CV

Ph.D.

Case Western Reserve University

Computer Science

• GPA: 3.83/4.0
• Expected graduation: May 2028

M.S.

New York University

Computer Engineering

• GPA: 3.93/4.0

Undergraduate and B.E.

Hunan University

Software Engineering (2015–2017); Digital Media Technology, a specialization of Software Engineering focusing on visual software and algorithm development (2017–2019)

• GPA: 83.16/100.0

Graduate Research Assistant

TAPP AI Lab, Case Western Reserve University

Conduct research on large language model reasoning, post-training, model architecture, and AI agent systems.

• Advisors: Xiaotian Han and Vipin Chaudhary

Teaching Assistant

Case Western Reserve University

Teaching assistant for Large Language Models and Database Systems: graded assignments and held office hours.

• Professors: Xiaotian Han (Large Language Models); Yinghui Wu (Database Systems)

Research Intern

NYU Game Innovation Lab

Explored research and methods for creative AI for video games.

• Developed enjoyment-focused game-playing agents for Street Fighter II via an extended PPO model and self-play
• Advisor: Julian Togelius, Professor, New York University

Research and Development Engineer

AsiaInfo Technology

Tested the CTDI project and improved its performance as a member of the QA team.

• Manager: Ted Wang

How to Loop MoE: Flatten the Experts, Untie the Attention

LLMs · Model Architecture · Mixture-of-Experts · Looped Model

• Proposed Foil, a looped MoE architecture that flattens expert layers, loops them more often, and gives each pass its own attention parameters while holding expert parameters and expert compute per token fixed
• Validated Foil with 20B-token pretraining and 100B-token continued training; every Foil model outperformed the unflattened baseline at 20B tokens, and the most flattened model finished 0.012 nat lower at 100B tokens with downstream accuracy on par or better
• Characterized expert utilization through coverage, load balance, and routing confidence, showing that wider expert layers and additional passes amplify one another and that routing-confidence peaks help diagnose diminishing returns from looping

Agent Team Work Zone

LLM Agents · Claude Code · Multi-Agent Workflow

[arXiv] · [GitHub] · [npm]

• Built and maintain an open-source persistence layer for Claude Code's Agent Teams, whose teammates vanish on session restart with no native recovery
• Designed per-agent workstations (role, working context, checkpoints, commitments) and a team registry, so one command rebuilds an interrupted team from disk
• Shipped 13 slash-command skills and hook-driven auto-checkpointing across four public releases, maintaining parallel editions through a breaking Claude Code API change
• Use the system end-to-end in its own development workflow: the agent team that builds it also runs on it

Path-Lock Expert: Efficient Hybrid Thinking Architecture

LLMs · Reasoning

• Accepted at COLM 2026
• Developed a parameter-efficient architecture approach to improve hybrid-thinking mode controllability (thinking vs. fast) without compromising overall task performance, together with an efficient training pipeline using curated data
• Experiments on AIME24, MATH500, and GPQA-Diamond indicate improved mode separation and controllability with competitive accuracy, and greatly improved accuracy for no-think mode

Demystifying Hybrid Thinking

LLMs · Reasoning · Post-Training

• Accepted at EMNLP 2026 Main
• Systematically analyzed the controllability of hybrid-thinking LLMs (think vs. no-think), showing incomplete mode separation and quantifying reasoning leakage via output length and reflection-token counts across MATH500, AIME24, GPQA, and MMLU-STEM
• Conducted controlled SFT ablations on data scale (20k–140k), pairing strategy, think:no-think ratio, and two-phase training; proposed a practical recipe that preserves accuracy while cutting no-think verbosity (1085→585 tokens on MATH500) and reflection-token frequency (5917→522)

Mid-Think: Token-Level Reasoning Budget Control

LLMs · Reasoning · RL

• Accepted at ACL 2026 Findings
• Identified via attention analysis that hybrid-thinking mode switching is dominated by a few token-level triggers—‘Okay’ activates reasoning while </think>\n\n suppresses it—rather than high-level Think/No-think instructions
• Proposed Mid-Think, a training-free prompting format combining both triggers to achieve Pareto-optimal accuracy–length trade-offs (MATH500: 92.1% accuracy at 2589 tokens vs. 94.4% at 4904), outperforming fixed-token and prompt-based baselines across Qwen3-4B/8B/14B/32B
• Applied Mid-Think to GRPO-based RL after SFT, cutting training time by about 15% (54h→46h) while improving Qwen3-8B on AIME (69.8%→72.4%) and GPQA (58.5%→61.1%)

Path-Lock Expert: Separating Reasoning Mode in Hybrid Thinking via Architecture-Level Separation

COLM 2026

Shouren Wang*, Wang Yang*, Chuang Ma, Debargha Ganguly, Vikash Singh, Chaoda Song, Xinpeng Li, Xianxuan Long, Vipin Chaudhary†, Xiaotian Han†

Demystifying Hybrid Thinking: Can LLMs Truly Switch Between Think and No-Think?

EMNLP 2026 Main

Shouren Wang*, Wang Yang*, Xianxuan Long, Qifan Wang, Vipin Chaudhary†, Xiaotian Han†

Mid-Think: Training-Free Intermediate-Budget Reasoning via Token-Level Triggers

ACL 2026 Findings

Wang Yang*, Shouren Wang*, Debargha Ganguly, Xinpeng Li, Chaoda Song, Vikash Singh, Vipin Chaudhary†, Xiaotian Han†

A Survey on Efficient Protein Language Models

TMLR 2026

Shouren Wang, Debargha Ganguly, Vinooth Rao Kulkarni, Wang Yang, Zhuoran Qiao, Daniel Blankenberg, Vipin Chaudhary†, Xiaotian Han†

How to Loop MoE: Flatten the Experts, Untie the Attention

Under Review

Shouren Wang*, Chuang Ma*, Mohsen Hariri, Debargha Ganguly, Wang Yang, Xiaoqing Tong, Qianying Liu, Xiaotian Han†, Vipin Chaudhary†

Agent Team Work Zone: An Automated, Persistent Workspace for Long-Lived Claude Code Agent Teams

Preprint, Open Source, npm Package

Shouren Wang

[arXiv] · [GitHub] · [npm]

Test-Time Scaling in Reasoning LLMs: Inference Regimes, Evaluation, and Reproducibility

Under Review

Mohsen Hariri, Weicong Chen, Nahal Shahini, Vikash Singh, Kai Ye, Amirhossein Samandar, Debargha Ganguly, Sreehari Sankar, Yanyan Zhang, Shouren Wang, Jerry Peng, Biyao Zhang, Michael Hinczewski, Vipin Chaudhary

When Domains Interact: Asymmetric and Order-Sensitive Cross-Domain Effects in Reinforcement Learning for Reasoning

Under Review

Wang Yang, Shouren Wang, Chaoda Song, Chuang Ma, Xinpeng Li, Nengbo Wang, Kaixiong Zhou, Vipin Chaudhary†, Xiaotian Han†

A Survey on Agent Skills for LLMs: A Lifecycle Perspective from Construction to Ecosystems

Under Review

Wang Yang, Chaoda Song, Xinpeng Li, Shouren Wang, Nengbo Wang, Yanyan Zhang, Chuang Ma, Debargha Ganguly, Vikash Singh, Shuai Xu, Jing Ma, Yu Yin, Vipin Chaudhary†, Xiaotian Han†

AgentCE-Bench: Agent Configurable Evaluation with Scalable Horizons and Controllable Difficulty under Lightweight Environments

Under Review

Wang Yang, Chaoda Song, Xinpeng Li, Debargha Ganguly, Chuang Ma, Shouren Wang, Zhihao Dou, Yuli Zhou, Vipin Chaudhary†, Xiaotian Han†

MemTrace: Probing What Final Accuracy Misses in Long-Term Memory

Under Review

Xianxuan Long, Zhikai Chen, Shenglai Zeng, Shouren Wang, Kai Guo, Jiliang Tang†

Reliability-Gated Source Anchoring for Continual Test-Time Adaptation

Under Review

Vikash Singh, Debargha Ganguly, Weicong Chen, Sabyasachi Sahoo, Sreehari Sankar, Biyao Zhang, Mohsen Hariri, Shouren Wang, Osama Zafar, Christian Gagné, Vipin Chaudhary†

CausalGuard: Conformal Inference under Graph Uncertainty

Under Review

Vikash Singh, Weicong Chen, Debargha Ganguly, Yanyan Zhang, Nengbo Wang, Sreehari Sankar, Mohsen Hariri, Alexander Nemecek, Chaoda Song, Shouren Wang, Biyao Zhang, Van Yang, Erman Ayday, Jing Ma, Vipin Chaudhary†