关键差异：通过审计模型发现并修补能力缺口
Differences That Matter: Auditing Models for Capability Gap Discovery and Rectification
用强化学习训练一个 MLLM 审计器，主动生成反事实图像与困难问题来最大化目标模型间的分歧，从而无标注地定位视觉语言模型的失效模式，并把这些样本直接当作微调数据来修补能力缺口——这是一种可复用的模型诊断与数据自举流程。
原文依据 fine-tunes an MLLM as an auditor via reinforcement learning to generate challenging questions and counterfactual images
摘要译文
由模型翻译 · 术语统一 · 请对照原文核对关键结论现有多模态大语言模型（MLLM）的评测方法缺乏可解释性，且往往不足以充分揭示模型之间显著的能力差距。为此，我们提出 AuditDM，一个自动化框架，通过审计模型之间的分歧来主动发现并修补 MLLM 的失效模式。AuditDM 通过强化学习将一个 MLLM 微调为审计器，使其生成能够最大化目标模型间分歧的困难问题与反事实图像。训练完成后，该审计器能够挖掘出多样且可解释的样例，这些样例既揭示了模型的弱点，又可作为无需标注的数据用于修补。在应用于 Gemma-3 和 PaliGemma-2 等最先进模型时，AuditDM 发现了 20 多种不同的失效类型。基于这些发现进行微调，可在 16 个基准上一致地提升所有模型，并使一个 3B 模型超越其 28B 的对应版本。我们的结果表明，随着数据扩展的收益递减，有针对性的模型审计为模型诊断与改进提供了一条有效路径。
英文原文摘要 1052 字符
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.