Xianliang Liu

Multi-turn agent RL · Livestream Q&A

Responsible for post-training a digital human livestream commerce Q&A agent powered by an in-house MoE model. Built multi-turn agent RL environments with verl, using GRPO and related methods including GiGPO to train ReAct-based reasoning and tool use.

• Designed reward components for output format, response length, trajectory turn count, and retrieval of relevant RAG information.
• Implemented separate product and store retrieval tools, combining keyword/BM25 retrieval with vector search to supply relevant information for answers.
• Connected multi-turn information gathering to livestream Q&A, with average end-to-end response time under 2 seconds.

Qwen assistant · Post-training & planner rewards

Contributed to Qwen assistant development and Qwen3 post-training for task planning and tool use. Used verl and GRPO for reinforcement learning, with responsibility for reward design for the Planner module.

• Contributed to multi-agent workflows spanning intent clarification, complex task decomposition, planning, and tool use across services.
• Worked on synthetic query generation, data filtering and distillation, and multi-source RL rewards using LLM-as-judge.
• Supported YaRN-based context extension and decoupled training/inference deployment, including routing scripts adapted to Alibaba’s infrastructure.
• Enabled 4B and 32B models to replace Qwen-Max in several modules.

Visual reasoning · Post-training & evaluation

Built 400,000 image-text training examples spanning math, code, and open-ended QA. Applied SFT, DPO, and GRPO, designed reference-based RL rewards, and developed automated business-task evaluations with OpenCompass.

Algorithm intern · Assisted driving

Developed vehicle pose estimation and parking-space detection and tracking for assisted parking, with deployment and testing on a modified BYD Han EV.