CV

Jialong Liu

• Mobile: (+86) 155-5821-6267
• Email: Galleons@whu.edu.cn
• Github: github.com/Galleons2029

Education

• B.S. in Safety Engineering, XiangTan University, 2020 - 2024
• M.S. in Data Science, Wuhan University, 2025 - 2027 (expected)

Professional Experience

Research Experience

SRPO: Self-Reflective Policy Optimization for Long-Horizon Reasoning

• First Author, Accepted

• Wuhan University / Shanghai Jiao Tong University

• ICML 2026

• Proposed the Self-Reflective Policy Optimization (SRPO) framework, enabling models to analyze their own reasoning trajectories and summarize mistakes into Reflection Patches. Combined with a Reset-with-Memory mechanism, SRPO transforms sparse terminal rewards into dense token-level learning signals ($O(1)\to O(T)$), addressing the credit assignment problem in long-horizon tasks without external critics, reward models, or larger teacher models. Based on Qwen3-8B, SRPO achieved state-of-the-art performance using only 8% of the training FLOPs, reaching 73.3% on AIME’24 and significantly outperforming prior methods on agent benchmarks including WebShop (64.7%), ALFWorld (76.8%), and SWE-Bench-Lite (31.2%).

TTPO: Turn-Level Trajectory Optimization for Multi-Turn LLM Reasoning

• First Author, Under Review at Meta 3.5

• Wuhan University / Chinese Academy of Sciences / Shanghai Jiao Tong University

• EMNLP 2026

• Proposed Turn-Level Trajectory Policy Optimization (TTPO), a critic-free reinforcement learning method that reparameterizes GRPO at the dialogue-turn granularity and treats each turn as a macro-action. TTPO introduces three key components: Turn-Level Importance Ratio Surrogates (PPO-style clipping), Trajectory Pruning (removing low-reward branches to suppress exponential growth), and Cross-Turn Reward Propagation (propagating terminal rewards back to earlier turns). Across six multi-turn dialogue tasks, TTPO reduced run-to-run variance by approximately 19–22 $U_{90\text{-}10}$ points, improved average performance by 9–12 percentage points, and was consistently preferred in human blind evaluations while maintaining high-percentile quality ($A_{90}$).

HELM: Hierarchical Epistemic Learned Memory for Long-Horizon Agents

• First Author, Under Review

• Wuhan University

• EMNLP 2026

• Proposed the Hierarchical Epistemic Learned Memory (HELM) framework, exposing memory through event-driven interfaces and constructing a Structured Hierarchical Nested Memory (SHNM) consisting of episodic traces (${\mathcal{M}}^{(1)}$), integrated recollections (${\mathcal{M}}^{(2)}$), and thematic indices (${\mathcal{M}}^{(3)}$). These memory layers are connected through provenance links and cognitive metadata (timestamps, sources, tool states). HELM further incorporates Epistemic Governance and a learned controller that decides READ/WRITE/CONSOLIDATE/PRUNE operations under budget constraints to prevent high-confidence misuse of weak evidence. Under matched budgets and scaffolding conditions, HELM significantly improved success rates on long-horizon benchmarks including GAIA (L2: $8.9\to 12.8$; L3: $0.8\to 1.7$) and WebArena ($19.5\to 22.4$), while providing auditable memory provenance diagnostics.

CHOP: Segment-Level On-Policy Distillation for Long-Horizon Agents

• First Author, Under Review

• Wuhan University / Shanghai Jiao Tong University / Xiaomi

• NeurIPS 2026

• Proposed CHOP, a hybrid distillation method for long-horizon agents. CHOP dynamically estimates the reliability of token-level supervision during training and switches to segment-level Wasserstein distribution matching when supervision quality deteriorates. This approach fundamentally mitigates the performance degradation commonly observed in traditional On-Policy Distillation methods on long-horizon reasoning and complex agent tasks. CHOP significantly outperformed existing OPD and Top-K LSM baselines on AIME 2024 long-form mathematical reasoning and SWE-bench Verified software engineering tasks, enabling more stable and efficient long-context capability transfer.

Project Experience

Bank-Copilot: AI-Powered Financial Management Platform

• Built long-horizon agents using the Harness architecture and deployed backend agent services based on a data-driven design pattern with Flink, enabling stable 24/7 operation.
• Constructed graph-based memory with Graphiti and FalkorDB, structuring key entities, accounting relationships, and business events into a knowledge graph while combining retrieval systems to form a dual-channel architecture of vector memory and graph memory.
• Developed a Unix-like execution environment exposing file system and CLI tool access, expanding agent action spaces while reducing dependence on context window length.
• Unified tool invocation, fault-tolerance mechanisms, and Langfuse observability into a single state graph, leveraging PostgreSQL Checkpointer for session-level rollback and disaster recovery.

Cascade-RAG: Deep Research Assistant

• Adopted a Multi-Agent architecture for context management and complex task decomposition, ensuring consistency in long-horizon memory settings while reducing state drift.
• Designed a recursive multi-agent scheduling mechanism for dynamic task allocation and quality control, supporting generation of reports exceeding 10,000 words.
• Built a four-stage closed-loop workflow (“Plan-Build-Verify-Repair”) using the LangGraph Deep Agent framework, enabling iterative refinement and feedback-driven self-improvement.
• Optimized tool-calling paths through progressive disclosure and caching mechanisms to reduce redundant queries, while improving execution stability via sub-agent isolation.

SGLang Open Source Community Contributions

• Participated in fixing bugs within the SGLang Diffusion ecosystem, improving module stability.
• Tested compatibility and usability in agent scenarios, identifying issues related to training-inference consistency.
• Contributed to the development and testing of SGLang RL for multimodal multi-turn dialogue tasks, covering both training and inference pipelines.

Industry Experience

Agent Post-Training Intern

• Location: Shanghai

• Company: Xiaohongshu (RedNote)

• Duration: Jun 2026 – Present

• Participating in the general-agent post-training of Xiaohongshu’s next-generation dots model, enhancing its agentic capabilities across multi-turn tool use, planning, and long-horizon interactive tasks.

Inference Acceleration Intern

• Location: Haidian, Beijing

• Company: Xiaomi Corporation

• Duration: Mar 2026 – Jun 2026

• Contributed to the design and development of Xiaomi’s inference acceleration framework, optimizing model efficiency and resource utilization for multiple internal AI products.

• Contributed to the open-source multimodal inference framework vLLM-Omni and adapted it for internal deployment, identifying and resolving bottlenecks across the inference pipeline.

• Conducted performance analysis of locally deployed foundation models and optimized bottlenecks using techniques such as CUDA Graphs and kernel fusion.

Algorithm Engineer

• Location: Changsha, Hunan

• Company: Peking University Changsha Institute for Computing and Digital Economy

• Duration: Jun 2025 – Dec 2025

• Developed the intelligent assistant “Xiaosuan” for the HESI computing resource scheduling platform, integrating knowledge-base QA and operational guidance capabilities. Built upon a multi-agent routing architecture deeply integrated with platform tools, achieving over 90% QA accuracy and more than 80% long-horizon task completion rates.

• Implemented an Agentic Document Extraction solution for production-grade enterprise document parsing, overcoming limitations of conventional OCR systems on extreme formats and low-quality scans, improving parsing accuracy from approximately 80% to 95%.

• Enhanced the performance of locally deployed small models in real-world agent tasks using online policy distillation techniques.

Algorithm Researcher

• Location: Changsha, Hunan

• Company: Kunlunyuan Artificial Intelligence Co., Ltd.

• Duration: Mar 2025 – Jun 2025

• Followed cutting-edge research in Reasoning-RAG and reproduced/evaluated retrieval-native models such as Research-o1, Research-R1, and R1-Researcher in large-scale enterprise document retrieval scenarios.

• Designed and implemented a RAG-based agent system supporting text, image, and table inputs with multi-turn interactions, enabling closed-loop retrieval and reasoning in enterprise knowledge environments.

LLM Application Engineer

• Location: Changsha, Hunan

• Company: Yunyan Technology Co., Ltd.

• Duration: May 2024 – Mar 2025

• Served as a core developer of a graduate employment recommendation platform built with FastAPI, RabbitMQ, Qdrant, and MongoDB, supporting nearly 10 million registered users, over 100,000 daily active users, and more than 300,000 online job postings.

• Developed LLM agents to analyze students’ academic backgrounds, grades, and internship experiences, generating structured feature labels and leveraging FFM-based feature crossing for personalized job recommendations.

• Utilized Apache Kafka and RabbitMQ for asynchronous feature generation and notification delivery, improving throughput and system robustness under high-concurrency workloads.

Technical Skills

Category	Skills
Programming Languages	Python, C, TypeScript, Java, Rust, LaTeX, Shell
AI Frameworks	PyTorch, LangGraph, Agno, Scikit-Learn, Keras
AI Infrastructure	vLLM, SGLang, VERL, Roll, SLIME
Infrastructure & Systems	Slurm, Distributed Training, Docker, Optimization, Git, CI/CD Pipelines

Honors & Awards

Certifications

• MIT MicroMasters Program in Data Science, 2024.
• Rustling Open Source Operating System Training Camp Certificate, Spring 2023.

Awards

• Outstanding Winner Nominee, Mathematical Contest in Modeling (MCM/ICM), 2023.
• First Prize, Provincial Division, China Undergraduate Mathematical Contest in Modeling, 2023.
• Second Prize, National Graduate FinTech Competition, 2025.

Research Interests

My research focuses on algorithmic advances in the post-training stage of large language models, with a particular emphasis on reinforcement learning for intelligent agent systems.

I currently have four first-author RL papers accepted or under review. My work primarily investigates optimization techniques for multi-turn dialogue and interactive scenarios in long-horizon agent tasks, including topics such as Memory, Multi-turn Conversation, Long-Horizon Tasks, and Continual Learning.

In addition, I actively contribute to open-source communities, including projects such as vLLM-Omni, SGLang Diffusion, and SGLang RL. I am experienced in using vibe coding methodologies to manage and maintain large-scale production systems.