Guorui Sang

arXiv 2609.38348 · Submitted to ICLR 2027

• Proposed Recursive FST for operator learning to preserve local detail and capture long-range dependencies with a compact latent representation.
• Designed input-dependent spatial anchors and recursive attention over a continuous feature field. The model supports queries at arbitrary coordinates.
• Experiments evaluate the framework on PDE prediction and image classification.

Extended paper submitted to ICRA 2027

• Proposed Hamiltonian-guided diffusion to generate physically plausible robotic trajectories across variable horizons without a simulator at inference.
• Formulated an energy function from Hamiltonian residuals for guided sampling. Gradient guidance and importance resampling use this energy to steer the diffusion denoising process.
• Evaluation uses simulated robot-arm trajectories in MuJoCo. This ongoing project extends our ICLR 2026 ReALM-GEN Workshop paper.

M.S. thesis, Beijing University of Technology

• Built a diffusion model for pose-guided person image generation with a three-branch U-Net and mixed classifier-free guidance.
• Achieved the best SSIM and PSNR compared with PIDM, UPGPT, NTED, and DPTN.

ICASSP 2025 · Equal contribution

Jul 2024 - Dec 2024

• Co-developed a consistency model for efficient singing voice synthesis. Designed a quality scorer to select the denoising starting point.
• On PopCS: 2.8× faster than DiffSinger (RTF 0.018 vs. 0.052), better FAD (0.32 vs. 0.88), and MOS 3.88 vs. 3.81.