Renyuan LIU

Renyuan LIU

MPhil student in AI at CUHK-Shenzhen

Renyuan LIU (刘稔rěn远)

Nice to meet you here! I am a Research Master’s (MPhil) student in Artificial Intelligence at the School of Artificial Intelligence, The Chinese University of Hong Kong, Shenzhen, under the supervision of Prof. Junjie Hu.

My current work spans AI infrastructure, AI agents, and LLM post-training, with a continuing interest in embodied intelligence and machine vision. I care about building efficient, reliable systems and turning research ideas into working implementations.

During my undergraduate studies, I was fortunate to be advised by Prof. Qinbing Fu at the Machine Life and Intelligence Research Centre, Guangzhou University, and Prof. Pengcheng Liu at the Department of Computer Science, University of York, UK.

I welcome collaborations across computer science.

Curriculum Vitae · TEAM NEXUS

Open source

GitHub Activity

Contribution calendar · last 8 months

GitHub contribution calendar for Ryannnice from 2026-01-01 to 2026-08-16
473 contributions in the last year · Updated Aug 16, 2026 Less More

News

Research Interest

I am interested in the full path from model behavior to efficient deployment:

  • AI Infrastructure: efficient inference and training, accelerator adaptation, profiling, and kernel optimization;
  • AI Agents: tool retrieval and planning, multi-turn context orchestration, memory, evaluation, and embodied decision-making;
  • LLM Post-Training: supervised fine-tuning, preference optimization, parameter-efficient fine-tuning, and data-centric evaluation;
  • Machine Vision (Undergraduate): neural models of insect vision for real-time collision detection and robotic navigation.

My undergraduate research focused on brain-inspired neural networks and machine vision. I modeled insect visual circuits for robust motion perception and translated them into collision-detection systems for robots. I enjoy hands-on coding, debugging, profiling, and end-to-end system implementation. My experimental platforms include TurtleBots equipped with a binocular camera and the micro ground robot Colias.

Selected code for neuro-embodied micro-robot Colias is available at:

Experience

Selected AI Systems Projects

100-Tool E-Commerce Shopping Assistant and Multi-Turn Tool Planning Agent System

Jul. 2026 – Present

Line chart showing the full tool-planning pipeline improving from 19.29% to 92.62% on 9,000 requests
Full-pipeline gains on the frozen 9K test set
Line chart showing multi-turn Agent accuracy improving to 94.56% on the independent test set
Multi-turn Agent ablation on the independent test set

View the complete project page

View project details

Project Background: As a core developer, I worked on an e-commerce shopping Agent spanning 100 tools and eight business domains, addressing full-tool overhead, confusion among similar tools, incorrect ordering of multi-intent requests, and cross-turn history contamination.

  • Intent Routing (100.0% Classification Accuracy): Built 48K e-commerce and casual-conversation samples and LoRA-tuned Qwen3-0.6B, raising accuracy on 9,000 isolated test cases from 84.8% to 100.0%. Compared with Qwen3-1.7B, median inference latency fell by 34.3% and throughput rose by 52.0%.
  • Efficient Tool Retrieval (97.4% Top-10 Full Tool Recall): Built 70K positive pairs and 96K typed negatives, fine-tuned bge-large with multi-negative contrastive learning, and trained bge-reranker-v2-m3 with 1:7 positive-to-hard-negative groups. Implemented dual BM25/vector recall, two-stage RRF fusion, and Cross-Encoder reranking with a “recall 40 → rerank 20 → plan 10” funnel. Top-10 full recall rose from 47.6% to 97.4%; compared with “rerank 30 → plan 20,” planner input tokens and median reranking latency fell by 36.8% and 33.7%.
  • Multi-Step Tool Planning (92.6% End-to-End Accuracy): Built 78K multi-style tool-call trajectories and LoRA-tuned Qwen3-8B with MS-SWIFT to output ordered tool chains from a fixed Top-10 candidate set. Constructed 4,000 preference pairs from real parent-model errors, using RPO for main training and low-learning-rate DPO for convergence. Exact tool-chain match improved from 85.7% to 88.9%, with 92.6% end-to-end accuracy on 9,000 static tests.
  • Multi-Turn Agent Framework (92.8% Context-Dependent Request Accuracy): Designed an external context orchestrator whose rule-based state machine schedules recent state, rule summaries, and in-session BM25 retrieval. It compresses necessary history into a standalone current request before invoking the original planner, while request boundaries and full-turn recovery suppress history contamination. A 12-stage ablation on nearly 2,000 natural conversations of 2–7 turns raised independent-test overall accuracy from a 50.5% direct-concatenation baseline to 94.6%; context-dependent and context-independent request accuracy reached 92.8% and 97.4%, while version-validation caching reduced planner calls by 60.3%.

Core Results: 100.0% intent accuracy on 9,000 isolated tests; 97.4% Top-10 full tool recall (47.6% baseline); 92.6% end-to-end accuracy; 92.8% context-dependent request accuracy; and 60.3% fewer planner calls.

Project 1 | DeepSeek-V4 Domestic-Accelerator Adaptation and Ascend Kernel Optimization

Mar. – Jun. 2026 · Shenzhen Research Institute of Big Data

Architecture of the Ascend kernel optimization Agent, including planning, coding, verification, knowledge retrieval, profiling, diagnosis, optimization, and testing
Ascend kernel optimization Agent architecture
View project details

Project Background: As a core developer, I participated in the full-parameter post-training of DeepSeek-V4-Pro on domestic accelerators, jointly conducted by Shenzhen Hetao College, Huawei, and 智子芯元. For DeepSeek-V4 inference and post-training on domestic hardware, I worked on Ascend 910C deployment evaluation, kernel-optimization Agent development, and fused backward-kernel optimization.

  • 910C Inference Adaptation and Evaluation (DeepSeek V4 Flash/Pro): Participated in deploying and evaluating DeepSeek V4 Flash/Pro on an Ascend 910C cluster. Compared Prefill/Decode performance between vLLM Ascend and SGLang on ShareGPT and other datasets to support inference-framework selection and deployment configuration.
  • Kernel Optimization Agent (Ascend Migration and Open Source): Addressed the manual analysis and repeated trial-and-error required by Ascend kernels by rebuilding MIT Kernel Design Agents for 910B/910C with a kernel knowledge base, optimization rules, and performance-diagnostic tools. Connected retrieval → generation → compilation → accuracy/performance verification → retain/rollback into a closed loop and open-sourced Kernel-Agent-Ascend.
  • Fused Backward-Kernel Optimization (20.9% Lower Latency): Used msprof to locate scalar address calculation on AIV, sparse memory access, synchronization waits, and load imbalance. Simplified address calculation, pruned low-hit data-movement branches, removed redundant synchronization, optimized tiling/on-chip caching, redistributed vector tasks, and fixed tail-block out-of-bounds access introduced by cache expansion. Kernel latency fell from 22.603 ms to 17.88 ms (1.26×), while cosine similarity remained 1.0 for all four outputs.

Core Results: Under the target configuration (BSND/BF16, B=1, S=4096, K=1024), fused backward-kernel latency fell from 22.603 ms to 17.88 ms, a 20.9% reduction (1.26×). All four output cosine similarities were 1.0, with a maximum relative error of 8.39e-4.

Project 2 | LLM Semantic Routing and AI Agent Applications

Mar. – May 2026 · Shenzhen Research Institute of Big Data

View project details

Project Background: To control multi-model serving costs and deliver an AI education product, I developed a lightweight semantic router and the backend of an AI education platform.

  • Lightweight Semantic Routing (62% Cost Reduction): Helped design domain partitioning from queries and domain-prototype embeddings, then combined model-internal representations to estimate question difficulty and assign models of different sizes. Reduced inference cost by 62% while keeping accuracy loss within 10%.
  • AI Agent Application (Backend Development): Built a Web AI education platform and Android client from scratch with Vue 3/Vite/Vue Router + FastAPI/SQLite/Docker. Implemented SSE streaming and multi-turn sessions, user/session/token monitoring, attachment upload, Markdown/code/web preview, and mobile adaptation, supporting news filtering, learning, and a Coding Agent workshop.

Manuscripts Under Review

  • R. Liu, H. Zhou, C. Fang and Q. Fu, [Manuscript under double-blind review.] Under review in The 2026 International Conference on Robotics and Automation (ICRA)(CCF-B).

  • M. Wang*, R. Liu*, and Q. Fu, [Manuscript under double-blind review.] Under review in IEEE Transactions on Cognitive and Developmental Systems (JCR Q1, IF = 4.9).

  • [Manuscript under double-blind review.] Under review in IEEE Robotics and Automation Letters (RA-L), in the area of embodied AI and semantic navigation.

Publications

Honors and Awards

  • First Prize (Provincial; Top 3%), Chinese Collegiate Computing Competition (4C)
    May 2025 [Evidence]

  • Honorable Mention (International), Mathematical Contest in Modeling (MCM)
    Jan. 2025 [Evidence]

  • First Prize (National; Top 5%), Asia and Pacific Mathematical Contest in Modeling (APMCM)
    Nov. 2024 [Evidence]

  • First Prize & Innovation Silver Award (Provincial; Top 2 out of 1,167 Teams),
    “Greater Bay Area Cup” Guangdong-Hong Kong-Macao Financial Mathematics Modeling Competition
    Nov. 2024 [Evidence]

  • Second-Class Scholarship (Top 8%), Guangzhou University
    Dec. 2025 [Evidence]

  • Third-Class Scholarship (Top 12%), Guangzhou University
    Nov. 2024 [Evidence]

  • First-Class Scholarship (Top 5%), Guangzhou University
    Dec. 2023 [Evidence]

  • Commendation Letter for Outstanding Performance in the Winning Team,
    Interdisciplinary Programme - New Media, Technological Innovation, Art and Data Mining at University of Hong Kong and The University of Macao
    Nov. 2023 [Evidence]

  • Provincial Key College Students’ Innovative Entrepreneurial Training Plan Program: Bio-Inspired LGMD Collision Detection Model Leveraging Optical Flow and Learning-Based Optimization (S202411078014) [Evidence]

Skills

  • LLM Post-Training & Agents: Qwen3, Transformers, SFT, DPO/RPO, LoRA, tool calling, external memory, multi-turn orchestration, and agent evaluation
  • Retrieval & Model Serving: embeddings, rerankers, BM25/vector hybrid retrieval, RRF, cross-encoders, vLLM, and SGLang
  • AI Infrastructure: CUDA, Ascend 910B/910C, profiling with msprof, inference benchmarking, and kernel optimization
  • Engineering: Python, C/C++, PyTorch, Hugging Face, MS-SWIFT, FastAPI, SQLite, Docker, Git, Vue 3, and Linux
  • Robotics & Embedded Systems: ROS, STM32, Keil, Webots, MATLAB, and binocular/micro-robot platforms
  • AI-Assisted Development: intensive daily use of Codex and Claude Code, with monthly usage on the order of 10 billion tokens
  • Languages: IELTS 6.5 (R8.0, L6.5, W6.0, S5.5) and CET-6 564 (242/248.5 in reading)
  • Documentation & Media: LaTeX, Markdown, MS Office/Visio, Adobe Photoshop, and Premiere Pro

  • Learning is one of the happiest things in the world.

    I enjoy movies, music, photography, basketball, jogging, badminton, hiking, and cooking.

Conference

TAROS 2025, York, United Kingdom

  • Fly-Inspired Ultra-selective Looming Perception and Avoidance on Resource-Constrained Micro-Robots, Poster.

Living Machines 2025, Sheffield, United Kingdom

  • A Biomimetic Collision Detection Visual Neural Model Coordinating Self-and-Lateral Inhibitions, 30-min Oral.

IJCNN 2025, Rome, Italy

  • Attention-Driven LPLC2 Neural Ensemble Model for Multi-Target Looming Detection and Localization, Video.

ICMPSO 2024, Guangzhou, China

  • A Broader Academic World.



Inspiring Articles

Visitor Map

Map of visitor countries