AgentLTV: An Agent-Based Unified Search-and-Evolution Framework for Automated Lifetime Value Prediction

A research team including Chaowei Wu, Mengzhuo Guo, and seven other authors has introduced AgentLTV, a novel automated framework for building Lifetime Value (LTV) prediction models. LTV prediction is crucial for advertising, e-commerce, and recommendation systems, but traditionally requires expensive, scenario-specific pipelines that are difficult to transfer between projects. AgentLTV addresses this by treating each candidate solution as an executable pipeline program, with LLM-driven agents that autonomously generate code, run pipelines, analyze feedback, and make repairs. This represents a significant shift toward automating one of the most labor-intensive aspects of applied machine learning.

The framework employs a two-stage search strategy: first, a Monte Carlo Tree Search (MCTS) stage explores a wide space of modeling choices under a fixed budget, guided by a Pareto-aware multi-metric value function. Second, an Evolutionary Algorithm (EA) stage refines the best programs found by MCTS using island-based evolution with crossover, mutation, and migration. Experiments on a large-scale proprietary dataset and a public benchmark show AgentLTV consistently discovers strong models across both ranking and error metrics. The system has already been deployed in production, where it demonstrated improved ranking consistency and value calibration, particularly for high-value and negative-LTV customer segments—areas where manual models often struggle. The researchers submitted their findings to KDD 2026, providing practitioners with clear takeaways: use MCTS for rapid adaptation to new data patterns and EA for stable refinement.