Yuxiao Hu胡愉霄

1. 01

   Symbolic regression

   Finding compact, interpretable equations efficiently: generative equation spaces, auxiliary search subspaces shared across solvers, and agent harnesses whose prompts and workflows are optimized jointly.

   • DASS
   • GenSR
   • HarnessSR
2. 02

   LLM research agents

   Agents that pursue scientific conjectures rather than single expressions: they model, ask questions and run experiments, then revise their conjectures with the evidence.

   • SCOPE
3. 03

   Trustworthy evaluation

   Benchmarks that score the evidence behind a discovery, not only the recovered equation, grounded in my earlier work on adversarial robustness.

   • ClosureBench
   • Focus on Hiders
   • GMAA

EarlierLLMs for time series (ICLR 2025), PM_2.5 forecasting (Environment International 2024), and fast multi-track locating for single-event effect testing (Nuclear Science and Techniques 2023).

1. NeurIPS2026 · Accepted

   DASS: A Solver-Adaptable Dynamic Auxiliary Search Strategy for Symbolic Regression

   Yuxiao Hu, Qian Li, Ding Wang, Yekun Zheng, Jinyue Yan, Yuntian Chen

   A shareable search strategy that plugs into different SR solvers, building and refining a subspace of high-potential basis terms; on LLM-SRBench it nearly doubles average accuracy with a 24% runtime increase.

   arXivCode
2. ICLR2026

   GenSR: Symbolic Regression Based on Equation Generative Space

   Qian Li, Yuxiao Hu, Juncheng Liu, Yuntian Chen

   Maps equations into a generative latent space that is symbolically continuous and numerically smooth, then searches it coarse to fine, consistently on the rank-1 Pareto front of accuracy, simplicity and runtime.

   PaperCode
3. ICLR2025

   Context-Alignment: Activating and Enhancing LLMs Capabilities in Time Series

   Yuxiao Hu, Qian Li, Dongxiao Zhang, Jinyue Yan, Yuntian Chen

   Aligns time series with the linguistic structure and logic LLMs already understand, rather than at the token level, and is particularly effective in few-shot and zero-shot forecasting.

   PaperCode
4. CVPR2024

   Focus on Hiders: Exploring Hidden Threats for Enhancing Adversarial Training

   Qian Li, Yuxiao Hu, Yinpeng Dong, Dongxiao Zhang, Yuntian Chen

   Identifies “hiders”, samples that adversarial training defends and then loses again, and trains against them alongside worst-case examples to improve accuracy and robustness together.

   PaperCode
5. Environ. Int.2024

   Domain knowledge-enhanced multi-spatial multi-temporal PM_2.5 forecasting with integrated monitoring and reanalysis data

   Yuxiao Hu, Qian Li, Xiaodan Shi, Jinyue Yan, Yuntian Chen

   Environment International, 192, 108997CAS Q1 Top

   Links station-level and city-level graphs across several time scales, cutting 72-hour PM_2.5 forecast RMSE by 6–10% relative to the strongest of nine baselines.

   Paper
6. CVPR2023

   Discrete Point-Wise Attack Is Not Enough: Generalized Manifold Adversarial Attack for Face Recognition

   Qian Li*, Yuxiao Hu*, Ye Liu, Dongxiao Zhang, Xin Jin, Yuntian Chen

   Extends face-recognition attacks from one image to many states of an identity, and from discrete points to a continuous expression manifold, for adversarial examples that generalize better with higher visual quality.

   PaperCode* Equal contribution
7. Nucl. Sci. Tech.2023

   An online fast multi-track locating algorithm for high-resolution single-event effect test platform

   Yu-Xiao Hu, Hai-Bo Yang, Hong-Lin Zhang, Jian-Wei Liao, Fa-Tai Mai, Cheng-Xin Zhao

   Nuclear Science and Techniques, 34, 72CAS Q1 Top

   A slot-based encoder–decoder that locates every heavy-ion track online for the Hi'Beam-SEE platform, at 238 fps with a 1.62 μm standard error.

   Paper

1. arXiv2026 · 2610.18938

   Beyond Equation Search: Conjecture-Guided Discovery with SCOPE and ClosureBench

   Yuxiao Hu, Qian Li, Ding Wang, Siheng Chen, Linfeng Zhang, Jinyue Yan, Yuntian Chen

   Scientist agents pursue competing conjectures through equation modeling, questions and active experiments, evaluated on ClosureBench, 60 interactive closure-discovery tasks across five scientific domains.

   arXiv
2. arXiv2026 · 2610.16966

   Learning to Design SR Solvers: Harness Optimization for Symbolic Regression

   Qian Li, Yuxiao Hu, Qihao Duan, Jinyue Yan, Yuntian Chen

   Treats the prompts and workflow of an SR agent as the object of optimization; on LSR-Synth it reaches 61.33% mean Acc_0.01, a 45% relative gain over the next-best baseline.

   arXiv