Evolutionary Multi-Task Optimization for LLM-Guided Program Discovery

2026-05-21 · Source: Machine Learning · Field: Technology & Digital — Artificial Intelligence & Machine Learning · Depth: Expert, quick

Summary

Evolutionary Multi-Task Optimization (EMO) for LLM-guided program discovery introduces EMO-STA, a two-stage framework designed to enhance program discovery by leveraging shared structures across related tasks. Unlike traditional methods that optimize each task independently, EMO-STA first evolves a shared archive of executable programs across a task family. Subsequently, it adapts selected shared candidates to individual target tasks. The framework explores multiple adaptation strategies, including warm-starting and adapting the best-performing shared programs. Across eight task families, encompassing continuous optimization, geometric construction, modeling, and algorithmic optimization, EMO-STA consistently outperforms matched-compute single-task evolution. Specifically, STA Best-Local provides the strongest in-distribution adaptation, while STA Best-Shared offers robust transfer to unseen tasks. Compute allocation experiments indicate that dedicating a substantial, often balanced, budget to shared evolution is beneficial. Beyond efficiency, shared evolution mitigates overfitting in low-evidence scenarios, such as ARC tasks and time-series feature engineering, by promoting programs that generalize rather than exploit brittle task-specific artifacts.

Key takeaway

For Machine Learning Engineers developing LLM-guided program discovery systems, consider implementing Evolutionary Multi-Task Optimization (EMO-STA). This framework can significantly improve algorithm discovery efficiency and generalization compared to single-task approaches. By allocating a substantial, often balanced, budget to shared evolution, you can mitigate overfitting in low-evidence scenarios and achieve robust transfer to unseen tasks. Prioritize STA Best-Local for in-distribution adaptation or STA Best-Shared for broader applicability.

Key insights

EMO-STA improves LLM-guided program discovery by evolving shared program archives across task families before adapting them to specific tasks.

Principles

• Related tasks share reusable program structure.
• Shared evolution mitigates overfitting in low-evidence settings.
• Balanced shared and adaptation budgets are often optimal.

Method

EMO-STA employs a two-stage process: first, evolve a shared archive of executable programs across a task family; then, adapt selected candidates to each target task using strategies like warm-starting.

In practice

• Apply EMO-STA to continuous optimization problems.
• Use STA Best-Shared for robust transfer to unseen tasks.
• Mitigate overfitting in few-shot learning with shared evolution.

Topics

• Evolutionary Multi-Task Optimization
• LLM-Guided Program Discovery
• EMO-STA Framework
• Program Synthesis
• Overfitting Mitigation
• Low-Evidence Learning

Best for: Research Scientist, AI Scientist, Machine Learning Engineer

Related on AIssential

• Towards Automated Knowledge Transfer in Evolutionary Multitasking via Large Language Models — LLMs can autonomously design efficient and effective knowledge transfer models for multi-task optimization.
• Unleashing LLMs in Bayesian Optimization: Preference-Guided Framework for Scientific Discovery — LGBO integrates LLM semantic reasoning into every BO iteration via a preference mechanism, accelerating scientific discovery.
• Controlled Self-Evolution for Algorithmic Code Optimization — Controlled Self-Evolution significantly boosts code optimization efficiency by guiding LLM-based iterative refinement.
• MetaEvo: A Meta-Optimization Framework for Experience-Driven Agent Evolution — MetaEvo enables LLM agents to continually improve reasoning by optimizing how they learn principles from experience, not just what they store.
• Efficient Hyperparameter Optimization for LLM Reinforcement Learning — JF-HPO efficiently optimizes LLM RL hyperparameters by jointly adapting model size and training budget, using proxy models, early stopping, and checkpointing.
• EvoSelect: Data-Efficient LLM Evolution for Targeted Task Adaptation — EvoSelect efficiently adapts LLMs by selecting diverse, task-aligned synthetic data using optimal transport and diversity regularization.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Machine Learning.