WaferSAGE: LLM-Powered Wafer Defect Analysis via Synthetic Data and Rubric-Guided RL
WaferSAGE serves as a framework designed for visual question answering related to wafer defects, utilizing compact vision-language models. It tackles the issue of limited data in semiconductor production through a three-phase synthesis process that includes generating structured rubrics. Initially, it begins with a small number of labeled wafer maps, where clustering-based cleaning filters help eliminate noise. Subsequently, vision-language models create defect descriptions that are transformed into evaluation rubrics. These rubrics facilitate the synthesis of VQA pairs, addressing aspects like defect type, spatial distribution, morphology, and root cause analysis. A dual assessment framework integrates rule-based metrics with LLM-Judge scores through Bayesian optimization for automated evaluation, employing curriculum-based reinforcement learning alongside Group Sequence Policy Optimization (GSPO) and rubric-aligned rewards.
Key facts
- WaferSAGE is a framework for wafer defect visual question answering.
- It uses small vision-language models.
- A three-stage synthesis pipeline incorporates structured rubric generation.
- Clustering-based cleaning filters label noise from limited labeled wafer maps.
- Defect descriptions are generated by vision-language models and converted into rubrics.
- Rubrics guide synthesis of VQA pairs for defect type, spatial distribution, morphology, and root cause analysis.
- A dual assessment framework aligns rule-based metrics with LLM-Judge scores via Bayesian optimization.
- Curriculum-based reinforcement learning uses Group Sequence Policy Optimization (GSPO) and rubric-aligned rewards.
Entities
—