MetaCaster: Meta-Harness-Optimized Agent for End-to-End Few-Shot Learning of Lightweight Time Series Forecasters

arXiv:2608.23473 · cs.LG, cs.AI · Submitted 2026-08-24 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.

Tom: Next we'll be talking about the paper "MetaCaster: Meta-Harness-Optimized Agent for End-to-End Few-Shot Learning of Lightweight Time Series Forecasters".

Jane: The paper was written by ChengAo Shen, Wenchao Yu, Fangyu Wu, Dongjin Song, Hanghang Tong et al. from University of Houston and NEC Labs and University of Waterloo and University of Connecticut and University of Illinois at Urbana-Champaign and Singapore Management University.

Tom: Stay tuned as we take you through the paper and discuss its implications.

Summary: Tom: Now, let’s talk about the summary of MetaCaster: Meta-Harness-Optimized Agent for End-to-End Few-Shot Learning of Lightweight Time Series Forecasters, and how it actually functions under the hood.

Jane: The paper breaks down a multi-agent system that starts with a limited support set Dsup and some contextual description C.

Lu: The core mechanism involves two agents: MGAGENT, which generates a sufficient dataset D̄, and FTAGENT, which trains and selects the best lightweight forecasters using that data.

Meng: I’m interested in how they make this synthetic data—is it just random noise or something much smarter?

Lalam: It's much smarter; they are intentionally generating time series to improve forecasting performance, not just simulating random data.

Tom: That distinction is vital, as the authors emphasize that they aren’t just looking for data realism but better training input.

Jane: The summary highlights that this entire process is a form of task adaptation done at deployment time rather than through massive finetuning.

Lu: And it's all orchestrated through a Metaharness—a system prompt, skills, and tools—which is quite an elegant way to organize the agent's behavior.

Meng: This means that at deployment time, we get a highly tailored forecaster trained specifically for that domain using only those few examples.

Lalam: It’s a powerful marriage of generative AI and specific model training, making the concept feel very complete.

Tom: So, the system is designed to produce a highly specialized lightweight forecaster after all those steps.

Jane: That’s right; by focusing on generating data that specifically supports forecasting, they have created something quite robust.

Lu: I think this structure allows us to achieve performance that rivals much larger models while keeping the computational footprint incredibly small.

Meng: It sounds like a practical solution for edge computing devices where massive models would simply fail to run efficiently.

Lalam: We can hope that the ability to train specialized, lightweight forecasters will significantly reduce the energy consumption of our global AI infrastructure.

Improvements: Tom: We’ve seen how MetaCaster works, but let's dig deeper into the improvements it offers in this paper: MetaCaster: Meta-Harness-Optimized Agent for End-to-End Few-Shot Learning of Lightweight Time Series Forecasters.

Jane: The authors introduce a third agent, the Harness Proposer or HPAGENT, which is perhaps the most novel part of the entire system.

Lu: This HPAGENT is what actually optimizes the "Harness"—that system prompt and skills structure—of MGAGENT during an outer loop of optimization.

Meng: So, it’s not just running a fixed set of instructions; it’s actively improving the AI's internal logic through meta-harness engineering.

Lalam: It feels like the authors are proving that optimizing the *process* is just as important as optimizing the *data*.

Tom: And they do this by using a hinge-loss measure, which penalizes poor performance on generated data to improve it.

Jane: The goal of HPAGENT is to ensure that even if the initial few-shot examples are small, the resulting training data D̄ is better suited for building an effective forecaster.

Lu: This creates a feedback loop: the performance gap between authentic and generated data drives the refinement of the system prompt itself.

Meng: I’m impressed that this optimization is transferrable across different LLMs, which means we can swap out different backbones later on without rebuilding the whole pipeline.

Lalam: It gives us a blueprint for how to improve any agentic workflow, showing that adaptability is a core feature of AI design.

Tom: It’s not just about training the model; it’s about making sure the system generates high-quality training data at all.

Jane: The improvements in efficiency are enormous, too, because the final forecaster is lightweight and requires no further LLM calls after deployment.

Lu: This architecture is a massive step away from simply fine-tuning a large foundation model and is much more sophisticated.

Meng: It’s truly a practical improvement for deploying AI at scale where resource management is paramount.

Conclusion: Tom: We’ve covered the mechanics, the optimizations, and the improvements, but let's wrap up with the overall conclusion of MetaCaster: Meta-Harness-Optimized Agent for End-to-End Few-Shot Learning of Lightweight Time Series Forecasters.

Jane: The authors demonstrate that this system achieves high performance while maintaining both data efficiency and computational efficiency across eighteen different datasets.

Lu: I am really excited about the generalization to out-of-domain or OOD test sets, which shows the robust nature of this approach, proving it’s not just overfitting to one specific dataset.

Meng: The results on Table one show that even when only K=ten samples are available, MetaCaster remains competitive with state-of-the-art methods.

Lalam: This is a great validation of the idea that we can achieve high performance using compact, specialized models in a way that aligns with a more sustainable computational practice.

Tom: It’s clear that the "Agent-as-Engineer" paradigm offers a viable path forward for many complex, data-poor tasks.

Jane: We've seen how it works on real-world scenarios like electricity load diagrams and solar power data, which is very encouraging for applied research.

Lu: This work provides a strong foundation for the next wave of agentic AI by optimizing the infrastructure itself.

Meng: I think this means that we can finally move away from relying on giant, expensive models in many industries where those large-scale deployments aren's necessary.

Lalam: It’s a hopeful look at how AI can help us build better systems for everyone, using MetaCaster: Meta-Harness-Optimized Agent for End-to-End Few-Shot Learning of Lightweight Time Series Forecasters.

Tom: That is the perfect way to wrap up this discussion. Thank you so much to Lu, Meng, and Lalam for sharing your insights with us today on this groundbreaking work!

Conclusion: Tom: So, if I'm getting this right, what we’re seeing with MetaCaster is that they’ve fundamentally made time series forecasting much more accessible, even when you only have a small amount of data.

Jane: Exactly. It really changes the game because historically, you needed massive amounts of historical data to build reliable predictors, but now the model can figure things out with just a few examples.

Meng: And that’s huge for industrial applications right off the bat; many niche systems don't have petabytes of historical readings available, so the efficiency is critical for real-world deployment.

Lu: You nailed it, Meng. It’s not just about having data; it's about learning *how* to learn effectively from minimal examples, which moves us closer to true meta-learning capability in time domains.

Lalam: I think that ability to generalize from scarcity has massive implications for underserved communities that might not have centralized data collection infrastructure yet.

Tom: That’s a powerful point, Lalam. So, if we can deploy this kind of lightweight agent everywhere, it opens up possibilities we barely thought about in predictive analytics.

Jane: It means smaller companies and research groups don't need to spend millions on massive computational clusters just to get a decent forecast for their operations.

Lu: The focus on the 'Meta-Harness' really suggests that the model isn't just learning patterns, it’s learning the optimal *strategies* for pattern discovery itself.

Meng: From an engineering viewpoint, that optimized agent structure means we could potentially run this on edge devices with limited power consumption, which is a major win for embedded systems.

Lalam: Knowing that we can push sophisticated predictive intelligence to the edge changes how we design smart cities and local resource management systems entirely.

Tom: Wow, so instead of just predicting the next hour of electricity usage, we could predict everything from localized traffic flow to micro-climate changes using "MetaCaster: Meta-Harness-Optimized Agent for End-to-End Few-Shot Learning of Lightweight Time Series Forecasters."

Jane: It’s genuinely exciting to think about the sheer breadth of problems that improved time series prediction can solve across so many sectors.

Tom: We’ll definitely keep an eye on how this research translates into commercial products, because the potential here is truly massive.

Lu: I'm really looking forward to seeing how this foundational work influences future architectures for sequential data processing.

Meng: We need to start prototyping implementations with these lightweight models immediately; the practical impact is too big to wait on.

Lalam: The ability of AI to handle data sparsity will elevate global cultural understanding by making complex predictive tools universally available.

Tom: Alright, listeners, that wraps up our deep dive into this incredible paper. We hope you found as much excitement about time series prediction as we did! Next up, we’re shifting gears completely and talking about how AI is transforming genomics...

University of Houston · NEC Labs · University of Waterloo · University of Connecticut · University of Illinois at Urbana-Champaign · Singapore Management University

cs.LG, cs.AI

Submitted: 2026-08-24

Updated: 2026-09-03

Comments: Accepted by EMNLP 2026

Code: https://github.com/D2I-Group/metacaster

License: http://creativecommons.org/licenses/by/4.0/

Importance score: 76/100

The gist: Lightweight time series forecasting (TSF) is critical for resource-constrained environments, yet these specialized models typically require "substantial training data," which severely limits their

Key concepts

MetaCaster
MetaCaster is a multi-agent framework that orchestrates behavior using a 'Metaharness'—a system prompt and tools. This allows the forecasting model to adapt to specific domains at deployment time, avoiding the need for massive finetuning.
Few-Shot Learning
This is a method where the forecaster is trained using only a limited support set of data points. The system intelligently generates high-quality synthetic data from these few examples to build an effective, specialized model.
MGAGENT and FTAGENT
These are the two core agents within MetaCaster. MGAGENT is responsible for generating a sufficient dataset ($ar{D}$), while FTAGENT utilizes that generated data to train and select the best lightweight forecaster.

Terminology

Summary

Lightweight time series forecasting (TSF) is critical for resource-constrained environments, yet these specialized models typically require substantial training data, which severely limits their applicability in domains characterized by scarce, slowly accumulated, or privacy-sensitive time series. To overcome this fundamental dilemma, the paper introduces M ETAC ASTER, a novel multi-agent framework designed to solve the challenging problem of few-shot learning for lightweight forecasters. This system enables the training of highly specialized models using only a small set of examples and textual context, establishing a new paradigm where agents act as intermediary engineers rather than direct forecasters.

How it works

M ETAC ASTER operates as an end-to-end pipeline involving three distinct agents: the Meta-Generator (MG AGENT), the Forecaster Trainer (FTAGENT), and the Harness Proposer (HPAGENT). The process begins when a few-shot support set (D sup) and a contextual description (C) are provided. MG AGENT utilizes this input to generate a sufficient, domain-compliant training dataset, ensuring that the synthetic data is better suited for building effective time series forecasters rather than merely simulating realism. This generated dataset is then partitioned into training and validation sets for the next stage. FTAGENT takes this data and trains a panel of 23 state-of-the-art (SOTA) lightweight forecasters, selecting the best performer based on evaluation metrics like MSE.

The Core Innovation: Harness Optimization

The distinguishing feature of M ETAC ASTER is its integration of a meta-harness optimization loop managed by HPAGENT. Rather than relying on standard data generation techniques, the system employs HPAGENT to optimize the Harness (theta)—the infrastructure surrounding MG AGENT—to minimize a specific forecasting error loss (Eq. 1). This iterative process allows the system to automatically improve the quality of generated data by diagnosing performance degradation and updating skills within the agent's framework. This optimization is crucial because it ensures that the forecaster trained on the generated tr performs comparably to those trained on real data in a few-shot setting.

Experimental Validation and Contributions

The effectiveness of M ETAC ASTER was validated through comprehensive experiments across 18 diverse datasets, comparing its performance against 14 baselines and various time series generation models. The results demonstrate that the framework achieves both data efficiency and computational efficiency while maintaining high-quality TSF performance.

The paper enumerates the following key contributions:

  • Investigating the challenging problem of few-shot learning for lightweight forecasters.

  • Proposing M ETAC ASTER, a novel metaharness-optimized multi-agent framework that achieves large-model-like performance with efficient inference for TSF tasks.

  • Compiling 23 SOTA lightweight forecasters into LT-L IB, a unified library for training.

  • Conducting comprehensive experiments on 18 datasets against 14 baselines, demonstrating the effectiveness of M ETAC ASTER.

Improvements for AI systems

Based on a rigorous analysis of M ETAC ASTER, I have identified several key architectural and methodological advancements that can be generalized to significantly improve various AI systems, particularly in scenarios requiring efficiency, data scarcity handling, and automated task preparation.

Improvement: Transition from data synthesis focused on realism (e.g., generating statistically plausible noise or cycles) to synthesis focused on performance optimization.

What the Improved System Can Do:

  • Automated Performance Alignment: An automated agent (Meta-Generator) can take a small set of real examples (D sup) and generate synthetic data that is specifically tailored to maximize the performance of a target, resource-constrained model (e.g, achieving minimal Mean Squared Error on the generated set).

  • Deployment in Data Scarce Environments: This allows specialized models to be trained successfully in domains where collecting large amounts of historical data is impossible due to privacy constraints or slow accumulation (e.g., niche medical records, specific industrial failure logs).

The improved AI system moves away from the one-size-fits-all foundation model paradigm. It becomes an Agentic Data Engineer that automatically:

  1. Optimizes data generation for performance, not just realism.

  2. Refines its own operational logic (Harness) to achieve superior task performance iteratively.

  3. Select the most efficient specialized tool from a library to execute the final inference, ensuring high accuracy with minimal computational overhead (10 cubed to 10 5 times faster than TSFMs).

Abstract

Time series forecasting (TSF) is evolving toward multimodal and agentic settings, yet using foundation models remains uneconomical in resource-constrained scenarios, where compact, specialized forecasters are more desirable. However, lightweight forecasters typically require substantial training data, limiting their use in domains with scarce, slowly accumulated, or privacy-sensitive time series. To address this dilemma, we investigate the challenging problem of few-shot learning for lightweight forecasters. We propose MetaCaster, a meta-harness-optimized multi-agent framework that uses agentic data generation to automatically train specialized lightweight forecasters from only a few examples and textual contexts. Our work highlights a new TSF paradigm in which agents act not as forecasters but as intermediary engineers that prepare efficient, task-specific forecasters for deployment. Experiments on 18 datasets, 23 state-of-the-art lightweight forecasters, and 14 baselines demonstrate that MetaCaster achieves both data efficiency and computational efficiency while maintaining high-quality TSF performance.

Sources

Related papers