Reproduce “CIEM: Contrastive Instruction Evaluation Method for Better Instruction Tuning” on OpenResearch

About this paper

Hongyu Hu, Jiyuan Zhang, Minyi Zhao, Zhenbang Sun

Researchers from ByteDance Inc. developed CIEM, an automated method to evaluate visual hallucination in Vision-Language Models (VLMs), and CIT, a contrastive instruction tuning technique that generates factual and negatively-answered training samples with justifications. This approach mitigates hallucination in VLMs, demonstrating improved factual grounding without substantial degradation in other multi-modal task performance.

Key insights

• Many existing VLMs exhibit a strong bias towards positive responses, evidenced by high Recall but low Precision in hallucination evaluation, often generating ungrounded assertions.
• Integrating detailed Chain-of-Thought (CoT) justifications in contrastive training data is critical for VLMs to learn reasoning paths, leading to more robust hallucination alleviation compared to binary answers alone.
• Instruction tuning with automatically generated factual and contrastive samples effectively reduces visual hallucination in VLMs by promoting more discerning and factually grounded responses.

Summary by alphaXiv.