# Attention-Spectrum Regularization for Replay-Free Continual Multimodal LLMs

arXiv: https://arxiv.org/abs/2606.23063 · HTML: https://mlsanity.com/paper/2606.23063/

## Key facts

- Cited by 1 papers in the map (data through 14 Oct 2026).
- Placed in Prompt learning and adaptation of CLIP, one of 16 directions in Multimodal LLMs / vision-language models.
- Builds on 38 of its 52 references that are in the map.
- First posted on arXiv 22 Jun 2026.

## Placed in

- Multimodal LLMs / vision-language models / [Prompt learning and adaptation of CLIP](https://mlsanity.com/mllm/prompt-learning-and-adaptation-of-clip/)

## Builds on (most cited first, 30 of 38)

- [Visual Instruction Tuning](https://mlsanity.com/paper/2304.08485/) (2023)
- [Grad-CAM: Visual Explanations from Deep Networks via Gradient-based Localization](https://mlsanity.com/paper/1610.02391/) (2016)
- [Qwen2.5-VL Technical Report](https://mlsanity.com/paper/2502.13923/) (2025)
- [Overcoming catastrophic forgetting in neural networks](https://mlsanity.com/paper/1612.00796/) (2016)
- [Making the V in VQA Matter: Elevating the Role of Image Understanding in Visual Question Answering](https://mlsanity.com/paper/1612.00837/) (2016)
- [Learning without Forgetting](https://mlsanity.com/paper/1606.09282/) (2016)
- [ViLBERT: Pretraining Task-Agnostic Visiolinguistic Representations for Vision-and-Language Tasks](https://mlsanity.com/paper/1908.02265/) (2019)
- [InternVL3: Exploring Advanced Training and Test-Time Recipes for Open-Source Multimodal Models](https://mlsanity.com/paper/2504.10479/) (2025)
- [LXMERT: Learning Cross-Modality Encoder Representations from Transformers](https://mlsanity.com/paper/1908.07490/) (2019)
- [Paying More Attention to Attention: Improving the Performance of Convolutional Neural Networks via Attention Transfer](https://mlsanity.com/paper/1612.03928/) (2016)
- [Dark Experience for General Continual Learning: a Strong, Simple Baseline](https://mlsanity.com/paper/2004.07211/) (2020)
- [Orthogonal Subspace Learning for Language Model Continual Learning](https://mlsanity.com/paper/2310.14152/) (2023)
- [On Tiny Episodic Memories in Continual Learning](https://mlsanity.com/paper/1902.10486/) (2019)
- [Stacked Attention Networks for Image Question Answering](https://mlsanity.com/paper/1511.02274/) (2015)
- [LVLM-eHub: A Comprehensive Evaluation Benchmark for Large Vision-Language Models](https://mlsanity.com/paper/2306.09265/) (2023)
- [PLOP: Learning without Forgetting for Continual Semantic Segmentation](https://mlsanity.com/paper/2011.11390/) (2020)
- [Preventing Zero-Shot Transfer Degradation in Continual Learning of Vision-Language Models](https://mlsanity.com/paper/2303.06628/) (2023)
- [CoIN: A Benchmark of Continual Instruction tuNing for Multimodel Large Language Model](https://mlsanity.com/paper/2403.08350/) (2024)
- [Continual Vision-Language Representation Learning with Off-Diagonal Information](https://mlsanity.com/paper/2305.07437/) (2023)
- [TiC-CLIP: Continual Training of CLIP Models](https://mlsanity.com/paper/2310.16226/) (2023)
- [CL-MoE: Enhancing Multimodal Large Language Model with Dual Momentum Mixture-of-Experts for Continual Visual Question Answering](https://mlsanity.com/paper/2503.00413/) (2025)
- [Symbolic Replay: Scene Graph as Prompt for Continual Learning on VQA Task](https://mlsanity.com/paper/2208.12037/) (2022)
- [Psycholinguistics meets Continual Learning: Measuring Catastrophic Forgetting in Visual Question Answering](https://mlsanity.com/paper/1906.04229/) (2019)
- [SEFE: Superficial and Essential Forgetting Eliminator for Multimodal Continual Instruction Tuning](https://mlsanity.com/paper/2505.02486/) (2025)
- [HiDe-LLaVA: Hierarchical Decoupling for Continual Instruction Tuning of Multimodal Large Language Model](https://mlsanity.com/paper/2503.12941/) (2025)
- [CTP: Towards Vision-Language Continual Pretraining via Compatible Momentum Contrast and Topology Preservation](https://mlsanity.com/paper/2308.07146/) (2023)
- [Dynamic Mixture of Curriculum LoRA Experts for Continual Multimodal Instruction Tuning](https://mlsanity.com/paper/2506.11672/) (2025)
- [Cerberus Transformer: Joint Semantic, Affordance and Attribute Parsing](https://mlsanity.com/paper/2111.12608/) (2021)
- [VIBUS: Data-efficient 3D Scene Parsing with VIewpoint Bottleneck and Uncertainty-Spectrum Modeling](https://mlsanity.com/paper/2210.11472/) (2022)
- [KeepLoRA: Continual Learning with Residual Gradient Adaptation](https://mlsanity.com/paper/2601.19659/) (2026)

## Built on by (most cited first, 1 of 1)

- [MMOOC: A Comprehensive Benchmark for Out-of-Context Evaluation in Multimodal Large Language Models](https://mlsanity.com/paper/2607.27637/) (2026)

## Abstract (arXiv)

Multimodal large language models (MLLMs) are increasingly required to adapt to non-stationary streams of visual domains, question types, and user instructions, yet continual fine-tuning often causes severe forgetting of previously acquired multimodal skills. Existing continual vision-language methods mainly preserve outputs, replay data or pseudo-data, regularize embedding geometry, or allocate task-specific parameters, but they provide limited control over how internal cross-modal attention patterns supporting old skills drift during adaptation. We propose Attention-Spectrum Regularization (ASR), a replay-free continual learning framework that preserves skill-conditioned structures of cross-modal attention. ASR treats cross-attention maps as two-dimensional signals, summarizes their scale and directional properties into compact spectral statistics, and stores only skill-wise prototype distributions instead of replaying past image-question pairs, generated pseudo-examples, or old-stage teacher snapshots. In later stages, a phase-invariant spectral regularizer constrains harmful drift of these prototypes while allowing instance-level attention to adapt to new tasks. We provide theoretical analysis showing that skill-conditioned spectral drift controls forgetting under a spectral sufficiency assumption, and that Fourier power spectra are stable to spatial translations and bounded perturbations. Experiments on continual VQA and multimodal instruction-tuning benchmarks, including VQA v2, VQACL, CLT-VQA, CoIN, and UCIT, show that ASR consistently improves final performance and reduces forgetting over strong replay-, regularization-, and adapter-based baselines. Preserving skill-level attention structure is an effective and lightweight mechanism for continual MLLMs. Code is available at https://github.com/Creative-zcx/attention-spectrum-replay