arXiv preprint
2026
First author
Master Student of Computer Science and Technology
Qingyu Zhang
I’m Qingyu Zhang, a second-year master’s student at Chinese Information Processing Laboratory in the Institute of Software Chinese Academy of Sciences. My current research focuses on AI sales and customer-service agents, especially reliable multi-turn interaction and evaluation in real-world business scenarios. I am also actively exploring User Agents for realistic evaluation and multi-turn interaction.
My internship experience spans foundation-model pretraining at Baichuan Intelligence, post-training and dialogue optimization at Meituan, and building user agents for evaluation and multi-turn training at ByteDance. Along the way, I have worked on AI-Salesman for reliable LLM-driven telemarketing, ShortGPT for layer pruning and model efficiency, and open-source toolkits such as AutoAlign and ShortX.
Research interests
- Agent
- LLM Long Context
- LLM Compression & Efficiency
- LLM Post-training
Education
- M.S. in Computer Science and TechnologyInstitute of Software, CAS2024 - Present
- B.S. in Computer Science and TechnologyCollege of Computer and Data Science, Fuzhou University2020 - 2024
News
- Jul, 2026 One paper “ShortOPD” is released on arXiv as first author.
- Dec, 2025 One paper “AI-Salesman” is accepted by AAAI 2026 as first author.
- Nov, 2025 Open-sourced ShortX project, a unified pruning toolkit for AI models.
- Jun, 2025 Contributed to AutoAlign project, an open-source toolkit for automated LLM alignment.
- May, 2025 One paper “ShortV” is accepted by ICCV 2025.
- May, 2025 One paper “ShortGPT” is accepted by ACL Findings 2025.
Experience
- Led the R&D of a User Agent framework supporting multi-turn evaluation needs across business lines, with over 80% of the generated evaluation data being business-usable.
- Exploring viable paradigms for applying the User Agent to multi-turn RL for Sales Agents.
- Led the R&D of an RL-based dialogue optimization system for large models, building the full pipeline of training, inference, and evaluation.
- Deployed in a live business environment, increasing core business conversion rate by 10%~20%.
- Published as first author (AI-Salesman, AAAI, 2026).
- Investigated Transformer redundancy and proposed a layer-based pruning method (ShortGPT, ACL Findings, 2025).
- Researched the lower bounds of RoPE Base (Base of RoPE Bounds Context Length, NeurIPS, 2024).
- Proposed a variant of the “Needle in a Haystack” evaluation method (Patent Granted).
Research Intern
·
Institute of Software, Chinese Academy of Sciences
Oct 2023 –
Sep 2024 · Beijing, China
- Adapted and optimized SFT/DPO algorithms for the Megatron framework (ACL Demo, 2025).
- Implemented large-scale distributed training on Ascend 910b using the ModelLink framework.
Publications
Projects
ShortX
A unified pruning toolkit for AI models. Currently includes ShortGPT and ShortV for efficient layer pruning.
AutoAlign - Automated Alignment Toolkit for LLMs
An open-source toolkit for automated alignment of Large Language Models. I was responsible for adapting and optimizing SFT/DPO algorithms for the Megatron framework.
Inference Acceleration for the 70B LLaMA-2 Large Language Model
As a coaching assistant for the ASC24 Student Supercomputer Challenge, I helped the team optimize the inference performance of the LLaMA-2-70B model using efficient frameworks like vLLM and designed data parallelism strategies to significantly reduce latency.
Training the 10-Billion Parameter Yuan-1.0 LLM
As a key member of the ASC23 team, I trained a 10B-level large language model using the DeepSpeed-Megatron framework, combining tensor, pipeline, and data parallelism. Our work won the First Prize.
Awards
- Jun, 2024 Honored as an Outstanding Graduate at Fuzhou University.
- May, 2023 Won the First Prize in the 10th ASC Student Supercomputer Challenge.
- Nov, 2022 Won the First Prize in the 13th National College Student Mathematics Competition.