A four-stage synthetic data pipeline that teaches multimodal models to verify each rule in a prioritized rubric lifts strict rule-following accuracy on a new benchmark from 9.5% to 30.1% on Qwen3-VL-4B.
Title resolution pending
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.LG 1years
2026 1verdicts
CONDITIONAL 1representative citing papers
citing papers explorer
-
PRISM: Priority-aware Rubric Internalization via Structured Multimodal Data Synthesis
A four-stage synthetic data pipeline that teaches multimodal models to verify each rule in a prioritized rubric lifts strict rule-following accuracy on a new benchmark from 9.5% to 30.1% on Qwen3-VL-4B.