Yuqiao Tan

I am a first year master student at Institute of Automation, advised by Prof. Shizhu He. Before that, I received my B.E. from UESTC.

My recent research focuses on User-Agent Collaboration [SWE-Touch], LLM reasoning [DyPRAG/Zero-Step], LLM interpretability [LaTen/BuPO], and reinforcement learning [BuPO/PreRL].

Email  /  Github  /  Google Scholar

Yuqiao Tan portrait
Publications

2026
SWE-Touch: Benchmarking Coding Agents When Users Touch the Code
Yuqiao Tan, Jinxiang Meng, Fangyu Lei, Minzheng Wang, Shizhu He, Jun Zhao, Kang Liu
Preprint
TL;DR: SWE-Touch injects validated, task-conflicting user edits into live coding-agent trajectories to evaluate state awareness and adaptation in shared workspaces.
paper  /  project  /  code
SWE-Touch benchmark construction and shared-workspace evaluation framework
From P(y|x) to P(y): Investigating Reinforcement Learning in Pre-train Space
Yuqiao Tan*, Minzheng Wang*, Bo Liu, Zichen Liu, Tian Liang, Shizhu He, Jun Zhao, Kang Liu
Preprint
TL;DR: DSRL warms up models with negative-sample reinforcement on unconditional P(y), then reincarnates the policy into conditional P(y|x) reinforcement learning.
paper  /  code
Comparison between post-train-space and pre-train-space reinforcement learning

2025
Bottom-up Policy Optimization: Your Language Model Policy Secretly Contains Internal Policies
Yuqiao Tan*, Minzheng Wang*, Shizhu He, Huanxuan Liao, Chengfeng Zhao, Qiunan Lu, Tian Liang, Jun Zhao, Kang Liu
Preprint
TL;DR: BuPO first optimizes a lower-layer internal policy to rebuild foundational reasoning, then switches to full-model reinforcement learning for refinement.
paper  /  code
Bottom-up Policy Optimization internal policy framework
The Zero-Step Thinking: An Empirical Study of Mode Selection as Harder Early Exit in Reasoning Models
Yuqiao Tan, Shizhu He, Kang Liu, Jun Zhao
NeurIPS 2025 Efficient Reasoning Workshop
TL;DR: Mode selection is a harder early-exit problem because it must choose a reasoning mode before observing any question-specific reasoning trajectory.
paper  /  code
Comparison of iterative Early Exit and zero-step Mode Selection
Neural Incompatibility: The Unbridgeable Gap of Cross-Scale Parametric Knowledge Transfer in LLMs
Yuqiao Tan, Shizhu He, Kang Liu, Jun Zhao
ACL 2025
TL;DR: Cross-scale parametric knowledge transfer requires alignment, yet behavioral and structural differences create neural incompatibility that destabilizes both PostPKT and PrePKT.
paper  /  code
Language-based and parameter-based cross-scale knowledge transfer frameworks
Dynamic Parametric Retrieval Augmented Generation for Test-time Knowledge Enhancement
Yuqiao Tan, Shizhu He, Huanxuan Liao, Jun Zhao, Kang Liu
Preprint
TL;DR: DyPRAG trains a lightweight translator to convert retrieved documents into dynamic LoRA parameters, enhancing LLM knowledge efficiently at test time.
paper  /  code
DyPRAG offline parameter translation and online inference framework
RobustPT: Dynamic Disentanglement Prompt Tuning in Vision-Language Models with Missing Modalities
Ruiting Dai*, Yuqiao Tan*, et al.
ICMR 2025
TL;DR: RobustPT dynamically disentangles modality-specific prompts before softly aligning them, reducing channel bias and improving robustness when vision or language is missing.
paper  /  code
RobustPT multimodal prompt generation and fusion framework

2024
MuAP: Multi-step Adaptive Prompt Learning for Vision-Language Model with Missing Modality
Ruiting Dai*, Yuqiao Tan*, et al.
Preprint
TL;DR: MuAP generates modality-specific complete and missing prompts, fuses them into ViLT, and iteratively aligns text and vision through multi-step tuning.
paper
MuAP multimodal prompt generation, fusion, and multi-step tuning framework
G-SAP: Graph-based Structure-Aware Prompt Learning over Heterogeneous Knowledge
Ruiting Dai*, Yuqiao Tan*, et al.
ICMR 2024
TL;DR: G-SAP builds refined evidence graphs, derives structure-aware prompts for frozen language models, and performs heterogeneous graph-text message passing for commonsense reasoning.
paper
G-SAP evidence graph, structure-aware prompting, and message-passing framework
Education

M.E., Pattern Recognition and Intelligent System, Institute of Automation, UCAS, 2025 - present

B.E., Software Engineering, University of Electronic Science and Technology of China, 2021 - 2025

Internship

Tsinghua University, SIG, Research Intern, 2023.07 - 2024.05, Focus on GNN, IoT

ByteDance, DCar-AI-Y, Research Intern, 2024.01 - 2024.07, Focus on RAG, GenIR

Award

Outstanding Graduate of Sichuan Province, 2024

First Prize, Baidu Business AI Technology Innovation Competition (80000 RMB), 2024

Soong Ching Ling Scholarship, UESTC, 2023

National Scholarship, Ministry of Education, 2022

Invited Talk

NICE - Internal Policy of LLMs and Reinforcement Learning, 2026.01 [Video]

Reviewer

EMNLP 2026, COLM 2026, ICMR 2025, NeurIPS ER 2025