Continual Fine-Tuning with Provably Accurate and Parameter-Free Task Retrieval

2026-03-17 · Source: cs.LG updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Data Science & Analytics · Depth: Expert, extended

Summary

Researchers from Washington State University and Princeton University introduce Proteus, a novel parameter-adaptation method for continual fine-tuning that combines adaptive input embedding use with parameter-free task retrieval. Continual fine-tuning aims to adapt pre-trained models to new tasks sequentially while retaining performance on prior tasks without access to their data. Existing methods either suffer from forgetting in retrieval functions (input-adaptation) or lack representation adaptability (parameter-adaptation). Proteus addresses these issues by deriving theoretical guarantees for a clustering-based, parameter-free retrieval paradigm, linking low retrieval error to well-organized task-specific representation clusters. The method features an adaptive module composition strategy that learns orthogonal task-specific updates and a clustering-based retrieval mechanism capturing distinct representation signatures. Extensive experiments on benchmarks like CIFAR-100, ImageNet-R, and VTAB demonstrate that Proteus consistently outperforms state-of-the-art baselines, achieving up to 57% gains in retrieval and 30% in classification performance, along with the best average forgetting metric, while maintaining low GPU memory consumption.

Key takeaway

For research scientists developing continual learning systems, Proteus offers a robust framework to mitigate catastrophic forgetting and improve performance. You should consider implementing its parameter-free, clustering-based retrieval and adaptive LoRA fine-tuning with orthogonality constraints. This approach provides theoretical guarantees for low retrieval error and has demonstrated superior accuracy and scalability across diverse task scenarios, making it a strong candidate for adapting large pre-trained models to evolving task sequences.

Key insights

Proteus offers provably accurate, parameter-free task retrieval for continual fine-tuning by leveraging distinct representation signatures.

Principles

• Low retrieval error correlates with well-separated representation clusters.
• Orthogonal task-specific updates enhance cluster separation.
• Parameter-free retrieval mitigates catastrophic forgetting.

Method

Proteus uses LoRA-based adaptive fine-tuning with orthogonal knowledge transfer, combined with a Dirichlet Process Gaussian Mixture Model (DP-GMM) for parameter-free, clustering-based task retrieval at inference.

In practice

• Use DP-GMM for multi-modal input embedding distributions.
• Apply L1/L2 norm penalties to promote selective knowledge transfer.
• Enforce orthogonality in LoRA updates for better task separation.

Topics

• Continual Fine-Tuning
• Parameter-Free Retrieval
• Low-Rank Adaptation
• Representation Clustering
• Catastrophic Forgetting

Code references

• zhoudw-zdw/RevisitingCIL

Best for: Research Scientist, AI Researcher, AI Scientist, Machine Learning Engineer

Related on AIssential

• Octopus: History-Free Gradient Orthogonalization for Continual Learning in Multimodal Large Language Models — Octopus enables continual learning in MLLMs by orthogonalizing gradients without needing historical data.
• Evaluating the Impact of Task Granularity on Catastrophic Forgetting in Continual Learning — Learning general categories before specific classes may reduce catastrophic forgetting in continual learning systems.
• RAFT: Data Refinement and Adaptive Distillation for Domain Fine-Tuning with Alleviated Forgetting — RAFT improves domain fine-tuning by refining data and preserving original model behavior through adaptive distillation.
• Catastrophic Forgetting is Low-Rank: A Function-Space Theory for Continual Adaptation — Catastrophic forgetting is a low-rank function-space phenomenon, predictable via a cross-task kernel in the NTK regime.
• Learning, Fast and Slow: Towards LLMs That Adapt Continually [R] — Combining slow parameter updates with fast context optimization enhances LLM adaptation and mitigates forgetting.
• Teaching AI to Remember: What I Learned Building a Continual Learning Framework for Drug Discovery — Continual learning models can avoid catastrophic forgetting by integrating new knowledge with awareness of existing representations.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.LG updates on arXiv.org.