Where Not to Learn: Prior-Aligned Training with Subset-based Attribution Constraints for Reliable Decision-Making

2026-01-30

arXiv.org

AI Summary

This work addresses the tendency of models to rely on spurious correlations rather than valid evidence, which often yields high accuracy but unreliable reasoning. To mitigate this, the authors propose an attribution-alignment training framework that encodes human priors—such as preferred attention regions—as explicit constraints during learning. By integrating a high-fidelity subset selection attribution method, the approach continuously monitors and penalizes model decisions grounded outside these prior-specified regions. This mechanism enables the first explicit guidance of model rationales in both image classification and MLLM-driven GUI agent click prediction tasks. The method not only improves task accuracy but also substantially enhances the faithfulness and plausibility of the underlying reasoning process.

Abstract

Reliable models should not only predict correctly, but also justify decisions with acceptable evidence. Yet conventional supervised learning typically provides only class-level labels, allowing models to achieve high accuracy through shortcut correlations rather than the intended evidence. Human priors can help constrain such behavior, but aligning models to these priors remains challenging because learned representations often diverge from human perception. To address this challenge, we propose an attribution-based human prior alignment method. We encode human priors as input regions that the model is expected to rely on (e.g., bounding boxes), and leverage a highly faithful subset-selection-based attribution approach to expose the model's decision evidence during training. When the attribution region deviates substantially from the prior regions, we penalize reliance on off-prior evidence, encouraging the model to shift its attribution toward the intended regions. This is achieved through a training objective that imposes attribution constraints induced by the human prior. We validate our method on both image classification and click decision tasks in MLLM-based GUI agent models. Across conventional classification and autoregressive generation settings, human prior alignment consistently improves task accuracy while also enhancing the model's decision reasonability.

Problem

• human prior alignment
• attribution constraints
• reliable decision-making
• shortcut correlations
• decision evidence

Innovation

• attribution-based alignment
• human priors
• subset selection
• reliable decision-making
• evidence grounding