Dawei Zhou

Selective verification can boost accuracy and reduce compute costs, achieving 76.3% accuracy while cutting verification tokens by 91.2% compared to always verifying.

LLMs can identify some discrimination signals in assessment items, but their predictions fall significantly short of human benchmarks.

CRAM achieves a breakthrough in parameter efficiency and task retention by dynamically allocating resources based on task-specific needs, avoiding the pitfalls of both shared updates and isolated expansions.

Format-aware task prototypes in ProtoAda prevent inter-task interference, leading to superior performance in multimodal instruction tuning.

Decomposing CLIP-based classification into attribute extraction and aggregation reveals a new path to mitigating catastrophic forgetting in class-incremental learning.

Stop wrestling with monolithic MLLM codebases: Prism offers a plug-and-play framework for multimodal continual learning research.