UESTC · Zhufeng Program · Computer Science

Tinghao Wang

Sophomore Undergraduate Student in Computer Science and Engineering, University of Electronic Science and Technology of China.

I am a second-year undergraduate student in the School of Computer Science and Engineering, Zhufeng Program, UESTC. My research interests include Multimodal Large Language Models, Large Language Models, Vision-Language-Action Models, Embodied AI, efficient inference, visual token pruning, and streaming video understanding.

Selected Work Email GitHub CV
3.99 GPA / 4.00
94.76 Weighted Average / 100
1st Rank in Computer Science and Technology
4 Preprints and submissions

News

recent updates
  • 2026 Four works are currently under review at NeurIPS 2026 / ACMMM 2026, covering LVLM token pruning, hallucination mitigation, and VLM-agent evaluation.
  • 2025 Won the Best Positive Side Award and National Second Prize in the 16th China Undergraduate Physics Tournament.
  • 2024 Joined the Zhufeng Program at UESTC and started research on multimodal models and efficient inference.
  • 2023 Won the Silver Medal in the 40th Chinese Physics Olympiad Final, ranked 10th in Henan Province.

Preprints and Submissions

selected research works
TOPS

TOPS: First-Principles Visual Token Pruning via Constructing Token Optimal Preservation Sets for Efficient LVLM Inference

Tinghao Wang, Yichen Guo, Rui Huang, Zheng Lu, Qizhe Zhang, Chenxi Li, Yuan Zhang, Jiajun Cao, Zhirong Shen, Yaosong Du, Guangyan Gan, Wenya Wang, Lin William Cong, Shanghang Zhang

NeurIPS 2026 Conference, under review

A training-free visual token pruning framework that formulates pruning as Token Optimal Preservation Set construction and studies token selection through task relevance, information coverage, and semantic diversity.

Visual Token Pruning Efficient LVLM Training-Free
PAFI

PAFI: Training-Free Hallucination Mitigation in LVLMs via Dual-Stream Attention and FFN Intervention

Renyuan Li, Yichen Guo, Tao Luo, Tinghao Wang, Kai Tang, Wenya Wang

ACMMM 2026 Conference, under review

A training-free hallucination mitigation method that combines image-token attention amplification with FFN output attenuation at critical layers.

Hallucination Mitigation LVLM Inference Intervention
DCLA

Mitigating Hallucinations via Inter-Layer Consistency Aggregation in Large Vision-Language Models

Kai Tang, Jinhao You, Yichen Guo, Tao Luo, Tinghao Wang, Yuhan Rui, Renyuan Li, Jixian Wu, Xiuqi Ge, Hanze Li, Wenya Wang, Xiande Huang

ACMMM 2026 Conference, under review

A training-free decoding method that enforces semantic consistency across transformer layers to suppress unreliable predictions in LVLMs.

Decoding Inter-Layer Consistency LVLM Reliability
WRB

WildRoadBench: A Wild Aerial Road-Damage Grounding Benchmark for Vision-Language Models and Autonomous Agents

Bingnan Liu, Chenhang Cui, Rui Huang, Jiani Luo, Zhirong Shen, Tinghao Wang, Xiande Huang, Lingbei Meng, Fei Shen, An Zhang

NeurIPS 2026 Evaluations and Datasets Track, under review

A wild aerial road-damage grounding benchmark for evaluating direct VLM grounding and LLM-driven autonomous agent workflows.

Benchmark VLM Grounding Autonomous Agents

Research Experience

projects and contributions
NeurIPS 2026 Submission

TOPS: First-Principles Visual Token Pruning

Topic: Efficient inference and visual token pruning for Multimodal Large Language Models.

  • Designed a training-free framework with coarse visual-token selection after the vision encoder and progressive refinement inside the LLM decoder.
  • Formulated visual token pruning as a Token Optimal Preservation Set construction problem from a first-principles perspective.
  • Conducted experiments across 7 MLLM backbones and 14 benchmarks under different token reduction ratios.
  • Demonstrated that TOPS removes 77.8% of visual tokens on LLaVA-NeXT while preserving 100.0% performance on the 7B model and achieving 100.6% relative performance on the 13B model.
ACMMM 2026 Submission

PAFI: Dual-Stream Attention and FFN Intervention

Topic: Hallucination mitigation in Large Vision-Language Models.

  • Conducted mechanistic analysis showing that attention aggregates visual evidence while FFNs at critical layers may introduce overriding language priors.
  • Proposed a single-forward-pass training-free intervention that amplifies image-token attention and attenuates FFN outputs at critical layers.
ACMMM 2026 Submission

DCLA: Inter-Layer Consistency Aggregation

Topic: Training-free decoding and hallucination mitigation in LVLMs.

  • Proposed Decoding with Inter-Layer Consistency via Layer Aggregation for reducing hallucinations in LVLMs.
  • Validated that inter-layer consistency can improve factual reliability during inference.
NeurIPS 2026 Submission

WildRoadBench

Topic: Vision-language grounding, UAV road-damage perception, and autonomous research agents.

  • Contributed to a 1,061-image UAV road-damage grounding benchmark with professionally annotated ground truth.
  • Supported the design of a two-track evaluation protocol and a controlled agent evaluation harness.
  • Benchmarked 25 VLMs and 15 frontier LLM-driven agents.

Education

academic background
Sept. 2024 – Present

University of Electronic Science and Technology of China (UESTC)

Undergraduate Student in Computer Science and Engineering, Zhufeng Program.

  • GPA: 3.99/4.00; Weighted Average Score: 94.76/100.
  • Rank: 1st in Computer Science and Technology.
  • Selected coursework: College Physics I 99, College Physics II 97, Mathematical Analysis II 97, C Programming 96, Linear Algebra 95.

Honors & Awards

competitions and scholarships

Academic Competitions

  • Silver Medal, 40th Chinese Physics Olympiad Final, 2023; ranked 10th in Henan Province.
  • Provincial First Prize, 40th CPHO Semifinal Contest; ranked 12th in Henan Province and selected for the 2023 Henan Provincial Team.
  • Provincial Second Prize, 39th CPHO Semifinal Contest, 2022.
  • Outstanding Camper, Peking University Physics Golden Autumn Camp, 2023.
  • Second Prize, Yingcai Cup Electronic Design Competition, UESTC, 2024.
  • Third Prize, Mengsheng Cup Electronic Design Competition, UESTC, 2024.

CUPT and Academic Performance

  • Best Positive Side Award, 16th China Undergraduate Physics Tournament, 2025.
  • National Second Prize as a team member, 16th CUPT, 2025.
  • First Prize, CUPT Southwest Regional Tournament, 2025.
  • UESTC Special Scholarship for outstanding academic performance.
  • Built a strong mathematical and physical foundation through long-term physics olympiad training.