Recon: A New Method for User Modeling via Reconstruction-Guided Reasoning Synthesis
A research paper available on arXiv (2605.26969) presents Recon, an innovative technique for user modeling that employs action reconstruction to evaluate reasoning traces. The goal of user modeling is to replicate individual behaviors based on historical context-action pairs through language models, applicable in fields such as behavioral science, human-AI collaboration, and market research. Current methods create reasoning traces by relying on both context and action, which leads to post-hoc rationalization rather than genuine reasoning, as these traces merely justify actions without reflecting the true causal decision-making process. In contrast, Recon utilizes a reconstruction model to forecast actions based on context and potential reasoning, with the quality of reasoning hinging on reconstruction accuracy. In four domains, Recon boasts a 54.7% success rate over Backward Synthesis, a conventional post-hoc approach.
Key facts
- Recon uses action reconstruction to score reasoning traces for user modeling.
- Existing methods generate reasoning traces by conditioning on both context and action, leading to post-hoc rationalization.
- Recon's reconstruction model predicts the action given context and candidate reasoning.
- Recon achieves a 54.7% win rate over Backward Synthesis across four domains.
- User modeling has applications in behavioral science, human-AI collaboration, and market research.
- The paper is available on arXiv with ID 2605.26969.
- The approach aims to encode underlying latent causal decision paths.
- Recon stands for Reconstruction-Guided Reasoning Synthesis.
Entities
Institutions
- arXiv