VideoThinker framework improves lightweight MLLMs' video reasoning via causal debiasing

Researchers have developed VideoThinker, a novel framework designed to enhance the reasoning capabilities of lightweight multimodal language models (MLLMs) in video analysis. This approach addresses the issue of perceptual bias, where models tend to rely on superficial data patterns rather than genuine understanding. VideoThinker employs a two-stage debiasing process, first creating a 'bias model' to capture shortcut behaviors and then using a Causal Debiasing Policy Optimization (CDPO) algorithm to steer the primary model towards accurate reasoning. AI

影响 Introduces a method to improve video reasoning in lightweight MLLMs, potentially enabling more efficient on-device AI applications.

排序理由 This is a research paper detailing a new framework and algorithm for improving MLLM video reasoning. [lever_c_demoted from research: ic=1 ai=1.0]

在 arXiv cs.CV 阅读 →

• MLLMs
• MVBench
• TempCompass
• VideoMME
• VideoRFT-3B
• VideoThinker
• Video-UTR-7B

• 论文
• 模型发布

AI 生成摘要 · Google Gemini · 来自 1 个来源。 我们如何撰写摘要 →