Jiarong Xu

ChartWalker reveals significant performance gaps in cross-chart RAG tasks, challenging the status quo of existing benchmarks and paving the way for more robust multi-modal reasoning.

PrivacyCredit achieves top-tier credit risk prediction accuracy without compromising consumer privacy or model confidentiality, setting a new standard for secure financial assessments.

LLM agents can explore more effectively by retrieving and reasoning over off-policy step-level traces, leading to significant performance gains and faster training.