Skip to content
Book Open access

STEER: Domain-Routed Multi-LoRA Teacher On-Policy Distillation for Cross-Domain PCBA Visual QA

Nov 2026 · Proceedings of the 34th ACM International Conference on Multimedia · 0 citations · 20 references

Abstract

The PCBA Standard-to-Real Grand Challenge evaluates multimodal models on a form of industrial visual question answering (VQA) that is difficult for generic vision-language models: a system must transfer manufacturing knowledge from standards-derived examples to cluttered real production imagery, while answering not only perceptual questions but also defect-cause and handling questions. This paper presents STEER, a domain-routed on-policy distillation framework designed around this cross-domain structure. Instead of training a single model on a naively mixed corpus, STEER first specializes two homologous LoRA domain teachers: a standard-domain LoRA teacher for rule, cause, and handling knowledge, and a real-world LoRA teacher for high-resolution PCBA perception, template comparison, and dense counting. Since both teachers are lightweight adapters over the same base model, the cost of training multiple teachers is much lower than maintaining independent full checkpoints. A single deployable student is then trained on its own generated trajectories, with each prompt routed to its corresponding domain teacher and dense token-level supervision supplied by teacher prefill probabilities. The on-policy design reduces exposure bias from static teacher completions, while answer-level verifiers keep the optimization aligned with normalized challenge metrics. We use Qwen3.5-9B for controlled ablations and scale the final submission to Qwen3.5-27B. On the official public evaluation, our final system obtains an overall score of 90.920129 and ranks first on the leaderboard.

Read PDF

We use cookies to run the site and, with your consent, for analytics and to show ads. See our Cookie Policy.