Differences That Matter: Auditing Models for Capability Gap Discovery and Rectification
This addresses the need for better model diagnosis and improvement in multimodal AI, offering a targeted approach as data scaling becomes less effective.
The paper tackles the problem of interpretable evaluation for multimodal LLMs by introducing AuditDM, an automated framework that discovers and rectifies capability gaps, leading to a 3B model surpassing its 28B counterpart and consistent improvements across 16 benchmarks.
Conventional evaluation methods for multimodal LLMs (MLLMs) lack interpretability and are often insufficient to fully disclose significant capability gaps across models. To address this, we introduce AuditDM, an automated framework that actively discovers and rectifies MLLM failure modes by auditing their divergence. AuditDM fine-tunes an MLLM as an auditor via reinforcement learning to generate challenging questions and counterfactual images that maximize disagreement among target models. Once trained, the auditor uncovers diverse, interpretable exemplars that reveal model weaknesses and serve as annotation-free data for rectification. When applied to SoTA models like Gemma-3 and PaliGemma-2, AuditDM discovers more than 20 distinct failure types. Fine-tuning on these discoveries consistently improves all models across 16 benchmarks, and enables a 3B model to surpass its 28B counterpart. Our results suggest that as data scaling hits diminishing returns, targeted model auditing offers an effective path to model diagnosis and improvement.