A new survey paper explores Unsupervised Post-Training (UPT), a method for adapting foundation models using unlabeled data and internal model artifacts rather than external human labels or oracles. The paper categorizes 80 distinct UPT methods based on the source of their learning signal, such as prediction statistics or self-generated targets. It also introduces a framework that maps deployment regimes by considering input visibility and update persistence, aiming to guide the selection and evaluation of UPT techniques and prevent recursive error amplification. AI
IMPACT Provides a structured overview of unsupervised adaptation techniques for large models, aiding researchers in selecting and evaluating methods.
RANK_REASON The cluster contains a survey paper on a specific machine learning technique. [lever_c_demoted from research: ic=1 ai=1.0]
- alphaXiv
- arXiv
- CatalyzeX
- DagsHub
- foundation model
- Gotit.pub
- Hugging Face
- ScienceCast
- Unsupervised Post-Training
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
