Hui Chu

Low-quality modalities can quietly sabotage multi-modal networks because they receive minimal training gradients yet dominate test-time skip connections—a failure mode solved by explicitly filtering and rebalancing degraded feature streams.

Medical LLMs can be significantly improved by explicitly modeling and injecting fine-grained clinical criteria, leading to substantial gains in both accuracy and safety compliance, rivaling proprietary models.