Small VLMs achieve high accuracy in industrial vision with new CoT distillation technique

Researchers have developed a new method called answer-conditioned chain-of-thought (CoT) distillation to efficiently adapt small vision-language models (VLMs) for industrial visual inspection tasks. This technique uses minimal labeled data by having a larger VLM generate explanations for correct labels, which are then used to fine-tune a smaller 3B-parameter model via LoRA. The method significantly improves performance over direct fine-tuning and even surpasses GPT-4.1 on specific tasks like weld radiograph classification, demonstrating its effectiveness with limited data. AI

IMPACT Enables rapid deployment of AI visual inspection in manufacturing with minimal data, outperforming larger models on specific tasks.

RANK_REASON The cluster contains an academic paper detailing a new method for adapting AI models. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.AI →

• AI-based visual inspection
• Answer-Conditioned Chain-of-Thought Distillation
• arXiv
• Few-Shot Industrial Vision
• GPT-4.1
• Lora
• Small VLMs
• Vision--Language Models

• paper
• model release

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →