Evolutionary Multi-Task Optimization for LLM-Guided Program Discovery
Proposes EMO-STA, a multi-task evolutionary framework leveraging shared program archives, improving efficiency and generalization in LLM-guided program discovery.
Key Findings
Methodology
This paper introduces EMO-STA (Shared-Then-Adapt), a framework combining evolutionary search with LLM-generated executable programs. It first conducts shared evolution across a task family to optimize a reusable program archive, using a group-average objective to capture common structures. Then, for each target task, it applies three adaptation strategies—Warmstart, Best-Shared, and Best-Local—to initialize local evolution from the shared archive, refining task-specific programs. Experiments across eight diverse task families demonstrate superior performance over isolated single-task evolution, especially in low-data scenarios, with notable improvements in solution quality and transfer robustness.
Key Results
- Across most task families, EMO-STA significantly outperforms single-task baselines, with performance gains of 5-15%. For example, in Heilbronn triangle tasks, scores increased from 0.547 to 0.650 (Haiku-4.5), and in function minimization, average improvements of 0.058 points were observed. Shared evolution strategies showed enhanced generalization, particularly in low-sample settings like ARC and time-series forecasting, reducing overfitting.
- Different initialization strategies yielded varied benefits: Best-Local was most stable within tasks, while Best-Shared excelled in transfer to unseen tasks. Resource allocation experiments indicated that a balanced split (e.g., 60/15/120) between shared and local evolution maximized performance, confirming the importance of shared structure in efficient search.
- Overall, EMO-STA effectively captures shared task structures, improving program transferability and robustness, especially under limited data conditions, demonstrating its potential in scientific discovery and automated algorithm design.
Significance
This work addresses the inefficiency and limited generalization of traditional single-task evolutionary methods by introducing a multi-task shared-then-adapt framework. It leverages the structural similarities among related tasks, enabling more resource-efficient search and better transferability of discovered programs. Such advancements are crucial for scientific research, automated algorithm synthesis, and industrial automation, especially when data is scarce or tasks are complex. The approach offers a new paradigm for multi-task learning in program synthesis, with broad implications for AI-driven scientific discovery.
Technical Contribution
The core innovation lies in integrating a shared program archive with multi-strategy local adaptation, facilitated by LLM-generated executable programs. This approach moves beyond fixed representation spaces, allowing flexible, scalable multi-task optimization. The framework's design supports diverse task families, combining shared evolution with task-specific refinement, backed by experimental validation showing improved solution quality, robustness, and transferability. It also introduces resource allocation strategies that balance shared and local compute, optimizing overall efficiency.
Novelty
This is the first systematic implementation of a shared-then-adapt multi-task evolutionary framework utilizing LLM-generated programs. Unlike prior work limited to fixed encoding spaces, EMO-STA exploits shared program structures across tasks, with multiple initialization strategies enhancing transfer and generalization. Its ability to improve performance in low-data regimes and mitigate overfitting distinguishes it as a significant step forward in automated program synthesis and multi-task learning.
Limitations
- The framework relies on designing compatible shared interfaces; tasks with highly heterogeneous structures or incompatible interfaces may limit effectiveness, requiring further interface adaptation techniques.
- Scaling to very high-dimensional or complex tasks could lead to increased search space and computational costs, necessitating more efficient search heuristics.
- Dependence on LLM quality means that poor program generation could impair overall performance; integrating validation or verification mechanisms remains an open challenge.
Future Work
Future directions include automating interface design for broader task applicability, integrating reinforcement or meta-learning to improve adaptation efficiency, and extending the framework to multi-modal inputs and more complex real-world scenarios. Exploring scalability and real-time applications will also be key to deploying EMO-STA in industrial settings and scientific research.
AI Executive Summary
The advent of large language models (LLMs) has revolutionized program synthesis, enabling automated discovery of complex algorithms. However, traditional evolutionary algorithms often operate in isolated, single-task settings, leading to inefficiencies and limited transferability. This paper addresses these challenges by proposing EMO-STA, a novel multi-task evolutionary framework that leverages shared program structures across related tasks.
EMO-STA begins with a shared evolution phase, where a population of candidate programs is evolved against an aggregated objective across a task family. This process produces a repository of reusable programs that encapsulate common structures and strategies. Subsequently, for each target task, the framework employs three adaptation strategies—Warmstart, Best-Shared, and Best-Local—to initialize local evolution from the shared archive, refining programs to task-specific requirements.
Experimental results across eight diverse task families—including continuous optimization, geometric construction, and algorithmic tasks—demonstrate that EMO-STA consistently outperforms traditional single-task evolution under equivalent compute budgets. Notably, in low-data scenarios such as ARC tasks and time-series forecasting, shared evolution mitigates overfitting, leading to more robust solutions. The resource allocation studies reveal that balancing shared and local compute (e.g., 60/15/120) maximizes performance gains.
This approach introduces a significant advancement in automated program discovery, combining structural sharing with targeted adaptation. It offers a scalable, generalizable solution for complex multi-task environments, with promising implications for scientific discovery, automated algorithm design, and industrial automation. Future work will focus on interface automation, scalability, and integration with reinforcement learning to further enhance its capabilities.
Deep Dive
Abstract
Recent LLM-guided evolutionary search methods have shown that iterative program mutation can discover strong algorithms, but they typically optimize each task independently, even when related tasks share reusable structure. We introduce Evolutionary Multi-Task Optimization (EMO) for LLM-guided program discovery, and propose EMO-STA (Shared-Then-Adapt), a two-stage framework that first evolves a shared archive of executable programs across a task family and then adapts selected shared candidates to each target task. Within EMO-STA, we explore multiple adaptation strategies, including warm-starting from the shared archive, adapting the best average shared program, and adapting the shared program that performs best on each target task. Across eight task families spanning continuous optimization, geometric construction, modeling, and algorithmic optimization, EMO-STA improves over matched-compute single-task evolution in most settings, with STA Best-Local providing the strongest in-distribution adaptation and STA Best-Shared yielding robust transfer to unseen tasks. Compute-allocation experiments show that allocating a substantial fraction of the family-level budget to shared evolution is consistently beneficial, with roughly balanced shared and adaptation budgets often being optimal. Beyond compute efficiency, we show that shared evolution can mitigate overfitting in low-evidence settings (e.g. few training data), including ARC tasks and time-series feature engineering, by favoring programs that generalize across all tasks rather than exploiting task-specific brittle artifacts.