Specialist Model Composition
Methods for combining multiple task-specific models into a single model without full joint retraining.
6 papers
Where this stands
The written synthesis of this thread is for subscribers. Subscribe.
How this thread developed
Nov 2026 · Google
Image restoration models can share skills without full joint retraining
Combines image restoration specialists via small layer-specific adjustments without full retraining.
Nov 2026 · Rutgers University–New Brunswick
Image models can outperform specialized models at reading brain recordings
Shows an image model pretrained on general data can match or exceed specialized EEG models when paired with a small display module, demonstrating transfer across modalities.
Nov 2026 · Tohoku University
Merging AI models improves by removing conflicting feature directions
Presents Joint Null-Space Projection to merge fine-tuned checkpoints by removing conflicting feature directions, achieving high accuracy without data.
2 further papers
Nov 2026 · University of Illinois Urbana-Champaign, ANL
AI agents benefit more from ordered training histories than extra data
Shows that replaying training histories of specialists can consolidate them into a single agent without joint retraining.
4 of 6 papers shown