Multimodal decision models show greater stability under presentation changes
MM-SystemOneBias: Evaluating Bias in Multimodal Jev-Style System One Decision Models
Artificial Intelligence
Summary
Multimodal decision models use different types of input, like images and text, to make decisions. The authors found that these models tend to keep their answers more stable when irrelevant changes are made, such as swapping option orders or adding unrelated information. They tested 13 models across nine datasets and various task scenarios, showing that these models handle tricky situations better than their simpler versions, though some input changes still affect their confidence levels in subtle ways.
What this means in practice
- •For software engineers: Build systems that rely on multimodal decision models with insights on how to reduce bias from input presentation changes.
- •For data annotation teams: Design evaluation protocols for labeling tasks to assess consistency of model decisions under diverse input conditions.
Authors
Jinzhe Li, Qi Wang, Dandan Guo, Li Liu, Yuan Wu, Yi Chang
Abstract
When multimodal decision models determine actions, moderation outcomes, or quality judgments, their reliability depends on preserving semantic decisions across presentation changes, ignoring irrelevant additions, and distinguishing information consistent with task evidence from information that conflicts with it. Existing evaluations study decision-interface biases and multimodal model behavior in separate settings, leaving no common protocol to compare multimodal decision models with their backbones on practical tasks. We introduce \textbf{MM-SystemOneBias}, which evaluates 13 models on 5,323 original instances from nine public datasets and 11 task scenarios using six paired interventions: option order, option label, external judgment, image--text description, irrelevant information, and rating scale. Across seven matched decision-model--backbone pairs, decision models are more stable under option-label, irrelevant-information, and rating-scale changes in all seven pairs; six pairs also show lower option-order flip rates. They incur smaller accuracy losses under incorrect external judgments in all seven pairs and conflicting descriptions in six. In six pairs, they gain less accuracy from correct external judgments and show less favorable accuracy changes under consistent descriptions. Among pairs with unchanged semantic decisions, mean absolute changes in answer probabilities range from 4.56 to 11.55 percentage points, revealing input effects that accuracy alone misses.