I am a first-year Ph.D. student in Statistics and Data Science at the Wharton School of the University of Pennsylvania, advised by Professor Yuxin Chen. Before joining Wharton, I received my B.S. in Statistics from Peking University and was a visiting student at the University of California, Berkeley. Throughout my academic journey, I have been fortunate to work closely with Professors Lei Wu, Zeyu Zheng, and Peijie Zhou. My research focuses on the theoretical foundations of machine learning, including optimization theory & scaling laws, LLM reasoning & post-training & agentic RL, and applications of AI.

• [1] Momentum in Large-Batch Training: Polyak Enlarges the Critical Batch Size, Nesterov Improves Data Efficiency [Paper]
  Jia-Nan Wang*, Zixun Huang*, Kairui Li*, Lei Wu
  Preprint
  ► TL;DR: We characterize how Polyak and Nesterov momentum reshape one-pass large-batch training: Polyak enables larger batches without sacrificing data efficiency, while Nesterov further improves data efficiency in the large-batch regime.

  • Theory: In power-law kernel regression under a fixed data budget, we derive sharp critical-learning-rate scalings and full risk dynamics from the early transient through power-law decay to the noise floor. Optimizing the admissible learning rate and momentum yields a three-regime batch-size phase diagram: Polyak enlarges the critical batch size that preserves optimal small-batch scaling, whereas Nesterov's look-ahead mechanism suppresses noise accumulation and achieves better large-batch data scaling.
• [2] Functional Scaling Laws in Kernel Regression: Loss Dynamics and Learning Rate Schedules [Paper]
  Binghui Li*, Fengling Chen*, Zixun Huang*, Lean Wang*, Lei Wu
  NeurIPS 2025 Spotlight
  ► TL;DR: We propose a Functional Scaling Law using an intrinsic-time view of SGD, analyze learning-rate schedules, and validate the framework to obtain effective training schedules.

  • Theory: We analyze one-pass SGD in power-law kernel regression through an intrinsic-time reparameterization and establish a Functional Scaling Law for the full loss trajectory under arbitrary learning-rate schedules. The schedule enters through a convolutional functional, yielding explicit data- and compute-limited scaling relations for constant, exponential-decay, and warmup–stable–decay schedules.
  • Experiments: Power-law kernel simulations validate the predicted loss dynamics and scaling behavior. Pre-training experiments on 0.1B–1B LLMs further show that the Functional Scaling Law can fit and forecast loss trajectories across schedules, while confirming the efficiency gains of larger models, learning-rate decay, and warmup–stable–decay schedules.

• [1] FlowBalance: Verifier-Grounded Self-Improvement from On-Policy Reasoning Experience [Paper] [GitHub] [Blog]
  Zixun Huang*, Kishan Panaganti*, Haitao Mi, Leowei Liang
  Preprint
  ► TL;DR: We introduce FlowBalance, which uses verifier-calibrated privileged-hindsight self-guidance and trajectory balance to learn a normalized distribution over complete responses, improving reasoning performance, training stability, and correct-strategy diversity.

  • Method: For each on-policy rollout group, a frozen privileged-hindsight view of the same policy converts token-level log-probability gains into a trajectory-level self-guidance score. Verifier-derived advantages retain this guidance on positive-advantage trajectories, reverse it on negative-advantage ones, and disable it when the group provides no outcome preference. Profiled trajectory balance then fits the resulting reference-supported distribution with one log-partition estimate per group and no separate token-level imitation loss.
  • Theory: We prove that the induced target preserves within-group probability contrasts, uniquely minimizes reverse-KL displacement from the reference at its attained energy level, responds monotonically to verifier strength, and exactly corrects false-positive self-guidance on rejected responses.
  • Experiments: Across mathematical-reasoning benchmarks, FlowBalance achieves the strongest overall averages on Qwen3-4B and Qwen3-8B, improving the four-benchmark average over FlowRL by 1.67 and 1.98 points, respectively. It also trains faster and more stably, avoids direct OPSD's response-length collapse, and yields greater correct-strategy diversity on AIME24.
• [2] Recursive Synthesis for Long-Horizon Terminal Tasks [Paper] [Hugging Face] [Website]
  Zhongzhi Li*, Yucheng Shi*, Zongxia Li*, Ruhan Wang, Anhao Li, Zixun Huang, Junyao Yang, Lei Ke, Ninghao Liu, Haitao Mi, Leowei Liang
  Preprint
  ► TL;DR: We recursively evolve verified terminal tasks into progressively harder executable tasks, providing scalable, low-cost training data for long-horizon agents.

  • Method: Each accepted task becomes a seed for the next synthesis round. The pipeline expands the executable solution, realigns the instruction and verifier, and admits a task only when its reference solution runs successfully in a fresh sandbox and every semantic requirement is publicly specified or discoverable.
  • Task Evolution: Recursive synthesis produces 37,484 verified tasks through 15 rounds at about $0.05 per task. Across rounds, solution length, command count, tool use, and control-flow complexity grow substantially, while domain coverage remains stable and a fixed solver's success rate drops sharply.
  • Training: Trajectories collected from the synthesized tasks improve Qwen3.5 agents through both supervised fine-tuning and reinforcement learning on held-out long-horizon terminal benchmarks, showing that the generated tasks are not only harder but also useful for training.
• [3] Variance-Aware Baselines and Adaptive Learning Rates for Reinforcement Learning with Verifiable Rewards [Paper]
  Zixun Huang*, Jiayi Sheng*, Zeyu Zheng
  Under review
  ► TL;DR: We replace heuristic baseline and learning-rate choices in RLVR with a variance-optimal baseline and an SNR-adaptive learning-rate rule, whose combination yields OBLR-PO and the strongest overall performance.

  • Theory: Under KL-regularized policy optimization, we prove that the policy-gradient estimator is unbiased, derive exact variance expressions including the KL cross-covariance, establish an optimization-loss upper bound, and provide convergence guarantees.
  • Method: The analysis yields two complementary components: an adaptive learning-rate schedule governed by the policy gradient's signal-to-noise ratio and a variance-optimal, gradient-weighted baseline for the KL-regularized reward. Combining them gives Optimal Baseline and Learning-Rate Policy Optimization (OBLR-PO).
  • Experiments: On Qwen3-4B-Base, each component independently improves policy-optimization performance. The learning-rate rule further improves existing policy-optimization methods, while the complete OBLR-PO method achieves the strongest overall results.

• [1] TracingFlow: A Simulation-Free Trajectory Inference Framework Based on Second-Order Dynamics [Paper]
  Yuhao Sun, Zekun Wu, Zixun Huang, Peijie Zhou
  Under review
  ► TL;DR: We introduce TracingFlow, a simulation-free trajectory-inference framework that models continuous evolution with second-order dynamics by learning acceleration fields.

  • Theory: The framework decouples optimal transport-plan design from single-particle control and provides an iterative strategy for settings where initial velocities are unavailable.
  • Experiments: Results on synthetic and single-cell RNA-sequencing datasets show improved distribution reconstruction, trajectory fidelity, and preservation of biological priors.

• University of Pennsylvania, Wharton School

  Ph.D. in Statistics and Data Science

  Aug. 2026 – Present
• Peking University, School of Mathematical Sciences

  B.S. in Statistics, Elite Undergraduate Program for Applied Mathematics

  Sept. 2022 – Jun. 2026
• University of California, Berkeley

  Visiting Student

  Jan. 2025 – Aug. 2025

• Tencent HY – Qingyun Talent Program

  Research Intern

  Jun. 2026 – Aug. 2026

• Grand Prize in Challenge Cup Competition2026
• Weiming Scholar in Peking University2026
• Hong Sheng Scholarship2024
• First Prize in Chinese Mathematics Contest2024
• Yau Contest Groups Prize2024
• Xiaomi Scholarship2023
• Gold Prize in Chinese Mathematical Olympiad2021
• Silver Prize in Chinese Mathematical Olympiad2020

• A Guided Tour of Modern Statistics (in progressing)
• Random Matrix Theory (StatC206)