Qingzhuo Wang

Qingzhuo Wang

M.S. student at Tongji University. Interested in LLM post-training, quantitative research, and agentic RL.

Posts by Collection

portfolio

publications

Understanding and Defending VLM Jailbreaks via Jailbreak-Related Representation Shift

Published in NeurIPS 2026, 2026

For explicitly harmful inputs, VLMs can distinguish harmfulness yet shift into a distinct jailbreak state when images are added. JRS-Rem removes the jailbreak-related component of this representation shift at inference time to improve safety.

Recommended citation: Zhihua Wei, Qiang Li, Jian Ruan, Zhenxin Qin, Leilei Wen, Ruiyang Qin, Qingzhuo Wang, Dongrui Liu, Wen Shen. (2026). "Understanding and Defending VLM Jailbreaks via Jailbreak-Related Representation Shift." NeurIPS 2026.
Download Paper

TME-PSR: Time-aware, Multi-interest, and Explanation Personalization for Sequential Recommendation

Published in arXiv 2026, 2026

TME-PSR jointly models individual temporal rhythms, fine-grained interests, and personalized alignment between recommendations and explanations. It improves recommendation accuracy and explanation quality with lower computational cost on the evaluated datasets.

Recommended citation: Qingzhuo Wang, Leilei Wen, Juntao Chen, Kunyu Peng, Ruiyang Qin, Zhihua Wei, Wen Shen. (2026). "TME-PSR: Time-aware, Multi-interest, and Explanation Personalization for Sequential Recommendation." arXiv 2026.
Download Paper

Mitigating Action-Relation Hallucinations in LVLMs via Relation-aware Visual Enhancement

Published in ACL 2026, 2026

Action-Relation Sensitivity (ARS) identifies attention heads that localize action-relevant image regions. Relation-aware Visual Enhancement (RVE) boosts attention to these regions without additional training, reducing action-relation hallucinations with negligible inference overhead.

Recommended citation: Zhenxin Qin, Qiang Li, Qingzhuo Wang, Ruiyang Qin, Zhihua Wei, Wen Shen. (2026). "Mitigating Action-Relation Hallucinations in LVLMs via Relation-aware Visual Enhancement." ACL 2026.
Download Paper

Multilingual Safety Alignment via Self-Distillation

Published in NeurIPS 2026, 2026

Multilingual Self-Distillation (MSD) transfers safety from high-resource to low-resource languages using multilingual queries without external response data. It supports on-policy and off-policy training, with Dual-Perspective Safety Weighting (DPSW) emphasizing safety-critical tokens.

Recommended citation: Ruiyang Qin*, Qingzhuo Wang*, Dongrui Liu, Qiang Li, Zhihua Wei, Wen Shen. (2026). "Multilingual Safety Alignment via Self-Distillation." NeurIPS 2026.
Download Paper

A Unified Approach to Interpreting Knowledge Distillation for Large Language Models via Interactions

Published in ICML 2026, 2026

We identify interaction sparsification as a common pattern across knowledge distillation methods. Complex Interaction Penalty (CIP) explicitly promotes sparsity in complex interactions and improves distillation performance.

Recommended citation: Qingzhuo Wang*, Ruiyang Qin*, Zhenxin Qin, Wen Shen, Zhihua Wei. (2026). "A Unified Approach to Interpreting Knowledge Distillation for Large Language Models via Interactions." ICML 2026.
Download Paper

Evaluating and Explaining Prompt Sensitivity of LLMs Using Interactions

Published in ICML 2026, 2026

Interaction-based Prompt Sensitivity (IPS) reveals changes in LLM inference patterns that output-level metrics miss. Across 50 models, supervised fine-tuning, larger scale, dense architectures, and few-shot learning primarily stabilize low-order interactions.

Recommended citation: Ruiyang Qin, Qingzhuo Wang, Tian Wang, Zhihua Wei, Wen Shen. (2026). "Evaluating and Explaining Prompt Sensitivity of LLMs Using Interactions." ICML 2026.
Download Paper

AlphaDiverse: Post-Training Local Quantitative Research Agents for Diverse Exploration in Alpha Factor Mining

Published in arXiv 2026, 2026

AlphaDiverse post-trains local Planner and Realizer agents with diverse research traces, supervised fine-tuning, and joint GRPO. Across four Chinese stock universes, it combines competitive prediction with broader exploration in alpha factor mining.

Recommended citation: Qingzhuo Wang, Zikun Wei, Zhihua Wei, Wen Shen. (2026). "AlphaDiverse: Post-Training Local Quantitative Research Agents for Diverse Exploration in Alpha Factor Mining." arXiv 2026.
Download Paper

Bridging the Gap Between Harmfulness Belief and Refusal Behavior for Safety Alignment

Published in NeurIPS 2026, 2026

Bridging Harmfulness and Refusal (BHR) trains a LoRA adapter to carry harmfulness information from the instruction to the response-start position. Belief and belief-gated refusal losses improve jailbreak robustness while limiting over-refusal and preserving general capabilities.

Recommended citation: Lu Zhang, Chen Feng, Qingzhuo Wang, Wen Shen, Zhihua Wei. (2026). "Bridging the Gap Between Harmfulness Belief and Refusal Behavior for Safety Alignment." NeurIPS 2026.

Authorization Closure Graph: Minimal Repair for LLM Agents with Evolving User Instructions

Published in arXiv 2026, 2026

Authorization Closure Graph (ACG) tracks authorization dependencies as user instructions evolve, preserving valid permissions and identifying the minimal evidence or authority needed to proceed. Across three LLMs and two tasks, it improves action safety and task success.

Recommended citation: Qingzhuo Wang, CaiYi Wang, Jinglu Meng, Ruiyang Qin, Kunyu Peng, Zhihua Wei, Wen Shen. (2026). "Authorization Closure Graph: Minimal Repair for LLM Agents with Evolving User Instructions." arXiv 2026.
Download Paper

talks

teaching

Teaching experience 1

Undergraduate course, University 1, Department, 2014

This is a description of a teaching experience. You can use markdown like any other post.

Teaching experience 2

Workshop, University 1, Department, 2015

This is a description of a teaching experience. You can use markdown like any other post.