Low-rank adaptive connections improve learning in frozen AI models
LARC: Low-Rank Adaptive Residual Connections for Learning in Frozen Models
Machine LearningComputation and Language
Summary
Many AI models are fixed after training and cannot learn from new feedback without retraining. This paper shows a way to add a small, adaptable piece to a large frozen model that lets it improve its decisions quickly based on feedback. The authors test this on tasks like choosing programs and show it can significantly reduce errors after a few updates. However, they also find that this method has limits in longer-term learning and decision consistency.
What this means in practice
- •For machine learning engineers: Add compact adaptive components to frozen large models for rapid task-specific improvements using feedback gradients.
- •For ai system operators: Deploy frozen AI models that can adapt quickly to changing inputs without full retraining by updating small trainable residual parameters.
Authors
Junyi Zou, Avrova Donz
Abstract
Low-Rank Adaptive Residual Connections (LARC) give a frozen model a compact numerical state that can learn from feedback. The map $h+BAh$ adds a low-rank correction to a hidden representation. A slow state $ρ$ learns starting factors across tasks; a private fast state $Φ$ copies them, changes with feedback, and resets to the trained initialization. This report specifies an input-side realization of the numerical policy carrier in Memory-Mediated Learning Architecture and examines its factor-space dynamics and learning lifetime. We study a rank-4 input residual with 12,288 trainable parameters on a frozen MiniCPM5-1B-SFT substrate. In a four-candidate program-selection task, two feedback-gradient steps reduce expected query execution error by 24.65 and 36.65 percentage points relative to resetting to the respective trained static and post-adaptation initializations. These development results cover 16 parameter groups and three paired training seeds. A direct support-loss selection rule is much more accurate, reaching 0.78125% error. In a repository-balanced chronological replay of public continuous-integration jobs, retaining online updates raises half-Brier loss from 0.1274 to 0.1808. A fixed follow-up intervention records same-batch non-descent and inconsistent future benefit from shrinking updates. Together, the algebra and measurements distinguish residual capacity, adaptation relative to a starting point, and usefulness on later decisions.