Beyond Action Imitation: Learning a Decision-Aware User Simulator for Online Advertising

2026-07-29

AI Summary

Existing user simulators rely solely on single-domain interaction histories to model observable behaviors—such as clicks—which inadequately capture user preferences, often leading to model shortcuts that compromise simulation fidelity and diagnostic utility. To address this limitation, this work proposes DASH, a decision-aware user simulator that introduces, for the first time, a decision-perception mechanism integrating heterogeneous multi-domain historical data. DASH explicitly models users’ reasoning processes via chain-of-thought distillation and jointly predicts behavioral actions. The approach employs context engineering to fuse cross-domain information, leverages large language models to generate reasoning trajectories as supervision signals, and incorporates a multidimensional reward model—encompassing form, content, and logic—for reinforcement learning. Experiments on five heterogeneous advertising datasets from Tencent demonstrate that DASH significantly outperforms existing methods in effectiveness, efficiency, simulation fidelity, and diagnostic capability.

Abstract

Recent advances in LLM-based user simulation have shown promise for offline evaluation of recommendation and advertising systems. However, existing simulators typically infer user preferences from single-domain interaction histories and are primarily optimized to reproduce observable actions such as clicks. Consequently, they capture only a partial view of user preferences, while action-only prediction easily induces model shortcuts and limits both the fidelity and diagnostic value of simulation. To address these challenges, we propose DASH, a decision-aware user simulator that jointly generates thinking traces and predicts behavioral actions from heterogeneous cross-domain histories. DASH first introduces a Context Engineering stage that folds heterogeneous cross-domain histories into decision-relevant context, together with prompt optimization for effective reasoning over the folded context. To train a user simulator, DASH distills thinking trajectories from strong LLMs as SFT data, and further tailors a rubric-based reward model that evaluates thinking traces along form, content, and logic for RL training. Combined with the action reward, these signals jointly improve action prediction and thinking quality. Extensive experiments on real-world Tencent advertising data spanning five heterogeneous content domains demonstrate the effectiveness, efficiency, fidelity, and diagnostic value of DASH.

Problem

• user simulation
• online advertising
• action imitation
• cross-domain history
• decision-aware modeling

Innovation

• decision-aware simulation
• thinking traces
• cross-domain user modeling
• context engineering
• LLM distillation