Dynamic Meta-Prompting for Continual Learning

• Dynamic meta-prompting is a framework that uses real-time prompt generation to adapt large pre-trained models to sequential task shifts without relying on rehearsal.
• It balances learning plasticity and stability through prompt synthesis, selection, grouping, and dynamic allocation to mitigate catastrophic forgetting.
• Empirical results across vision, language, multimodal, and reinforcement learning benchmarks demonstrate state-of-the-art performance with minimal parameter overhead.

Dynamic meta-prompting for continual learning refers to a structured family of algorithms that dynamically generate, modulate, or select prompt representations to drive rehearsal-free, high-fidelity adaptation of large pre-trained models subjected to sequential task or domain shifts. These strategies have emerged as a parameter-efficient alternative to naive fine-tuning or buffer-based replay, designed to balance plasticity and stability and directly combat catastrophic forgetting. The paradigm is applicable across vision, language, multimodal, and reinforcement learning scenarios, and has yielded state-of-the-art results through a variety of prompt synthesis, selection, mixture, and meta-grouping mechanisms.

1. Formal Definition and Core Concepts

Dynamic meta-prompting frameworks operate within a continual learning (CL) protocol: the model is exposed to a stream of tasks $\{D_1, ..., D_T\}$, where each $D_t$ may introduce new classes, domains, or data distributions. The backbone model (typically a large transformer) is either entirely frozen or only lightly adapted (e.g., using adapters or LoRA), and a prompt parameterization—either fixed-size or dynamically expandable—is leveraged as the only or primary mechanism for encoding new knowledge and transferring or preserving previously learned representations.

Meta-prompting differs from static prompt selection by:

• Synthesizing prompt tokens on the fly conditioned on the current input and/or task state (e.g., Prompt Customization (Dai et al., 28 Apr 2024), DPaRL (Kim et al., 9 Sep 2024));
• Dynamically grouping, fusing, or evolving prompt sets based on measured task similarity (e.g., AdaPromptCL (Kim et al., 2023), ConvPrompt (Roy et al., 29 Mar 2024), POP (Hu et al., 2023));
• Architecting hierarchical or expert mixtures that allow both generalization across tasks and specialization for current data (e.g., SMoPE (Le et al., 29 Sep 2025), RainbowPrompt (Hong et al., 30 Jul 2025)); or
• Using gradient-based or correlation-based meta-objectives to regulate prompt pool growth, sharing, or attention (e.g., LW2G (Feng et al., 27 Sep 2024), ModalPrompt (Zeng et al., 8 Oct 2024)).

The prompts themselves may be shared, task-specific, group-specific, or represented as a sparse composition or convex combination of the above. The meta-prompting process thus encapsulates both prompt generation (at training and inference), prompt adaptation, and, if applicable, prompt recycling and lifelong update strategies.

2. Representative Architectures and Algorithms

Numerous dynamic meta-prompting frameworks have been established in state-of-the-art CL literature:

• Prompt Of Prompts (POP): Explicitly separates global (meta) and task-specific prompt pools. Each new task learns a dedicated prompt set while a global meta-prompt pool is continually adapted to integrate cross-task knowledge, and the output is composed by concatenation and mean-pooling of each prompt block (Hu et al., 2023).
• Prompt Customization (PC): Maintains a fixed prompt codebook $M$ that is re-combined on a per-instance basis by a Prompt Generation Module (PGM) and modulated by a Prompt Modulation Module (PMM), thus synthesizing instance- and task-adaptive prompts via soft attention and correlation-based scaling (Dai et al., 28 Apr 2024).
• ConvPrompt: Implements prompt generators as convolutional filters over shared embeddings (enabling fine-grained, layer-specific prompt composition) and uses LLM-generated text to estimate inter-task similarities and dynamically allocate prompt capacity, which greatly reduces unnecessary prompt proliferation (Roy et al., 29 Mar 2024).
• AdaPromptCL: Groups tasks semantically via “assign-and-refine” clustering over prompt-derived embeddings—using coarse/fine clustering to decide meta-prompt sharing and supervised key-matching for inference (Kim et al., 2023).
• RainbowPrompt: Aggregates diverse, task-specific base prompts into an evolved, aligned unified meta-prompt per layer, with a learnable probabilistic gate to determine which prompt representations to insert at each transformer block (Hong et al., 30 Jul 2025).
• SMoPE: Views a long shared prompt as a set of “prompt experts” and uses a prompt-MoE (mixture-of-experts) gating architecture to dynamically select a sparse, input-dependent subset for activation, enhanced by expert load balancing and a prototype loss (Le et al., 29 Sep 2025).

In all these frameworks, the dynamic or meta aspect lies in learning to allocate representational and compositional power adaptively across tasks, instances, or input conditions.

3. Meta-Prompt Synthesis, Selection, and Modulation

Dynamic meta-prompting approaches can be differentiated by how they synthesize, select, and modulate prompts:

• Synthesis by soft compositionality: Instance-level prompt tokens are generated as convex combinations across a codebook with instance- or task-specific coefficients (e.g., PC (Dai et al., 28 Apr 2024)).
• Prompt group assignment and refinement: New tasks are assigned to prompt groups or clusters based on semantic similarity (from prompt embeddings or LLM-generated descriptions), and group prompts are tuned coarsely and then refined within group or via clustering (e.g., AdaPromptCL (Kim et al., 2023), ConvPrompt (Roy et al., 29 Mar 2024)).
• Attention or mixture-of-experts selection: Large prompt pools are organized so only a subset is used per input/task—via attention-guided routing (RainbowPrompt (Hong et al., 30 Jul 2025), ModalPrompt (Zeng et al., 8 Oct 2024)) or sparse gating (SMoPE (Le et al., 29 Sep 2025)).
• Gradient-based meta-control: Metrics such as the Hinder Forward Capability (HFC) are computed by projecting gradients into the orthogonal complement of previous task subspaces to determine whether to allocate a new prompt set or reuse existing sets (LW2G (Feng et al., 27 Sep 2024)).
• Contrastive/correlation-based modulation: Secondary alignment or contrastive objectives act on the prompt representations to encourage diversity, cross-task coherence, or to mitigate redundancy (e.g., contrastive task interaction in multimodal CL (Guo et al., 1 Mar 2025)).

A key property emerging from these strategies is the balance between representational sufficiency (plasticity) and containment (stability): prompt allocation adapts to task similarity, and freeze/adaptation decisions are made so as not to pollute old task knowledge.

4. Empirical Results and Performance Impact

Dynamic meta-prompting has demonstrated broad empirical improvements in diverse CL benchmarks. Selected results include:

Method	Dataset	Avg. Accuracy (%)	Forgetting (%)	Parameter Cost (%)
ConvPrompt	CIFAR-100	88.87	4.75	2.0 (of model)
L2P	CIFAR-100	83.8	7.6	0.7
POP	CIFAR-100	82.0	—	m=1 per task + global
PC	CIFAR-100	87.20	—	≪1
RainbowPrompt	ImageNet-R	79.09	—	—
SMoPE	ImageNet-R	79.32	—	0.38M (much lower than L2P++)

Source: (Roy et al., 29 Mar 2024, Dai et al., 28 Apr 2024, Hu et al., 2023, Hong et al., 30 Jul 2025, Le et al., 29 Sep 2025, Wang et al., 2021).

Critical observations from these results:

• Dynamic meta-prompting achieves both parameter efficiency (2–4% overhead in best cases) and high accuracy, outperforming or matching buffer-free and many rehearsal-based methods.
• Prompt pool growth is either mitigated by adaptive allocation or managed by rich sharing and grouping strategies—prompt inflation is avoided.
• Forgetting is consistently lower compared to regularization or naive prompt allocation baselines.
• Strategies are robust to task and domain shift, including open-world and task-agnostic settings (Kim et al., 9 Sep 2024, Kim et al., 2023).

5. Applications beyond Standard Supervised CL

Dynamic meta-prompting is not limited to image classification or canonical class-incremental learning but generalizes to:

• Multimodal continual learning: ModalPrompt integrates dual-modality guided prompt selection (image/text) and prompt fusion to enable LMMs to transfer knowledge efficiently and mitigate forgetting; retrieval is modality-aware and computationally constant in task size (Zeng et al., 8 Oct 2024).
• Missing modality adaptation: Three-level prompt architectures (modality-specific, task-aware, task-specific) combined with contrastive alignment enable continual adaptation under missing modalities and outperform both prompt and replay-based baselines (Guo et al., 1 Mar 2025).
• Reinforcement learning: In sparse-prompted meta-policy frameworks, continual task allocation is realized by inferring sparse mask prompts per task, inducing disjoint sub-networks in a shared policy (Yang et al., 2023).
• Parameter-efficient backbones: Dynamic meta-prompting naturally integrates with adapter-tuning (as in FM-LoRA), dynamic rank selection, and hybrid architectures, providing further robustness against catastrophic forgetting (Yu et al., 9 Apr 2025).

This versatility suggests dynamic meta-prompting is a substrate for efficient continual adaptation in domains with evolving data, incomplete input, or shifting supervisory signal.

6. Theoretical Underpinnings, Hyperparameters, and Practical Considerations

The principal theoretical insights underpinning dynamic meta-prompting include:

• Stability–plasticity balance: Inductive isolation of learned representations—via prompt freezing or sparse activation—effectively partitions the parameter space and drastically reduces interference (Yang et al., 2023, Feng et al., 27 Sep 2024).
• Task similarity estimation: Meta-prompt sharing is optimized by clustering or similarity metrics (from LLM-generated attributes or prompt embeddings), ensuring parameter sharing tracks real semantic overlap (Roy et al., 29 Mar 2024, Kim et al., 2023).
• Dynamic capacity allocation: Adaptive growth heuristics (e.g., HFC, DRS) prevent unbounded prompt pool expansion and ensure computational efficiency even in long task sequences (Feng et al., 27 Sep 2024, Yu et al., 9 Apr 2025).

Key hyperparameters include:

• Prompt length and shape (layerwise, prefix), typically proportional to the number of tasks or drift magnitude (Yu et al., 9 Apr 2025).
• Pool size/codebook width for PC or L2P-style architectures, tuned based on task diversity (Dai et al., 28 Apr 2024, Wang et al., 2021).
• Sparsity (number of active experts) in mixture-of-experts designs (Le et al., 29 Sep 2025).
• Grouping threshold, regularization multipliers (orthogonality, diversity), and overlap parameters for prompt allocation and refinement (Roy et al., 29 Mar 2024, Kim et al., 2023).

A core best practice is to jointly train meta-prompts across tasks if possible and to employ small learning rates and weight decay for meta-prompt parameters to promote stability without excessive overspecialization.

7. Current Limitations and Open Problems

Despite the demonstrated advances, several limitations remain:

• Prompt pool scaling: Task-specific or task-pairwise prompt allocation, if unregulated, risks linear (or worse) parameter growth; thus, continual prompt sharing/grouping or capacity regularization remains an active research area (Le et al., 29 Sep 2025, Feng et al., 27 Sep 2024).
• Selection accuracy: Effective prompt selection (retrieval accuracy) remains a challenge, particularly with high task overlap or in the “open-world” regime; enhanced routing and group adaptation are areas of active investigation (Feng et al., 27 Sep 2024, Kim et al., 2023).
• Diversity–integration trade-off: Over-aggregation of prompts can dilute specificity, while over-specialization limits transfer; two-stage aggregation/alignment (e.g., RainbowPrompt (Hong et al., 30 Jul 2025)) and mixture-of-experts strategies (SMoPE (Le et al., 29 Sep 2025)) represent approaches to balance these forces.
• Generalization to other domains: Application of dynamic meta-prompting beyond vision and text, e.g., to structure, time series, or multi-agent problems, remains underexplored.
• Complex continual learning regimes: Handling extremely long task streams, task-agnostic settings, or high-modality inputs efficiently without loss of retrieval accuracy or prompt collapse remains nontrivial.

Overall, dynamic meta-prompting for continual learning forms a rapidly maturing yet still evolving framework that enables parameter-efficient, robust continual adaptation across diverse domains through flexible prompt synthesis, selection, grouping, and evolution. It stands as a central component in next-generation rehearsal-free continual learning algorithms in both unimodal and multimodal contexts, as documented across a series of foundational and recent works (Hong et al., 30 Jul 2025, Le et al., 29 Sep 2025, Kim et al., 2023, Roy et al., 29 Mar 2024, Dai et al., 28 Apr 2024, Zeng et al., 8 Oct 2024, Yu et al., 9 Apr 2025, Feng et al., 27 Sep 2024, Hu et al., 2023, Wang et al., 2021, Kim et al., 9 Sep 2024, Guo et al., 1 Mar 2025, Yang et al., 2023).