few-shot time-series adaptation

This work addresses the significant performance degradation of conventional deep learning models in time series classification due to distributional shifts—commonly referred to as data drift—between training and test data, which often necessitates extensive labeled data and costly retraining. To systematically evaluate adaptation strategies under such conditions, the authors introduce SeisTask, a task-oriented benchmark based on seismic data, and compare optimization-based meta-learning against standard fine-tuning. Experimental results demonstrate that meta-learning achieves faster and more stable adaptation in low-data and small-model regimes; however, this advantage diminishes as both dataset size and model capacity increase. Furthermore, the study reveals that aligning task distributions is more critical for performance gains than maximizing task diversity.

This work addresses the high computational cost and inefficiency of traditional meta-reinforcement learning, which relies on trajectory collection and gradient updates in the inner loop. The authors propose LA-MAML, a novel approach that leverages task-specific language instructions as direct signals within the inner loop, enabling rapid task adaptation without requiring environmental interaction or gradient-based updates. By embedding language instructions to modulate policy network parameters in a single step, LA-MAML deeply integrates linguistic guidance with policy learning within the MAML framework. Evaluated on the BabyAI benchmark, the method achieves performance comparable to or better than existing baselines while substantially reducing per-iteration training time, thereby significantly enhancing meta-training efficiency.

To address the weak cross-task generalization and insufficient robustness of meta-learning models in few-shot learning scenarios, this paper proposes DGS-MAML—a meta-learning algorithm integrating domain generalization and sharpness-aware optimization. Within a bilevel optimization framework, DGS-MAML jointly models gradient matching (to improve task adaptability) and sharpness minimization (to enhance parameter robustness). It provides, for the first time for MAML-style methods, PAC-Bayes generalization bounds and convergence guarantees. Extensive experiments on multiple few-shot benchmark datasets demonstrate that DGS-MAML consistently outperforms mainstream approaches—including MAML and MetaReg—with average accuracy gains of 2.3–5.1%. Moreover, it exhibits superior robustness under distributional shift. The implementation is publicly available.

Existing LLM-based agent workflows rely on static templates or manual design, suffering from limited generalization and scalability. Method: We propose a natural-language-driven meta-learning framework—the first to integrate Model-Agnostic Meta-Learning (MAML) into language agent workflow optimization. Our approach employs subtask-level adaptive initialization and bi-level optimization: inner-loop fine-tuning for task-specific adaptation and outer-loop updates of the shared initialization to enable dynamic workflow evolution. Crucially, workflow modifications are guided entirely by LLM-generated feedback and natural-language instructions, eliminating manual intervention. Contribution/Results: Evaluated on question answering, code generation, and mathematical reasoning, our method consistently outperforms both handcrafted and automated search baselines, achieving multiple state-of-the-art results. It significantly enhances cross-task and cross-model generalization, demonstrating robust adaptability without human-designed structures.

This work challenges the prevailing consensus that pretraining (PT) universally outperforms meta-learning in few-shot learning. Method: We conduct a systematic, fair comparison between PT and model-agnostic meta-learning (MAML) across 21 diverse few-shot benchmarks—including Meta-Dataset and large-scale language benchmarks (e.g., GPT-2 + OpenWebText)—using identical architectures, optimizers, and convergence criteria. Contribution/Results: We identify dataset structural diversity as the decisive factor governing relative performance: PT slightly dominates under low diversity, whereas MAML surpasses PT under high diversity; no method exhibits consistent superiority across all 21 benchmarks. Crucially, we adopt Cohen’s *d* effect size (*d* < 0.2) instead of *p*-values to rigorously quantify negligible average performance differences. Results hold robustly across both vision and language domains, demonstrating that neither paradigm is universally superior—performance hinges critically on task distribution characteristics.

Addressing key bottlenecks in the Algorithm Selection Problem (ASP)—including high evaluation costs, insufficient pipeline diversity, and severe sampling bias in existing meta-datasets (e.g., OpenML)—this work introduces PIPES, a large-scale, balanced meta-dataset. PIPES encompasses 300 diverse datasets and 9,408 systematically designed machine learning pipelines, spanning broad preprocessing–model combinations. It is the first to achieve comprehensive, bias-mitigated, standardized pipeline metadata collection, fully recording module specifications, execution time, predictions, performance metrics, and failure logs. A unified metadata schema enables deep cross-pipeline and cross-dataset analysis. PIPES substantially enhances reproducibility and scalability in meta-learning research. All code and experimental data are publicly released.

This work addresses the challenges faced by learning-based systems in heterogeneous, dynamic, and long-running environments, where environmental shifts often lead to high retraining costs, substantial labeling overhead, performance degradation, and sluggish adaptation. To tackle these issues, the paper introduces EMA—a lightweight, adaptive framework that supports diverse system and model architectures through a system-driven, data-centric approach. EMA employs a state transformer to align representations between old and new environments, enabling warm-start model adaptation, and intelligently prioritizes high-utility data samples for labeling based on their expected contribution to performance. Experimental evaluation across eight representative systems demonstrates that EMA reduces adaptation costs—such as GPU training time—by 14.9% to 42.4% while simultaneously improving system performance metrics, including network throughput, by 6.9% to 31.3%.

This work addresses the inefficiencies in large-scale recommendation systems caused by maintaining separate models for different scenarios and objectives, which hinders development velocity and delays technology adoption. To overcome this, the authors propose the Standardized Model Template (SMT) framework, which leverages composable, standardized machine learning components to enable “design once, deploy everywhere,” uniformly accommodating diverse data distributions and optimization objectives. By decoupling model architecture from scenario-specific configurations, SMT reduces the complexity of technology deployment from O(n·2ᵏ) to O(n+k), breaking away from the conventional “one objective, one model” paradigm. Empirical evaluation on Meta’s ad ranking system demonstrates that SMT improves average cross-entropy by 0.63%, reduces engineering time per model iteration by 92%, and increases the throughput of technology-model pair adoption by 6.3×.

Existing time series pretraining methods struggle to generalize effectively across multiple datasets due to discrepancies in input length and channel dimensions. This work proposes ADAPT, a novel pretraining paradigm that enables unified modeling across 162 time series classification datasets by adaptively aligning the physical attributes of time series data. Integrating self-supervised learning with a hybrid batch training strategy, ADAPT overcomes the generalization limitations inherent in conventional many-to-one pretraining approaches. The method achieves state-of-the-art performance on multiple benchmarks, establishing a foundational framework for developing general-purpose foundation models for time series analysis.

This study addresses the challenge of rapid cross-task adaptation in robotic manipulation tasks. Method: We evaluate a meta-reinforcement learning approach combining Model-Agnostic Meta-Learning (MAML) with Trust Region Policy Optimization (TRPO) on the MetaWorld ML10 benchmark, proposing a MAML-based framework for universal policy initialization that enables one-step gradient adaptation to novel manipulation tasks—including pushing, grasping, and drawer opening—thereby substantially reducing adaptation overhead. Contribution/Results: During meta-training, the method achieves a task success rate of 21.0%; in zero-shot transfer to held-out test tasks, it attains 13.2% success, confirming effective single-step adaptation. Further analysis reveals heterogeneous generalization performance across tasks, indicating that structured policy representations—such as modular architectures or task embeddings—are critical for enhancing cross-task robustness. This work provides empirical validation and design insights for efficient meta-policy learning targeting diverse robotic manipulation behaviors.