge engineering and neurosymbolic AI — CShorten30 · 2026-09-29
- Anthropic opens Bay Area wet lab where Claude-driven robots run biology experiments — dl_weekly · 2026-09-29
- MATM framework lets LLM agent populations share and reuse task trajectories — 841io · 2026-09-29