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

2025-08-07 · Source: cs.AI updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Data Science & Analytics · Depth: Expert, extended

Summary

EVIL (EVolving Interpretable algorithms with LLMs) is a novel approach that employs LLM-guided evolutionary search to discover simple, interpretable Python/NumPy algorithms for dynamical systems inference. Unlike deep learning methods that require extensive training on large datasets, EVIL evolves pure Python programs capable of zero-shot, in-context inference across diverse datasets. The system was applied to three distinct tasks: next-event prediction in temporal point processes, rate matrix estimation for Markov jump processes, and time series imputation. In each case, a single evolved algorithm generalized across all evaluation datasets without per-dataset training, often achieving competitive or superior performance compared to state-of-the-art deep learning models. These evolved algorithms are orders of magnitude faster, fully interpretable, and significantly cheaper to discover, with evolution taking minutes and costing less than $1 in API fees.

Key takeaway

For AI Scientists and Machine Learning Engineers developing solutions for dynamical systems, EVIL offers a compelling alternative to complex deep learning models. You should consider applying LLM-guided evolutionary search to discover simple, interpretable algorithms, especially when zero-shot generalization, rapid inference, and full transparency are critical. This approach can significantly reduce development costs and time, while often achieving competitive performance, providing a grounded baseline to justify the complexity of more elaborate models.

Key insights

LLM-guided evolution can discover compact, interpretable algorithms for zero-shot dynamical systems inference, outperforming complex neural networks.

Principles

• Simpler algorithms can generalize better across diverse datasets.
• LLMs provide informed starting points for algorithm design.
• Direct statistical computations can surpass learned approximations.

Method

EVIL uses an LLM-guided evolutionary search (AlphaEvolve-style) to iteratively refine Python/NumPy functions. The LLM proposes code modifications, which are scored against a fitness function, and promising solutions are stored in an evolutionary database.

In practice

• Use EVIL for rapid, cost-effective algorithm discovery.
• Incorporate domain expert knowledge via system prompts.
• Prioritize interpretability for safety-critical applications.

Topics

• LLM-guided Program Evolution
• Zero-Shot Inference
• Dynamical Systems
• Temporal Point Processes
• Markov Jump Processes

Code references

• algorithmicsuperintelligence/openevolve

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

Related on AIssential

• EVIL: Evolving Interpretable Algorithms for Zero-Shot Inference on Event Sequences and Time Series with LLMs — LLM-guided evolutionary search can discover interpretable, zero-shot algorithms for dynamical systems inference.
• Data Machina #261 — Generative AI and LLMs are being adapted to achieve state-of-the-art performance in diverse time-series tasks.
• LLM-Guided Evolution for Medical Decision Pipelines — LLM-guided evolution optimizes medical decision pipelines at inference-time, outperforming manual baselines through interpretable program-level changes.
• IEEE Transactions on Artificial Intelligence, Volume 7, Issue 2, February 2026 — The volume showcases diverse AI advancements, from LLM applications in autonomous driving to specialized models for protein language processing.
• Mutation Without Variation: Convergence Dynamics in LLM-Driven Program Evolution — LLM-driven program mutation inherently converges to limited structural forms, hindering open-ended exploration.
• GUIDE: Guided Updates for In-context Decision Evolution in LLM-Driven Spacecraft Operations — GUIDE enables LLMs to adapt in real-time by evolving natural-language decision rules without weight updates.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.AI updates on arXiv.org.