Closing Cost-Quality Gap in Document VLMs: Difficulty-Aware Data Curation and Quality-Adjusted Deployment Economics
A 35B-parameter MoE document VLM, with only 3B active at inference, is presented as cheaper than both human annotation and stronger open-source rivals.
The paper says the system is built for regulated document workflows where external models are off-limits and OCR cascades do not cover enough cases. Its training mix combines in-house production data with open-domain documents selected for layout diversity, extractable facts, and cross-model consistency. The authors report that it fits on a single H100 and beats deployable non-reasoning baselines up to much larger sizes. Their production-calibrated cost model puts expected savings above 80% versus humans and above 50% versus the best competing open-source model. ArXiv · AI/CL/LG's note
The paper says the system is built for regulated document workflows where external models are off-limits and OCR cascades do not cover enough cases. Its training mix combines in-house production data with open-domain documents selected for layout diversity, extractable facts, and cross-model consistency. The authors report that it fits on a single H100 and beats deployable non-reasoning baselines up to much larger sizes. Their production-calibrated cost model puts expected savings above 80% versus humans and above 50% versus the best competing open-source model. ArXiv · AI/CL/LG's note
score 5