EVIL: Evolving Interpretable Algorithms for Zero-Shot Inference on Event Sequences and Time Series with LLMs

Researchers introduce EVIL, an LLM-guided evolutionary approach that discovers interpretable Python algorithms for zero-shot inference on time series and event sequences without traditional neural network training. The evolved algorithms match or exceed deep learning performance while remaining transparent and significantly faster, demonstrating a novel paradigm for dynamical systems inference.

EVIL represents a meaningful shift in how machine learning approaches complex inference problems. Rather than relying on black-box neural networks trained on massive datasets, the method uses large language models to guide evolutionary search toward discovering compact, human-readable algorithms. This approach tackles three challenging domains simultaneously—temporal point process prediction, Markov jump process estimation, and time series imputation—with a single evolved algorithm that generalizes across datasets without per-dataset fine-tuning.

The broader context reveals growing frustration with deep learning's opacity and computational inefficiency. As organizations increasingly face regulatory pressure around model explainability and struggle with deployment costs, methods that deliver comparable performance with interpretability gain traction. EVIL's zero-shot generalization capability is particularly valuable; it sidesteps the expensive retraining cycle that typically accompanies new data or domains.

For the AI industry, this work challenges assumptions about necessary model complexity. If interpretable algorithms can match neural network performance while running orders of magnitude faster, organizations may reconsider the ROI of deep learning pipelines. The approach is especially relevant for financial time series and event prediction, where both interpretability and speed directly impact decision-making and regulatory compliance.

The long-term significance depends on whether EVIL generalizes beyond the three tested domains. If the method scales to higher-dimensional problems and more complex dynamical systems, it could reshape how inference models are developed and deployed, particularly in resource-constrained or regulation-heavy environments. The next critical test involves real-world deployment scenarios where transparency requirements and computational constraints strongly favor evolved algorithms over traditional approaches.

• →EVIL discovers interpretable algorithms through LLM-guided evolution, achieving competitive or superior performance to deep learning without dataset-specific training
• →The method produces orders-of-magnitude speed improvements while maintaining full interpretability, addressing key deployment and regulatory concerns
• →A single evolved algorithm generalizes across multiple datasets and domains (temporal processes, Markov chains, time series imputation) without retraining
• →Results suggest neural networks may be unnecessarily complex for certain dynamical systems inference tasks, potentially reshaping model selection practices
• →The approach is particularly relevant for financial and regulated industries where both speed and explainability are critical requirements