Self-Directed Discovery: How LLMs Explore and Optimize Configurable LLM Solutions

ICLR 2026 Conference Submission19513 Authors

Keywords: meta-learning, self-improving systems, LLM agents, automated prompt engineering, meta-reasoning, iterative improvement

TL;DR: We show that large language models can systematically optimize other LLM-based workflows through a configuration-driven framework enabling improved prompt templates, input processing logic, and reasoning strategies.

Abstract: LLM solutions are increasingly replacing specialized machine learning models across various industry domains. While offering simplicity and maintainability advantages, optimizing these workflows remains heavily dependent on expert-driven experimentation. In this paper we test off-the-shelf powerful LLMs for the task of automatically building and iteratively improving LLM-based solutions. We propose a configuration-driven framework which defines such workflows by specifying model parameters, prompt templates, and data transformations. We show that this standardized representation enables automatic iterative improvement loops via optimization agents which are themselves defined within the same framework. When evaluated on challenging datasets, it discovers improved solutions while maintaining interpretability and human verifiability. The improvement loop we instantiate is generic and minimal (using prompts that have not been engineered) and self-referential (improved solutions are discovered with improved documentation and examples). The proposal is a self-improving system that bridges the gap between generic code-generating automatic optimization and more narrowly-focused techniques such as prompt engineering.

Primary Area: foundation or frontier models, including LLMs

Submission Number: 19513