TUNE: Frequency-Guided Token Gating for Robust Continual Learning in LLMs

Sumin Park, Noseong Park

Keywords: Continual Learning, Catastrophic Forgetting, Spectral Theory, Frequency, Model Robustness, LLMs

TL;DR: We propose TUNE, a frequency-guided token gating method that leverages wavelet-based spectral decomposition to stabilize LoRA updates and mitigate catastrophic forgetting in continual LLM training.

Abstract:

Continual learning (CL) in large language models (LLMs) remains a critical challenge, as sequential training often results in catastrophic forgetting of previously learned knowledge. To our knowledge, no prior work has approached CL in LLMs from a frequency perspective, despite strong evidence that spectral properties directly govern model robustness and vulnerability to forgetting.
Recent methods based on Low-Rank Adaptation (LoRA) have shown promise for parameter-efficient CL, but remain preliminary, relying on task-specific subspace expansion with additional regularization.
We propose {TUNE} (Token Update via Noise-robust Frequency Encoding), a frequency-guided token modulation mechanism that stabilizes LoRA residual updates. TUNE employs a stationary wavelet transform (SWT) to decompose token embeddings into multi-resolution subbands, where token saliency is derived from high-frequency activations and reliability is assessed through cross-scale agreement. These signals are fused into token-wise scaling that amplify reliable updates while suppressing noisy fluctuations.
Without introducing additional trainable parameters beyond LoRA expansion, TUNE achieves significant improvements over the SOTA baselines, establishing frequency-aware token adaptation as a promising direction for CL in LLMs.

Primary Area: transfer learning, meta learning, and lifelong learning

Submission Number: 8554