Homepage
I am currently a master's student in Computer Application Technology at the University of Chinese Academy of Sciences. My work focuses on large language model training and evaluation, including SFT, RLHF/RLVR, reward modeling, rubric-based reinforcement learning, and agentic evaluation.
I have worked on AI Office agentic evaluation, safety expert model training, customer-service agent training, and multimodal/audio-language model research. I am also interested in practical training insights around KL placement, rollout staleness, off-policy effects, multi-reward optimization, online policy distillation, and token-level credit assignment.
Updates
Research
ICML 2025
An audio language model trained under text-only supervision, with Santa for transferring audio embeddings into the text embedding space while preserving audio semantics.
A rubric-based LLM-as-a-Judge reward system with Pairwise Adaptive Meta-Rubrics and Pointwise Verifiable Rubrics for open-domain reward supervision.
IEEE TIP
Region-level no-reference image quality assessment with component-level and object-level mask data, local quality scores, and a mask-based feature extractor.
ICLR 2026
A benchmark for probing multimodal large language models across human-centric perception, comprehension, and reasoning tasks.
AAAI 2027
A two-stage multi-reward RL method that first trains reward-specialized teachers and then merges their capabilities into one student through online policy distillation.
AAAI 2027
A two-stage plan-then-act framework that learns an explicit manipulation plan before generating future bimanual motion with a flow-matching diffusion transformer.
AAAI 2027
Counterfactual replay converts rubric-conditioned token likelihood differences into bounded weights for fine-grained credit assignment, without introducing an auxiliary token scorer.
A survey that organizes AI Office agents around operation, evaluation, and training, and studies the shift from AI inside Office to AI as Office.
Work
Worked on full-scenario customer-service ChatBot agent training and evaluation, including SFT, RLHF, data cleaning, distillation data construction, model training, and online effect tracking.
Trained Qwen3-32B with ms-swift, FSDP, and Megatron-LM, and built multi-stage RLHF workflows with point-wise and pair-wise reward models.
Background
M.S. student, Computer Application Technology.
B.E. in Software Engineering. Ranked 3/169, received three special scholarships, and was named an Excellent Graduate of Beijing and Beijing Jiaotong University.
Honors