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

TL;DR

This work addresses the mismatch between correct predictions and plausible decision evidence by introducing attribution-guided, prior-constrained training that leverages highly faithful subset-based attribution. Human priors are encoded as input regions, and the model is penalized when its top attribution evidence lies outside these regions, via deviation and redundancy losses. The approach, instantiated with LIMA for image tasks and EAGLE for MLLM-based GUI agents, yields consistent gains in both accuracy and attribution reasonability across image classification and GUI decision tasks, including robustness to noise. The findings suggest that constraining decision evidence to human-recognized regions improves reliability and interpretability without sacrificing performance, with potential impact in safety-critical, data-constrained domains.

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.

Figures (7)

• Figure 1: Correct outputs do not guarantee reasonable decision evidence: both a classifier and an MLLM-based GUI agent can succeed while their attribution maps violate human priors. We propose attribution-guided, prior-constrained alignment training to push evidence toward human-prior regions and improve causal reasonableness and decision reliability.
• Figure 2: Qualitative comparison on Saliency-Bench and ImageNet-S. For each method, we visualize LIMA-based attributions on the same inputs; white masks indicate human priors (target object regions).
• Figure 3: Impact of training hyperparameters on model performance. A. Effect of the attribution interval on validation accuracy. B. Effect of the loss balancing coefficient $\lambda_1$ on validation accuracy.
• Figure 4: Qualitative GUI agent results comparing SFT with LoRA adaptation and the proposed our method, together with attribution heatmaps, predicted click locations (stars), and human-prior target bounding boxes.
• Figure 5: GUI clicking example comparing SFT with LoRA adaptation and our method, showing attribution maps and correct clicks on the target checkmark.

Theorems & Definitions (1)

• Definition 3.1: Subset-Selection-Based Attribution