Efficient learning
Parameter-efficient adaptation, expert fusion, and policy distillation for transferring knowledge into capable models with practical training costs.
Post-training · PEFT · DistillationPh.D. candidate · Shanghai Jiao Tong University
Jiacheng Ruan阮佳程
Efficient learning. Multimodal reasoning.
I am a Ph.D. candidate at Shanghai Jiao Tong University. My research focuses on efficient post-training, reasoning, and evaluation of multimodal large language models.
I work on helping models learn with fewer resources, use visual evidence more faithfully, and demonstrate their capabilities through carefully designed benchmarks.
01
Learning · Reasoning · EvaluationResearch interests
Multimodal reasoning
Inference methods that keep generated answers grounded in visual evidence and reduce hallucinations in complex multimodal tasks.
Visual grounding · Decoding · Reward modelsRigorous evaluation
Benchmarks and datasets that reveal what models understand, how they reason, and where their visual and scientific capabilities fall short.
Scientific reasoning · Benchmarks · Data02
Recent updates
- Updated our Contrastive On-Policy Distillation preprint.
- MMGist was accepted to Findings of EMNLP 2026.
- ExFusion was accepted by IEEE Transactions on Multimedia.
- MME-SCI was accepted by AAAI 2026.
Earlier updates
- MPI-CD was accepted by ACM Multimedia 2025.
- VLRMBench was accepted by ICCV 2025.
- MM-CamObj and TTE were accepted by AAAI 2025.
- LLaMA-MoE was accepted by EMNLP 2024.
- GIST was accepted by ACM Multimedia 2024.
03
Full list on Google Scholar ↗Selected publications
* Equal contribution. Preprints are labeled separately from published papers.
Contrastive On-Policy Distillation
Contrasts low- and high-effort teacher distributions to train multimodal models toward concise, efficient reasoning.
MMGist: A Comprehensive Multimodal Benchmark for 2027
Curates 7,262 visually grounded, discriminative questions from 18 benchmarks for reliable multimodal evaluation.
ExFusion: Efficient Transformer Training via Multi-Experts Fusion
Uses multiple experts during training and fuses them into a dense Transformer with no extra inference architecture.
MME-SCI: A Comprehensive and Challenging Science Benchmark for Multimodal Large Language Models
Evaluates scientific reasoning across four subjects, five languages, and three input modes with 1,019 annotated questions.
VLRMBench: A Comprehensive and Challenging Benchmark for Vision-Language Reward Models
Tests vision-language reward models on process understanding, outcome judgment, and critique generation across 12 tasks.
MPI-CD: Multi-Path Information Contrastive Decoding for Mitigating Hallucinations in Large Vision-Language Models
Combines contrastive visual branches and hidden-layer memory recall to reduce hallucinations without extra training.
TTE: Two Tokens Are Enough to Improve Parameter-Efficient Tuning
Regularizes parameter-efficient tuning with two learnable tokens to reduce overfitting on limited downstream data.
MM-CamObj: A Comprehensive Multimodal Dataset for Camouflaged Object Scenarios
Builds a multimodal dataset and evaluation suite for understanding camouflaged objects in challenging visual scenes.
More publications & collaborations
MEMO: Multimodal Evidence Memory Organization for Long-Horizon LLM Agents
Selects and organizes textual and visual evidence to support long-horizon agents under a memory budget.
Edit2TikZ: A Comprehensive and Challenging Benchmark for Scientific Figure Editing with TikZ
Benchmarks instruction-guided scientific figure editing through compilable TikZ code on 1,548 annotated samples.
VM-UNet: Vision Mamba UNet for Medical Image Segmentation
Uses visual state-space models in a U-Net architecture for medical image segmentation.
GIST: Improving Parameter Efficient Fine-Tuning via Knowledge Interaction
Improves parameter-efficient adaptation through interaction between pretrained and downstream knowledge.
LLaMA-MoE: Building Mixture-of-Experts from LLaMA with Continual Pre-Training
Builds sparse mixture-of-experts models from dense LLaMA checkpoints with continual pretraining.
iDAT: Inverse Distillation Adapter-Tuning
Transfers knowledge from an auxiliary model to improve adapter tuning.
LAMM: Label Alignment for Multi-Modal Prompt Learning
Learns label representations to improve alignment in multimodal prompt learning.
EGE-UNet: An Efficient Group Enhanced UNet for Skin Lesion Segmentation
Combines group-enhanced feature aggregation with a lightweight architecture for skin lesion segmentation.
MALUNet: A Multi-Attention and Light-weight UNet for Skin Lesion Segmentation
Uses lightweight multi-attention modules for efficient skin lesion segmentation.
04
Experience & education
Research internships
Qwen Team, Alibaba Group
Post-training of multimodal large language models.
ByteDance
Foundation models for video safety moderation and risk-frame selection.
Xiaohongshu
Multimodal understanding and multimodal reward models.
Institute for Advanced Algorithms Research
Multimodal large language models for camouflaged object scenarios.
Shanghai AI Laboratory
Mixture-of-Experts architectures and efficient Transformer pre-training.
Education
Shanghai Jiao Tong University
Ph.D. candidate in Computer Science and Technology
Wu Wenjun AI Honors Ph.D. ClassXidian University
Bachelor’s degree · School of Microelectronics
Outstanding Graduate of Shaanxi Province
Teaching
Teaching Assistant · SJTU
Integrated Circuit Design Practice in Microelectronics
Teaching Assistant · SJTU
Integrated Hardware Practice in Artificial Intelligence
05
Honors & academic service
Selected honors
- National Scholarship · Ph.D.
- IEEE TMI Certificate of Excellence · Bronze
- Yang Jiachi Scholarship
- Outstanding Graduate of Shaanxi Province
- National Scholarship · Undergraduate
- Xiaomi Scholarship
- National Scholarship · Undergraduate
Peer review
Conferences
ICLR 2025–2027 · NeurIPS 2024–2025 · ICML 2025 · AAAI 2026 · AISTATS 2026 · ECCV 2026 · ICCV 2025 · ACM MM 2024–2026 · ACL ARR 2025 · MICCAI 2023–2025 · ICASSP 2023–2025 · ICME 2024
Journals
IEEE TMI · IEEE TNNLS · IEEE TIP · IEEE TCSVT · IEEE JBHI · IJCV · Neurocomputing · Expert Systems with Applications · Engineering Applications of Artificial Intelligence
Program committee
IJCNN 2024 · Technical Program Committee
06
A little more about me
I collaborate with Suncheng Xiang, Mingye Xie, and Jingsheng Gao at SJTU. My research experiences have also included working with Yu Cheng and Xiaoye Qu at Shanghai AI Laboratory, Zhiyu Li at the Institute for Advanced Algorithms Research, Ye Guo at Xiaohongshu, and De Cai, Fubang Zhao, and Yangyang Kang at ByteDance.
Outside research, I enjoy swimming, bodyweight training, music, and playing the guitar.