Task-Anchored Representation Shaping for Pre-Trained Model-Based Continual Learning

2026-08-17

AI Summary

This study addresses cross-task prediction ambiguity and unreliable inference in continual learning with pretrained models by proposing the TAILS module. TAILS employs fixed task anchors to construct a latent recall mechanism that directly rectifies feature representations prior to prediction, thereby eliminating cross-task ambiguity and achieving representational decoupling. A key advantage of this approach is its compatibility with frozen base models and minimal computational overhead. Extensive experiments demonstrate that TAILS significantly enhances classification accuracy and task reasoning performance across multiple continual learning paradigms. By combining high parameter efficiency with negligible inference costs, this work establishes a novel and effective paradigm for efficient continual learning.

Abstract

Pre-trained models (PTMs) provide a strong foundation for continual learning by offering stable representations that facilitate lightweight adaptation to new tasks. However, adapting well to each task does not ensure reliable inference over all learned tasks. Since task boundaries are often artificial and semantically entangled, an input from an unknown task can remain ambiguous even with strong PTM features, making cross-task prediction a key bottleneck. We propose Task-Anchored Inference Latent Shaping (TAILS), a lightweight post-PTM module that can be integrated into diverse continual learners and optimized through a decoupled step. TAILS uses fixed task anchors as persistent references to accumulated knowledge. It interprets each sample's feature representation relative to these references, then composes relevant evidence across tasks into latent recall. Rather than selecting a task-specific path or adjusting classifier outputs, TAILS uses latent recall to directly correct the feature representation before prediction. It therefore resolves cross-task ambiguity at the representation level, while leaving the original PTM, method-specific modules, and classifier unchanged. Extensive experiments across multiple PTM-based continual learning paradigms show that TAILS can improve classification and task-inference performance with modest parameter overhead and negligible inference cost.

Problem

• Continual Learning
• Pre-Trained Models
• Cross-Task Ambiguity
• Task Inference

Innovation

• Continual Learning
• Pre-Trained Models
• Representation Shaping
• Task Anchors
• Latent Recall