Xu Zhang

IRDS introduces a novel data selection method that uses a verifier-coupled sparse autoencoder framework to efficiently select high-quality Reinforcement Learning with Verifiable Rewards (RLVR) training instances, achieving state-of-the-art performance on multiple reasoning benchmarks.

The paper proposes CAGE-CAL, a counterfactual graph calibration framework, to accurately assess the reliability and detect over-confidence in multi-agent LLM systems after agents communicate.

This paper proposes Privileged Hidden Flow (PHF), an extension to On-policy self-distillation (OPSD) that aligns token-to-token transition directions and trajectory geometry between a student and privileged teacher.

This paper introduces Active Task Driving Memory (ATMem), an actively maintained execution state for mobile GUI agents, and STR-GRPO, an online reinforcement learning method that uses ATMem selectively.

This paper introduces DynaKRAG, a method for multi-hop retrieval-augmented generation that learns a shared policy for evidence operations, achieving state-of-the-art results on three benchmarks.

This paper proposes SpikingMOT, a spike-driven multi-object tracking system that uses spiking neural networks and adaptively models sparse trajectory dynamics.