Consistency as Inductive Bias: Learning Cross-View Invariance for Robust Multimodal Reasoning

2026-06-29

AI Summary

This work addresses a critical limitation in current large multimodal language models—their lack of cross-view consistency as an inductive bias—rendering them prone to inconsistent or erroneous reasoning under semantically invariant input transformations. The authors formalize cross-view consistency as an online credit assignment signal within reinforcement learning and introduce ConsistRoll, a method that incurs no additional generation overhead or annotations. By constructing original–transformed view pairs via semantic-preserving data augmentation, ConsistRoll operates within a Group Relative Policy Optimization (GRPO) framework, awarding joint rewards only when outputs from both views are correct and consistent. This introduces a cross-view correction term absent in standard data augmentation, mitigating advantage collapse. Experiments demonstrate that ConsistRoll significantly enhances model robustness and consistency across multiple benchmarks, including mathematical reasoning, general-purpose tasks, and hallucination evaluation.

Abstract

Inductive biases steer learning toward generalizable solutions by encoding task structure. In this work, we identify a crucial missing bias in MLLMs: cross-view consistency, \textit{i.e.}, semantically invariant views of the same instance should lead to the same answer. Standard reinforcement learning with verifiable rewards (RLVR) objectives do not impose this constraint, but instead assign pointwise rewards to each visual input. Even with data augmentation (DA), transformed views are typically rewarded independently, providing little signal once within-view rewards saturate. We propose \textbf{ConsistRoll}, a simple but effective method that injects cross-view consistency into RLVR training by reusing the group-sampling mechanism of GRPO. Specifically, ConsistRoll places original and semantically invariant transformed views in the same generation group, and assigns a joint reward only when paired completions are both correct and consistent. In this way, ConsistRoll turns consistency into an online credit-assignment signal, \textbf{without extra generation overhead and annotations}. Theoretically, we show that cross-view consistency is a valid inductive bias, and ConsistRoll introduces a cross-view correction term absent from DA, penalizing view dependence and alleviating advantage collapse. Comprehensive benchmarks across math, general-purpose, hallucination domains confirm that ConsistRoll achieves robust improvements in multimodal reasoning.

Problem

• cross-view consistency
• multimodal reasoning
• inductive bias
• reinforcement learning
• data augmentation

Innovation

• cross-view consistency
• inductive bias
• multimodal reasoning
• reinforcement learning
• ConsistRoll