Recon: Reconstruction-Guided Reasoning Synthesis for User Modeling

Researchers introduce Recon, a method for improving user modeling by evaluating synthesized reasoning traces through action reconstruction rather than post-hoc rationalization. The approach achieves 54.7% win rates over baseline methods and demonstrates that reasoning should naturally elicit predicted actions from context, advancing AI's ability to simulate human behavior.

Recon addresses a fundamental limitation in current user modeling approaches: the gap between justification and causation. Traditional methods generate reasoning traces after observing actions, creating plausible-sounding explanations that may not reflect actual decision-making processes. This distinction matters because downstream applications—whether in behavioral science, market research, or human-AI collaboration—require models that capture authentic decision patterns rather than post-hoc rationalizations.

The research builds on growing recognition that language models struggle with causal reasoning. By inverting the problem through reconstruction-guided synthesis, Recon scores reasoning quality based on predictive power: can a reasoning trace, combined with context, reliably predict the original action? This framework transforms reasoning synthesis from a generation-then-rationalize problem into a prediction-then-validate problem, fundamentally reorienting how AI systems approach user behavior modeling.

The empirical results demonstrate substantial improvements—up to 70% win rates when using Recon-derived rewards for training synthesis models. Critically, the synthesized reasoning transfers across different models and improves performance beyond the reconstruction model itself, suggesting the method captures generalizable patterns rather than model-specific artifacts. This transferability indicates the approach identifies authentic causal pathways in user behavior.

For AI development, this work signals the limitations of purely generative approaches to reasoning and the importance of grounding synthetic explanations in predictive validation. The methodology could influence how future language models learn to explain their outputs and how researchers evaluate the faithfulness of AI-generated reasoning. The cross-domain validation across four different domains strengthens confidence in the approach's robustness and broader applicability to user simulation problems.

• →Recon uses action reconstruction to evaluate reasoning quality, achieving 54.7% improvement over post-hoc rationalization baselines.
• →The method demonstrates that authentic reasoning should naturally elicit predicted actions from context rather than justify them retroactively.
• →Recon-synthesized reasoning transfers across different models, improving user modeling performance beyond the reconstruction model.
• →The approach addresses a fundamental limitation in current user modeling: distinguishing between causal decision paths and plausible-sounding justifications.
• →Results span four domains, indicating the methodology's robustness for diverse user behavior simulation applications.