AlphaOPT: Formulating Optimization Programs with Self-Improving LLM Experience Library

Researchers introduce AlphaOPT, an AI system that automatically learns to translate complex optimization problems into executable code through a self-improving experience library. The method achieves 72% accuracy on optimization benchmarks and outperforms existing LLM approaches by 8-9% without requiring model retraining or gold-standard annotations.

AlphaOPT addresses a fundamental challenge in AI automation: converting natural language problem descriptions into precise mathematical formulations and executable solver code. Traditional approaches either rely on fragile prompt engineering or computationally expensive model retraining, both limiting practical deployment at scale. The research demonstrates that structured experience accumulation—rather than parameter updates—enables LLMs to progressively improve on constrained reasoning tasks.

The system operates through a two-phase learning cycle that mirrors human expertise development. During the Library Learning phase, failed problem-solving attempts are analyzed for insights that solver verification confirms as correct. The Library Evolution phase then refines which insights apply to new problems by tracking aggregate evidence across tasks. This design maintains bounded memory growth while capturing reusable optimization principles. The approach bypasses the need for annotated training data or gold-standard reference programs, making it practical for domains where expert solutions are scarce or expensive to obtain.

For the AI industry, AlphaOPT represents progress toward autonomous reasoning systems that improve through feedback loops rather than expensive retraining cycles. The 65-72% accuracy improvement across 100-300 training items suggests meaningful scaling potential, while 8.2-9.1% out-of-distribution gains indicate genuine generalization rather than memorization. This matters for enterprises deploying optimization across supply chains, resource allocation, and financial modeling—areas where custom model fine-tuning remains cost-prohibitive.

Looking forward, similar self-improving architectures could extend beyond optimization to other structured reasoning domains. The open-source release enables broader experimentation and signals potential integration into commercial AI platforms seeking to reduce retraining overhead while maintaining performance gains.

• →AlphaOPT learns optimization modeling through structured experience reuse rather than parameter retraining, reducing computational costs.
• →The system achieves 72% accuracy with 300 training items and beats baselines by 8-9% on out-of-distribution datasets.
• →Answer-only feedback and solver verification enable learning without gold-standard annotations or costly manual labeling.
• →Self-improving experience libraries represent a scalable alternative to continuous model fine-tuning for complex reasoning tasks.
• →Open-source availability at GitHub signals potential for enterprise adoption in supply chain and financial optimization applications.