01
Model Alignment
SFT, RLHF, RLVR, PPO, and GRPO pipelines designed for long-chain reasoning and robust policy improvement.
- Adaptive reward fusion
- Process & outcome supervision
AI researcher & engineer · USTC
Building language models that reason, verify, and scale.
I’m Songxin Qu (Jack), a graduate researcher focused on trustworthy reasoning systems—from verifiable reinforcement learning to efficient, multi-agent AI infrastructure.
LLM AlignmentVerifiable RewardsScientific ReasoningMulti-Agent SystemsDistributed Training
01 / Research
Making intelligence
more rigorous.
My work connects learning algorithms, reliable verification, and systems engineering to make advanced models more capable—and their reasoning more trustworthy.
02
Scientific Reasoning
Physics-consistent data synthesis and code-execution-grounded verification for specialized foundation models.
- Self-Instruct & rejection sampling
- Verifiable data engineering
03
AI Systems & Agents
Distributed training, high-throughput inference, and tool-using agent architectures built for real workloads.
- DeepSpeed ZeRO & vLLM
- LangGraph & code interpreters
02 / Selected work
Research in
motion.
Selected projects spanning model alignment, scientific data, and production-minded AI systems.
01
Process- & Verification-Aware Reward Modeling
An end-to-end alignment pipeline for scientific reasoning, introducing a Verification-Aware Reward Model and adaptive fusion of sparse execution rewards with dense semantic feedback.
+8.1% relative Pass@1
over SFT baseline
over SFT baseline
02
Multi-Agent Data Verification Framework
A Planner–Solver–Reviewer pipeline that translates scientific reasoning into executable checks, catches logical hallucinations, and self-corrects training data at scale.
92K+ verified scientific
reasoning samples
reasoning samples
03
Edge Vision & Multi-Agent Decision Systems
Deployed real-time detection and multi-agent control systems using PyTorch, YOLOv5, TensorRT, Vue, Spring Boot, and Docker—with national competition recognition.
98%+ recognition accuracy
in lab deployment
in lab deployment
03 / Profile
Research depth.
Engineering discipline.
I care about the full lifecycle of intelligent systems: how the data is constructed, how the model learns, how the reward is verified, and how the system performs under real constraints.
Currently pursuing an M.S. in Computer Technology at the University of Science and Technology of China, I aim to bridge theoretical alignment research and practical autonomous AI systems.
Start a conversation04 / Connect
Have a hard AI problem?
Let’s reason through it.
I’m always interested in conversations about LLM alignment, verifiable reasoning, agent systems, and open-source collaboration.
qsxjack@gmail.com