Evolutionary Multi-Task Optimization for LLM-Guided Program Discovery
For researchers in program discovery and evolutionary computation, this work provides a principled method to leverage shared structure across related tasks, improving efficiency and generalization.
The paper introduces EMO-STA, a two-stage framework for LLM-guided evolutionary multi-task optimization that first evolves a shared archive of programs across a task family and then adapts them to target tasks. Across eight task families, it improves over single-task evolution in most settings, with compute-allocation experiments showing that balanced shared and adaptation budgets are often optimal, and shared evolution mitigates overfitting in low-evidence settings.
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.