📝 Research & Publications

My research aligns users and agentic systems along three threads, with the infrastructure that runs them underneath:

• User–agent alignment. I externalize the LLM's latent understanding as an editable graph so users fix misunderstandings before code is generated (NeuroSync), and benchmark whether assistants learn users' cross-session ambiguity patterns to clarify less and code better (PACA). Ongoing work uses interaction history as verifiable rewards for RL.
• Augmenting users in data mining and sensemaking. A multi-agent system that quantifies the quality of qualitative data (QuAD), insight-guided retrieval for scientific surveys (InsightRAG), and conversational agentic media for deeper sensemaking than text-only chat (BirdRead).
• Augmenting agents. In a deployed WeChat agent, I moved from fixed workflows to skill-based agents and distilled them with sketch-of-thought SFT; I benchmark resilience to silent tool errors (RITE) and enable stable model replacement in latent-communication multi-agent systems (LatBridge).
• Infrastructure. Training and inference systems for running models on next-generation hardware (CUDA-Q ML, NVIDIA).

RITERITEICLR 2027 (under review), with WeChat Group, TencentA benchmark that injects silent tool errors into realistic e-commerce workflows to test whether agents notice, re-fetch and recover.LatBridgeLatBridgeACL ARR, to ACL 2027 (under review)Swap the model inside a latent-communication multi-agent system without breaking the messages the other agents rely on.InsightRAGInsightRAGACL ARR, to ACL 2027 (under review)Insight-guided retrieval that aggregates scholars' judgments scattered across papers to generate scientific surveys.PACAPACAAAAI Conference on Artificial Intelligence (under review)Benchmarks whether coding assistants learn a user's cross-session ambiguity patterns, so they ask fewer questions and write better code.BirdReadBirdReadCHI 2027 (under review), with Microsoft Research AsiaConversational agentic media that replace text-only chat with a probe-driven bird's-eye view for critical paper reading.QuADQuADNAACL (under review)A content-aware multi-agent system that quantifies the quality of qualitative data by aligning content and demographic distributions.CUDA-Q MLCUDA-Q MLISC High Performance 2026 (poster), done at NVIDIAMakes CUDA-Q quantum circuits differentiable nodes in PyTorch autograd, with distributed data-parallel training.NeuroSyncNeuroSyncACM Symposium on User Interface Software and Technology (UIST), 2025Externalizes the LLM's understanding as an editable graph so users fix misunderstandings before code is generated.SoTSoTDeployed · Paper to be releasedTencent WeChat (deployed)Sketch-of-thought agentic SFT: concise trajectories distilled into a small model that serves 21x more users per GPU in the WeChat agent. Deployed in production; the paper is still in progress and will be submitted to an industry track.

Hover a block for a summary, click it to jump to the paper.

All User–Agent Alignment Augmenting Users Augmenting Agents Infrastructure

ICLR 2027 · Under Review

RITE: Evaluating End-to-End Resilience to Silent Tool Errors in E-Commerce Agentic Systems

Wenshuo Zhang, Yuquan Lin, Yushi Sun, Yishi Lin, Rui Sheng, Peng He, Han Gao, Jikai Wang, Yifeng Rong

ICLR 2027 (under review), with WeChat Group, Tencent

arXiv

ACL 2027 · Under Review

LatBridge: Efficient and Stable Model Replacement in Latent-Communication Multi-Agent Systems

Jiajun Zhu, Haobo Li, Zixin Chen, Huamin Qu, Wenshuo Zhang* (corresponding author)

ACL ARR, to ACL 2027 (under review)

arXiv

ACL 2027 · Under Review

InsightRAG: Aggregating Distributed Scholarly Judgments for Scientific Survey Generation

Yuquan Lin*, Wenshuo Zhang* (corresponding author), Yilin Ye, Shuchang Xu, Zhao Chen, Weiqi Wang, Huamin Qu

ACL ARR, to ACL 2027 (under review)

arXiv

AAAI · Under Review

Fewer Clarifications, Better Code: Benchmarking Cross-Session Personalized Ambiguity Adaptation in Coding Assistants (PACA)

Zijian Xu*, Wenshuo Zhang*, Zisen Qin, Rui Sheng, Yushi Sun, Huamin Qu, Chuhan Shi

AAAI Conference on Artificial Intelligence (under review)

arXiv

CHI 2027 · Under Review

BirdRead: Probe-Driven Bird's-Eye View for Scope-Grounded Critical Paper Reading via Conversational Agentic Media

Wenshuo Zhang, Cheng Zhang, Baiqiao Zhang, Shuchang Xu, Huamin Qu, Yun Wang, Leixian Shen

CHI 2027 (under review), with Microsoft Research Asia

NAACL · Under Review

QuAD: Quantitative Assessment via Distribution-Alignment for Qualitative Data Quality Enabled by a Content-Aware Multi-Agent System

Wenshuo Zhang, Leixian Shen, Qiqi Zhou, Shuchang Xu, Feiyang Hao, Huamin Qu, Haobo Li, Zhao Chen, Linping Yuan

NAACL (under review)

arXiv

UIST 2026 · CCF-A

Sonic Stage: Auto-Generating Interactive Spatial Soundscapes to Facilitate Dialogue Video Comprehension for Blind Viewers

Shuchang Xu, Xiaofu Jin, Gaurav Jain, Wenshuo Zhang, Huamin Qu, Brian A. Smith, Yukang Yan

ACM Symposium on User Interface Software and Technology (UIST), 2026

🎖️ Best Paper Award (Top 1%)

arXiv

ISC 2026 · Poster

CUDA-Q ML: Enabling End-to-End Differentiable Quantum-Classical Computing

Wenshuo Zhang, Lyulin Kuang, Rita Zhang

ISC High Performance 2026 (poster), done at NVIDIA

PQE Exam

A Survey on LLM-Based Multi-Agent Workflow Automation (MAWA): Methods, Taxonomies, and Future Directions

Wenshuo Zhang, Huamin Qu

Ph.D. Qualifying Exam Survey Report, 2025

Slides

UIST 2025 (CCF-A)

NeuroSync: Intent-Aware Code-Based Problem Solving via Direct LLM Understanding Modification

Wenshuo Zhang, Leixian Shen, Shuchang Xu, Jindu Wang, Jian Zhao, Huamin Qu, Linping Yuan

ACM Symposium on User Interface Software and Technology (UIST), 2025

🎖️ Best Paper Honorable Mention Award (Top 1.5%)

WebsitearXiv

* Equal contribution, or corresponding author where marked. For the full list of publications and collaborations, please see my Google Scholar.