Xiang Zheng (郑想)

• Built ReGuide, a controllable planning framework that adapts classifier-free guidance to Tiny Recursive Models through exponential tilting and cross-window latent-state reuse.
• Achieved a state-of-the-art D4RL score of 111 on Hopper with 10× fewer parameters than the baseline while analyzing why recursive models can outperform diffusion models on reasoning tasks.

Advised by Dr. Fangkai Yang and Dr. Lu Wang.

• Contributed multimodal continuous-token rollouts for AgentLoop, including an end-to-end DeepEyes-v1/Qwen2.5-VL-7B GRPO recipe on 8×A100 GPUs, merged upstream in volcengine/verl#6804.
• Contributed On-Policy Self-Distillation and led reference-parity validation, reproducing official TRL training dynamics and Avg@12 results across AIME24/25 and HMMT25 in volcengine/verl#6909.

View verl on GitHub ↗

• Built ODE-1000, a 1,000-task benchmark with symbolic-equivalence diagnostics, and fine-tuned five Qwen and Llama models on 900 examples across 8 GPUs.
• On Qwen3-0.6B, increased execution rate from 27% to 97% and successful-run accuracy from 63.0% to 87.6%, while reducing mean relative L2 error from 0.371 to 0.104.
• Improved in-context operator learning across Transformer architectures by 50% through scientific data augmentation, with stronger generalization on forward and inverse ODE/PDE tasks.

Advised by Prof. Michael Mahoney.

• Built a multi-resolution wavelet-domain diffusion framework with inference-time guidance for indirect control of physical systems and released the first 2D incompressible fluid benchmark for this setting.
• Designed a BIC-based scoring framework that constructs per-token DAGs from discrete-diffusion probabilities and incorporates causal structure into sampling to improve accuracy and throughput.
• Developed a discrete diffusion model that learns inter-token masking schedules through maximum-entropy reduction, producing a more interpretable generation order.

Advised by Prof. Tailin Wu.