🌐 Language / 语言
English (current) · 中文
Hi, I’m Jialong Liu (柳佳龙), a M.S. student at Wuhan University (jointly trained with the SJTU BCMI Lab), advised by Prof. Zuchao Li. My research focuses on LLM post-training, with a particular interest in reinforcement learning for long-horizon agent tasks. I’m currently an Agent post-training intern at Xiaohongshu (RedNote).
📫 Galleons@whu.edu.cn · GitHub · 📞 (+86) 155-5821-6267
Research Interests
I work on post-training algorithms for large language models, with a main thread on optimizing LLMs in multi-turn / interactive settings within long-horizon agent tasks. Keywords:
Reinforcement Learning · Memory · Multi-turn Conversation · Long-Horizon Task · Continual Learning
I currently have 4 first-author papers accepted / under review, and actively contribute to open-source communities (vLLM-Omni, SGLang Diffusion, SGLang RL).
Selected Research
-
SRPO: Self-Reflective Policy Optimization for Long-Horizon Reasoning — ICML 2026 (Accepted) A self-reflective policy optimization framework that lets the model analyze its own reasoning trajectories and synthesize errors into a “Reflection Patch,” paired with a Reset-with-Memory mechanism that turns sparse terminal signals into dense token-level learning signals. Built on Qwen3-8B, it reaches SOTA with only 8% of the training FLOPs (AIME’24 73.3%) and leads clearly on agent benchmarks such as WebShop, ALFWorld, and SWE-Bench-Lite.
-
TTPO: Turn-Level Trajectory Optimization for Multi-Turn LLM Reasoning — EMNLP 2026 (Under Review) A critic-free RL method that reparameterizes GRPO at the dialogue-turn granularity. Across six multi-turn tasks it reduces run-to-run volatility by ~19–22 points and improves average performance by 9–12 percentage points.
-
HELM: Hierarchical Epistemic Learned Memory for Long-Horizon Agents — EMNLP 2026 (Under Review) A hierarchical epistemic learned-memory framework that exposes memory as an event-driven interface, building a three-tier nested memory store coupled with epistemic governance and a learned controller. It significantly improves success rates on GAIA and WebArena and supports auditable recall-provenance diagnostics.
-
CHOP: Segment-Level On-Policy Distillation for Long-Horizon Agents — NeurIPS 2026 (Under Review) A hybrid distillation method for long-horizon agents that dynamically assesses the reliability of token-level supervision and switches to segment-level Wasserstein distribution matching when reliability is low, fundamentally mitigating the degradation of conventional on-policy distillation.
Engineering & Projects
- Xiaohongshu (RedNote) · Agent Post-Training Intern (Jun 2026 – present): Working on the general-agent post-training of the next-generation dots model, improving its agentic capabilities across multi-turn tool use, planning, and long-horizon interactive tasks.
- Xiaomi · Inference-Acceleration Intern (Mar 2026 – Jun 2026): Contributing to Xiaomi’s inference-acceleration framework and the open-source omni-modal framework vLLM-Omni; optimizing in-house model deployment with CUDA Graphs, kernel fusion, and related techniques.
- Bank-Copilot: A long-horizon AI banking/finance agent built on a Harness architecture, combining Graphiti + FalkorDB into a dual “vector memory + graph memory” channel, running 24/7.
- Cascade-RAG: A multi-agent deep-research assistant built on LangGraph with a “plan–build–verify–repair” loop, reliably producing 10,000+ word reports.
Tech Stack
- Language: Python · C · TypeScript · Java · Rust · Shell
- AI Framework: PyTorch · LangGraph · Agno · Scikit-Learn
- AI Infra: vLLM · SGLang · VERL · Roll · SLIME
- Infrastructure: Slurm · Distributed Training · Docker · Git · CI/CD
Learning Notes
Notes from my self-study path in data science and machine learning, with a few classic open courses I recommend.
Math — Linear Algebra
- MIT 18.06 Linear Algebra
- MIT 18.065 Matrix Methods in Data Analysis, Signal Processing, and Machine Learning
- Introduction to Linear Algebra — Gilbert Strang
- Advanced Algebra — Weisheng Qiu
AI Infra Roadmap
- CS336: Large Language Models — Algorithms, Systems, and Applications
Further reading: A Survival Guide to a PhD — Andrej Karpathy
Navigation
- Full CV
- Tech blog: see the blog folder in the left sidebar
- Paper notes: see the paper folder in the left sidebar