When Model Priors Conflict with Visual Evidence: Mitigating Commonsense-Driven Hallucinations by Selective Prior Calibration

2026-07-31

AI Summary

This work addresses the tendency of vision-language models to generate hallucinations when confronted with counterfactual images that contradict commonsense priors despite clear visual evidence. To mitigate this issue, the authors propose Selective Prior Calibration (SPC), a method that estimates the commonsense prior preference for candidate answers and dynamically subtracts this bias from image-conditioned scores with instance-dependent strength. Prediction is adjusted only when the calibrated score strongly favors an unconventional answer. This approach innovatively enables selective suppression of commonsense priors, correcting errors on counterfactual inputs while preserving their beneficial role in conventional scenarios. Experimental results demonstrate that SPC significantly improves accuracy on counterfactual images across multiple commonsense conflict categories and benchmarks, without degrading performance on images consistent with commonsense expectations.

Abstract

In vision--language models, commonsense-driven hallucination (CDH) occurs when a model's commonsense prior overrides clear visual evidence of an atypical state. For example, a model may report that a visibly six-fingered hand has five fingers. We show that these errors are systematically directed: when a model answers a question about a counterfactual (CF) image incorrectly, its answer often coincides with the candidate it prefers without access to the image. Suppressing this prior indiscriminately can repair CF errors, but may also disrupt correct answers on matched commonsense (CS) images, where the same prior is helpful. We therefore propose Selective Prior Calibration (SPC), which subtracts candidate-level prior-preference estimates from image-conditioned scores with an instance-dependent strength and revises the original prediction only when the resulting score pattern strongly supports an alternative. Extensive experiments demonstrate that SPC substantially improves accuracy on CF images while largely preserving accuracy on matched CS images. Furthermore, these gains generalize across CDH categories, candidate-answer permutations, and other conflict benchmarks, while SPC rarely alters predictions on benchmarks without such conflicts.

Problem

• commonsense-driven hallucination
• vision-language models
• model priors
• visual evidence
• counterfactual images

Innovation

• commonsense-driven hallucination
• prior calibration
• vision-language models
• counterfactual reasoning
• selective intervention