Small Vision Models Now Rival Frontier AI in Factory Inspection
A new distillation method allows 3B-parameter models to outperform frontier VLMs in industrial defect detection using as few as 18 labeled images.
Industrial visual inspection is shifting from massive, static datasets to rapid, few-shot adaptation. This research introduces answer-conditioned chain-of-thought distillation, a process that allows small vision-language models (VLMs) to learn complex industrial tasks from a handful of examples. By forcing a frontier model to generate reasoning that leads specifically to a known correct label, the researchers created a high-quality training set for a 3B-parameter model.
This approach solves a critical failure point in AI distillation: the tendency of frontier models to hallucinate incorrect reasoning. Without answer-conditioning, wrong reasoning degraded performance by 17.8 percentage points. By ensuring the reasoning path always hits the correct conclusion, the small model achieves higher accuracy than direct fine-tuning across all tested industrial tasks.
For manufacturing, this removes the primary barrier to AI adoption: the need for thousands of labeled defect images. The method works with only 18 to 30 labeled images per task. This allows factories to deploy specialized inspection models for new product lines or emerging defect types in hours rather than months.
Hardware requirements for quality control are also dropping. The 3B-parameter model outperformed GPT-4.1 by 10.0 percentage points on weld radiograph classification. This means high-precision inspection can move from expensive cloud clusters to local, edge-computing hardware on the factory floor without sacrificing accuracy.
Sector-wide, this signals a move toward 'lean' AI deployment. Companies providing automated optical inspection (AOI) can now offer rapid customization for niche industrial modalities where data is scarce. The ability to outperform frontier models with minimal data suggests that specialized, small-scale models are becoming the optimal choice for industrial precision.
The efficiency of this distillation means the competitive advantage in industrial AI is shifting from who owns the most data to who has the best distillation pipeline. The focus now moves to the speed of adaptation for new manufacturing defects.