Can Jin

CBIM, Busch Campus can.jin@rutgers.edu

I am a Ph.D. Candidate in Computer Science at Rutgers University, New Brunswick, advised by Professor Dimitris N. Metaxas. My research interests include Efficient Pre-training/Post-training/Inference of Foundation Models.

I earned my M.S. and B.S. degrees in Mathematics from the University of Science and Technology of China (USTC). Before my doctoral studies, I worked as a Machine Learning Engineer at Meituan Dianping Corporation. Recently, as a Research Intern at Adobe Research, I focused on the efficient pre-training of foundation models (LLMs/DiTs/omni-models) via Dense/MoE architectures.

I am actively seeking research internship opportunities starting from 2027, focusing on Efficient Pre-training/Post-training/Inference of Foundation Models. You can find my CV here.

Selected Publications

My research is organized around three core stages of foundation model development: pre-training, post-training, and inference. Across these directions, I work on making large models more capable, efficient, and reliable for reasoning, coding, alignment etc.

Pre-training

Focus: Investigating foundation model pre-training through efficient architecture design, generalizable training strategies, and scaling laws.

ICML
DTop-p MoE: Sparsity-Controlled Dynamic Top-p MoE for Foundation Model Pre-training

Can Jin^*, Hongwu Peng^*, Mingcan Xiang, Qixin Zhang, Xiangchi Yuan, Amit Hasan, Ohiremen Dibua, Yifan Gong, Yan Kang^†, and Dimitris N. Metaxas^†

In Forty-third International Conference on Machine Learning, 2026

Propose DTop-p MoE, a dynamic routing mechanism that utilizes a Proportional-Integral controller and dynamic routing normalization to precisely control expert activation sparsity while adapting to varying token difficulty. DTop-p outperforms Top-k and Top-p MoE across Large Language Models and Diffusion Transformers.

PDF
@inproceedings{jin2025dtopp, title = {DTop-p MoE: Sparsity-Controlled Dynamic Top-p MoE for Foundation Model Pre-training}, author = {Jin, Can and Peng, Hongwu and Xiang, Mingcan and Zhang, Qixin and Yuan, Xiangchi and Hasan, Amit and Dibua, Ohiremen and Gong, Yifan and Kang, Yan and Metaxas, Dimitris N.}, booktitle = {Forty-third International Conference on Machine Learning}, year = {2026}, }

NeurIPS

Learning from Teaching Regularization: Generalizable Correlations Should be Easy to Imitate

Can Jin^*, Tong Che^*, Hongwu Peng^†, Yiyuan Li^†, Dimitris Metaxas^‡, and Marco Pavone^‡

In Advances in Neural Information Processing Systems, 2024

Develop Learning from Teaching (LoT), a novel regularization technique for deep neural networks that enhances model generalization by training a teacher model to prioritize features that are easier for a student model to imitate, thereby filtering out spurious correlations.

PDF Code Poster Website

@inproceedings{jin2024lot,
  author = {Jin, Can and Che, Tong and Peng, Hongwu and Li, Yiyuan and Metaxas, Dimitris and Pavone, Marco},
  booktitle = {Advances in Neural Information Processing Systems},
  editor = {Globerson, A. and Mackey, L. and Belgrave, D. and Fan, A. and Paquet, U. and Tomczak, J. and Zhang, C.},
  pages = {966--994},
  publisher = {Curran Associates, Inc.},
  title = {Learning from Teaching Regularization: Generalizable Correlations Should be Easy to Imitate},
  url = {https://proceedings.neurips.cc/paper_files/paper/2024/file/01ce1ae7f94d139e4917f9e4425a4f38-Paper-Conference.pdf},
  volume = {37},
  year = {2024},
}

Post-training

Focus: Investigating post-training techniques such as reinforcement learning, on-policy distillation, supervised fine-tuning, prompt-based adaptation, and pruning for efficient reasoning, coding, alignment, and adaptation.

EMNLP
Act More, Decide Less: Skill-Guided Adaptive Action Chunking for Long-Horizon LLM Agents

Yanting Yang^*, Can Jin^* , Jinman Zhao, Jiahao Wu, Yang Zhou , Zhepeng Wang , Zhendong Wang , Mu Zhou, and Dimitris N. Metaxas

In Findings of the Association for Computational Linguistics: EMNLP 2026, 2026

Propose SPACE, a skill-guided adaptive action-chunking framework with hybrid on-/off-policy training and chunk-aware credit assignment, improving success rates by 7.0%–31.3% over the strongest baselines on ALFWorld and ScienceWorld while reducing LLM decision rounds by up to 78.9%.

PDF
@inproceedings{yang2026act, title = {Act More, Decide Less: Skill-Guided Adaptive Action Chunking for Long-Horizon LLM Agents}, author = {Yang, Yanting and Jin, Can and Zhao, Jinman and Wu, Jiahao and Zhou, Yang and Wang, Zhepeng and Wang, Zhendong and Zhou, Mu and Metaxas, Dimitris N.}, booktitle = {Findings of the Association for Computational Linguistics: EMNLP 2026}, year = {2026}, url = {https://arxiv.org/abs/2609.02042}, }
arXiv
DARE: Difficulty-Adaptive Reinforcement Learning with Co-Evolved Difficulty Estimation

Yang Zhou^*, Can Jin^*, Zihan Dong , Zhepeng Wang, Yanting Yang, Shiyu Zhao , Lei Li, Runxue Bao, Yaochen Xie, and Dimitris N. Metaxas

2026

Introduce DARE, a difficulty-adaptive Reinforcement Learning framework that co-evolves policy-aligned difficulty estimation with dynamic data selection and difficulty-specific optimization, improving training efficiency, final accuracy, and inference-token efficiency for LLM reasoning.

PDF Code
@misc{zhou2026dare, title = {DARE: Difficulty-Adaptive Reinforcement Learning with Co-Evolved Difficulty Estimation}, author = {Zhou, Yang and Jin, Can and Dong, Zihan and Wang, Zhepeng and Yang, Yanting and Zhao, Shiyu and Li, Lei and Bao, Runxue and Xie, Yaochen and Metaxas, Dimitris N.}, year = {2026}, journal = {arXiv preprint arXiv:2605.09188}, }
ICML · AI4Math
Weak Critics Make Strong Learners: On-Policy Critique Distillation for Scalable Oversight

Can Jin^* , Jiakang Li^*, Rui Wu , Eddy Z. Zhang^†, and Dimitris N. Metaxas^†

In 3rd AI for Math Workshop: Toward Self-Evolving Scientific Agents, 2026

We propose weak-critic strong oversight and on-policy critique distillation (OPCD), showing that weak models can guide stronger models through useful critiques while improving alignment and reasoning performance.

PDF Code
@inproceedings{jin2026weak, title = {Weak Critics Make Strong Learners: On-Policy Critique Distillation for Scalable Oversight}, author = {Jin, Can and Li, Jiakang and Wu, Rui and Zhang, Eddy Z. and Metaxas, Dimitris N.}, booktitle = {3rd AI for Math Workshop: Toward Self-Evolving Scientific Agents}, year = {2026}, url = {https://openreview.net/forum?id=oEfedgUChS}, }
ACL
Reasoning over Precedents Alongside Statutes: Case-Augmented Deliberative Alignment for LLM Safety

Can Jin^*, Rui Wu^*, Tong Che^*, Qixin Zhang, Hongwu Peng , Jiahui Zhao, Zhenting Wang, Wenqi Wei, Ligong Han , Zhao Zhang, Yuan Cao, Ruixiang Tang^†, and Dimitris N. Metaxas^†

In The 64th Annual Meeting of the Association for Computational Linguistics, 2026

Propose CADA, a case-augmented deliberative alignment framework that leverages reinforcement learning on self-generated reasoning chains to transition from rigid rule enforcement to flexible case-based reasoning, significantly reducing over-refusal while enhancing robustness against jailbreak attacks.

PDF
@inproceedings{jin2026reasoning, title = {Reasoning over Precedents Alongside Statutes: Case-Augmented Deliberative Alignment for LLM Safety}, author = {Jin, Can and Wu, Rui and Che, Tong and Zhang, Qixin and Peng, Hongwu and Zhao, Jiahui and Wang, Zhenting and Wei, Wenqi and Han, Ligong and Zhang, Zhao and Cao, Yuan and Tang, Ruixiang and Metaxas, Dimitris N.}, booktitle = {The 64th Annual Meeting of the Association for Computational Linguistics}, year = {2026}, }
ICLR
LoR-VP: Low-Rank Visual Prompting for Efficient Vision Model Adaptation

Can Jin , Ying Li , Mingyu Zhao, Shiyu Zhao, Zhenting Wang, Xiaoxiao He, Ligong Han, Tong Che, and Dimitris N. Metaxas

In The Thirteenth International Conference on Learning Representations, 2025

Design LoR-VP, a low-rank visual prompting technique for efficient vision model adaptation that reduces trainable parameters while outperforming full fine-tuning and standard visual prompting methods on object detection and segmentation benchmarks.

PDF Code
@inproceedings{jin2025lorvp, title = {LoR-{VP}: Low-Rank Visual Prompting for Efficient Vision Model Adaptation}, author = {Jin, Can and Li, Ying and Zhao, Mingyu and Zhao, Shiyu and Wang, Zhenting and He, Xiaoxiao and Han, Ligong and Che, Tong and Metaxas, Dimitris N.}, booktitle = {The Thirteenth International Conference on Learning Representations}, year = {2025}, url = {https://openreview.net/forum?id=5btFIv2PNb}, }

AAAI

Visual Prompting Upgrades Neural Network Sparsification: A Data-Model Perspective

Can Jin^*, Tianjin Huang^* , Yihua Zhang, Mykola Pechenizkiy, Sijia Liu , Shiwei Liu, and Tianlong Chen

In Proceedings of the AAAI Conference on Artificial Intelligence, 2025

Propose VPNs, a novel data-model co-design framework that simultaneously optimizes visual prompts and network sparsity, significantly enhancing the performance and transferability of sparse vision models.

PDF Code

@inproceedings{jin2025visual,
  title = {Visual Prompting Upgrades Neural Network Sparsification: A Data-Model Perspective},
  author = {Jin, Can and Huang, Tianjin and Zhang, Yihua and Pechenizkiy, Mykola and Liu, Sijia and Liu, Shiwei and Chen, Tianlong},
  booktitle = {Proceedings of the AAAI Conference on Artificial Intelligence},
  volume = {39},
  number = {4},
  pages = {4111--4119},
  year = {2025},
  url = {https://arxiv.org/pdf/2312.01397},
}

Inference

Focus: Investigating inference-time techniques such as test-time search, refinement/critiquing, prompt engineering, and multi-agent systems for improving reasoning, coding, retrieval, and agentic performance.

EMNLP
Two Heads are Better Than One: Test-time Scaling of Multi-agent Collaborative Reasoning

Can Jin, Hongwu Peng, Qixin Zhang, Yang Zhou , Yujin Tang, Tong Che^†, and Dimitris N. Metaxas^†

In Findings of the Association for Computational Linguistics: EMNLP 2026, 2026

Develop MAS-TTS, a framework that integrates a specialized multi-agent training pipeline with an adaptive CEO agent to orchestrate collaborative reasoning, effectively optimizing test-time scaling for complex tasks.

PDF Code
@inproceedings{jin2025two, title = {Two Heads are Better Than One: Test-time Scaling of Multi-agent Collaborative Reasoning}, author = {Jin, Can and Peng, Hongwu and Zhang, Qixin and Zhou, Yang and Tang, Yujin and Che, Tong and Metaxas, Dimitris N.}, booktitle = {Findings of the Association for Computational Linguistics: EMNLP 2026}, year = {2026}, url = {https://arxiv.org/abs/2504.09772}, }
EMNLP
Latent Reward Steering: An Adaptive Inference-Time Framework that Implicitly Promotes Cognitive Behaviors in Reasoning LLMs

Jiakang Li^*, Guanyu Zhu^*, Can Jin^* , Chenxi Huang, Dexu Yu , Ronghao Chen, Yang Zhou, Hongwu Peng, Xuanqi Lan, Dimitris N. Metaxas^† , and Youhua Li^†

In Findings of the Association for Computational Linguistics: EMNLP 2026, 2026

Introduce LRS, an adaptive inference-time framework that learns reward gradients from correctness-labeled reasoning traces to selectively steer fragile SAE latent states, consistently improving reasoning across multiple LLM backbones and benchmarks.

PDF Code
@inproceedings{li2026latent, title = {Latent Reward Steering: An Adaptive Inference-Time Framework that Implicitly Promotes Cognitive Behaviors in Reasoning LLMs}, author = {Li, Jiakang and Zhu, Guanyu and Jin, Can and Huang, Chenxi and Yu, Dexu and Chen, Ronghao and Zhou, Yang and Peng, Hongwu and Lan, Xuanqi and Metaxas, Dimitris N. and Li, Youhua}, booktitle = {Findings of the Association for Computational Linguistics: EMNLP 2026}, year = {2026}, url = {https://arxiv.org/abs/2606.00726}, }

WWW · RelWeb

APEER: Automatic Prompt Engineering Enhances Large Language Model Reranking (Best Paper Award @ RelWeb)

Can Jin^*, Hongwu Peng^*, Shiyu Zhao, Zhenting Wang, Wujiang Xu, Ligong Han , Jiahui Zhao, Kai Zhong, Sanguthevar Rajasekaran, and Dimitris N Metaxas

In Companion Proceedings of the ACM Web Conference 2025, Sydney, NSW, Australia, 2025

Propose APEER, a novel automatic prompt engineering algorithm that iteratively generates and refines prompts to enhance the performance and transferability of Large Language Models in information retrieval reranking tasks.

DOI PDF Code

@inproceedings{jin2025apeer,
  title = {APEER: Automatic Prompt Engineering Enhances Large Language Model Reranking (<span class="award">Best Paper Award @ RelWeb</span>)},
  author = {Jin, Can and Peng, Hongwu and Zhao, Shiyu and Wang, Zhenting and Xu, Wujiang and Han, Ligong and Zhao, Jiahui and Zhong, Kai and Rajasekaran, Sanguthevar and Metaxas, Dimitris N},
  booktitle = {Companion Proceedings of the ACM Web Conference 2025},
  location = {Sydney, NSW, Australia},
  series = {WWW '25},
  year = {2025},
  doi = {10.1145/3701716.3717574},
  isbn = {979-8-4007-1331-6/2025/04},
  url = {https://arxiv.org/abs/2406.14449},
  publisher = {Association for Computing Machinery},
  address = {New York, NY, USA},
  keywords = {Prompt engineering, Information Retrieval, Large Language Model, ReRanking},
}

Academic Services

Teaching Assistant

Rutgers University: CS344: Algorithms (Spring 2026), CS211: Computer Architecture (Fall 2025), CS534: Computer Vision (Spring 2025), CS210: Data Management for Data Science (Fall 2024)

Peer Review

Conference: NeurIPS 25/26, ICLR 25/26, ICML 24/26, CVPR 25/26, ECCV 26, AAAI 26/27, etc.
Journal: Alexandria Engineering Journal, Information Fusion, Pattern Recognition, Signal Processing