New LARC method enables frozen models to learn from feedback

Researchers have developed Low-Rank Adaptive Residual Connections (LARC), a method to enable frozen models to learn from feedback by adding a compact numerical state. This approach utilizes a low-rank correction to a hidden representation, incorporating a slow state for learning starting factors across tasks and a private fast state that adapts with feedback. Experiments on a MiniCPM5-1B-SFT model demonstrated significant error reduction in program selection tasks and highlighted the distinction between residual capacity, adaptation, and future decision-making usefulness. AI

IMPACT Enables adaptation of pre-trained models without full retraining, potentially reducing computational costs for fine-tuning.

RANK_REASON The item describes a new research paper detailing a novel method for adapting frozen models. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.CL →

• alphaXiv
• arXiv
• CatalyzeX
• Connected Papers
• CORE Recommender
• DagsHub
• Gotit.pub
• Hugging Face
• IArxiv Recommender
• Litmaps
• Memory-Mediated Learning Architecture
• MiniCPM5-1B-SFT
• ScienceCast
• scite Smart Citations

• paper
• model release

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →