Zirui Zhu

PhD candidate in Computer Science at NUS

I am a PhD candidate in Computer Science at the National University of Singapore, where I am fortunate to be advised by Prof. Yang You. Before that, I received my B.Eng. in Electronics Engineering from Tsinghua University in 2022, where I was grateful to work with Prof. Yong Li and Prof. Xu Chen.

Research Interests: I develop structure-aware, resource-efficient methods for foundation models. My research studies how to exploit heterogeneity in data, inputs, and optimization to allocate limited compute, tokens, supervision, and model updates where they matter most. Recent work spans budgeted long-video understanding, confidence-gated reward modeling, and efficient LLM post-training.

I welcome collaborations and discussions of all kinds—feel free to reach out anytime!

News

Jul 2026 Received the NUS Research Achievement Award!
May 2026 CAMEL was accepted to ICML 2026.

Experience

Research Intern, TikTok Pte. Ltd, Singapore Apr 2025 – Apr 2026
Research Intern, 01.AI, Beijing, China May 2024 – Jul 2024
Research Intern, Kuaishou Technology, Beijing, China Sep 2021 – Feb 2022
Remote Research Assistant, The Robotics Institute, CMU Jun 2021 – Mar 2022
Research Assistant, FIB-Lab, Tsinghua University, Beijing, China Apr 2019 – Sep 2020

Publications

Representative papers are highlighted.

CAMEL: Confidence-Gated Reflection for Reward Modeling

Zirui Zhu, Hailun Xu, Yang Luo, Yong Liu, Kanchan Sarkar, Kun Xu, Yang You

ICML 2026 paper

TL;DROne-token verdicts when the reward model is confident, reflection only when it is not: a 14B model that beats 70B judges with 82.9% average accuracy on three reward-model benchmarks.

              @inproceedings{zhu2026camel,
                title     = {{CAMEL}: Confidence-Gated Reflection for Reward Modeling},
                author    = {Zhu, Zirui and Xu, Hailun and Luo, Yang and Liu, Yong and Sarkar, Kanchan and Xu, Kun and You, Yang},
                booktitle = {Proceedings of the 43rd International Conference on Machine Learning (ICML)},
                year      = {2026},
                url       = {https://openreview.net/forum?id=gl5xoOJJ22}
              }

MatchLM2Lite: A Scalable MLLM-to-Lite Framework for Reproduced Content Identification

Xiaotian Fan, Hiok Hian Ong, David Yuchen Wang, Zirui Zhu, Kanchan Sarkar, Kun Xu

KDD 2026 paper

TL;DRDistills a multimodal LLM that spots reproduced videos into a 35× cheaper real-time model, keeping most of its accuracy gain and cutting reproduced-video views by 2.5% in production.

              @inproceedings{fan2026matchlm2lite,
                title     = {{MatchLM2Lite}: A Scalable {MLLM}-to-Lite Framework for Reproduced Content Identification},
                author    = {Fan, Xiaotian and Ong, Hiok Hian and Wang, David Yuchen and Zhu, Zirui and Sarkar, Kanchan and Xu, Kun},
                booktitle = {Proceedings of the 32nd ACM SIGKDD Conference on Knowledge Discovery and Data Mining V.2 (KDD)},
                pages     = {7244--7255},
                year      = {2026},
                doi       = {10.1145/3770855.3818444}
              }

MLT-Dedup: Efficient Large-Scale Online Video Deduplication via Multi-Level Representations and Spatial-Temporal Matching

David Yuchen Wang, Haoying Li, Hailun Xu, Wei Chee Yew, Zirui Zhu, Sanjay Saha, Hao Hei, Kanchan Sarkar, Kun Xu

KDD 2026 paper

TL;DRSparse clip-level embeddings for retrieval, fine-grained frame matching for precision: an online video deduplication system that cuts repetition by 91% at 90% precision with 5× the index capacity.

              @inproceedings{wang2026mltdedup,
                title     = {{MLT-Dedup}: Efficient Large-Scale Online Video Deduplication via Multi-Level Representations and Spatial-Temporal Matching},
                author    = {Wang, David Yuchen and Li, Haoying and Xu, Hailun and Yew, Wei Chee and Zhu, Zirui and Saha, Sanjay and Hei, Hao and Sarkar, Kanchan and Xu, Kun},
                booktitle = {Proceedings of the 32nd ACM SIGKDD Conference on Knowledge Discovery and Data Mining V.2 (KDD)},
                pages     = {8100--8111},
                year      = {2026},
                doi       = {10.1145/3770855.3818450}
              }

FOCUS: Efficient Keyframe Selection for Long Video Understanding

Zirui Zhu, Hailun Xu, Yang Luo, Yong Liu, Kanchan Sarkar, Zhenheng Yang, Yang You

ICLR 2026 paper code

TL;DRCasts keyframe selection as bandit pure exploration to find the frames that matter under a strict token budget: training-free, model-agnostic, and up to 11.9% more accurate on LongVideoBench while reading under 2% of frames.

              @inproceedings{zhu2026focus,
                title     = {{FOCUS}: Efficient Keyframe Selection for Long Video Understanding},
                author    = {Zhu, Zirui and Xu, Hailun and Luo, Yang and Liu, Yong and Sarkar, Kanchan and Yang, Zhenheng and You, Yang},
                booktitle = {The Fourteenth International Conference on Learning Representations (ICLR)},
                year      = {2026},
                url       = {https://openreview.net/forum?id=1OQKqLFcbB}
              }

Sparse MeZO: Less Parameters for Better Performance in Zeroth-Order LLM Fine-Tuning

Yong Liu, Zirui Zhu, Chaoyu Gong, Minhao Cheng, Cho-Jui Hsieh, Yang You

NeurIPS 2025 paper code

TL;DRZeroth-order gradient noise hurts most on large weights, so apply it only to a chosen sparse subset: better accuracy and 3.5× faster convergence than MeZO at inference-level memory, fine-tuning LLaMA-30B on a single A100.

              @inproceedings{liu2025sparsemezo,
                title     = {Sparse {MeZO}: Less Parameters for Better Performance in Zeroth-Order {LLM} Fine-Tuning},
                author    = {Liu, Yong and Zhu, Zirui and Gong, Chaoyu and Cheng, Minhao and Hsieh, Cho-Jui and You, Yang},
                booktitle = {Advances in Neural Information Processing Systems (NeurIPS)},
                volume    = {38},
                year      = {2025},
                url       = {https://openreview.net/forum?id=Tjw0ACu3NL}
              }

POME: Post Optimization Model Edit via Muon-style Projection

Yong Liu, Di Fu, Yang Luo, Zirui Zhu, Minhao Cheng, Cho-Jui Hsieh, Yang You

arXiv 2025 paper code

TL;DRA zero-cost post-processing step: project the fine-tuning weight delta with a Muon-style truncated SVD to even out dominant directions and prune noise, gaining 2.5% on GSM8K with no extra data or training.

              @misc{liu2025pome,
                title         = {{POME}: Post Optimization Model Edit via Muon-style Projection},
                author        = {Liu, Yong and Fu, Di and Luo, Yang and Zhu, Zirui and Cheng, Minhao and Hsieh, Cho-Jui and You, Yang},
                year          = {2025},
                eprint        = {2510.06627},
                archivePrefix = {arXiv},
                primaryClass  = {cs.LG},
                url           = {https://arxiv.org/abs/2510.06627}
              }

SeedFT: Structure-Preserving Fusion for Multi-Seed LLM Fine-Tuning

Yong Liu, Di Fu, Hailun Xu, Yang Luo, Zirui Zhu, Kanchan Sarkar, Zhenheng Yang, Minhao Cheng, Cho-Jui Hsieh, Yang You

Preprint 2025 paper

TL;DRDifferent seeds learn complementary sub-skills, so SeedFT merges seed-specific fine-tuned models training-free while preserving matrix geometry: align their updates in a shared SVD subspace, then fuse them with a closed-form orthogonality-constrained step, for relative gains of 5.7% on GSM8K and 8.3% on MATH.

              @misc{liu2025seedft,
                title        = {{SeedFT}: Structure-Preserving Fusion for Multi-Seed {LLM} Fine-Tuning},
                author       = {Liu, Yong and Fu, Di and Xu, Hailun and Luo, Yang and Zhu, Zirui and Sarkar, Kanchan and Yang, Zhenheng and Cheng, Minhao and Hsieh, Cho-Jui and You, Yang},
                year         = {2025},
                howpublished = {Preprint, OpenReview},
                url          = {https://openreview.net/forum?id=XPxfT3ttAi}
              }

SeedLoRA: A Fusion Approach to Efficient LLM Fine-Tuning

Yong Liu, Di Fu, Shenggan Cheng, Zirui Zhu, Yang Luo, Minhao Cheng, Cho-Jui Hsieh, Yang You

ICML 2025 paper

TL;DRLoRA runs of the same task from different random seeds have complementary strengths: fusing them closes much of the gap to full fine-tuning on math, code, and instruction following at LoRA’s memory cost.

              @inproceedings{liu2025seedlora,
                title     = {{SeedLoRA}: A Fusion Approach to Efficient {LLM} Fine-Tuning},
                author    = {Liu, Yong and Fu, Di and Cheng, Shenggan and Zhu, Zirui and Luo, Yang and Cheng, Minhao and Hsieh, Cho-Jui and You, Yang},
                booktitle = {Proceedings of the 42nd International Conference on Machine Learning (ICML)},
                series    = {Proceedings of Machine Learning Research},
                volume    = {267},
                pages     = {38384--38396},
                year      = {2025},
                url       = {https://proceedings.mlr.press/v267/liu25o.html}
              }

MERIT: Maximum-normalized Element-wise Ratio for Language Model Large-batch Training

Yang Luo, Zangwei Zheng, Ziheng Qin, Zirui Zhu, Yong Liu, Yang You

ICML 2025 paper code

TL;DRA large-batch optimizer that computes trust ratios with the max-norm to tame the max attention logit, plus element-wise ratios for finer scaling: GPT-2 Medium trains at a 6k batch size with no loss in quality.

              @inproceedings{luo2025merit,
                title     = {{MERIT}: Maximum-normalized Element-wise Ratio for Language Model Large-batch Training},
                author    = {Luo, Yang and Zheng, Zangwei and Qin, Ziheng and Zhu, Zirui and Liu, Yong and You, Yang},
                booktitle = {Proceedings of the 42nd International Conference on Machine Learning (ICML)},
                series    = {Proceedings of Machine Learning Research},
                volume    = {267},
                pages     = {41566--41585},
                year      = {2025},
                url       = {https://proceedings.mlr.press/v267/luo25u.html}
              }

How Does the Textual Information Affect the Retrieval of Multimodal In-Context Learning?

Yang Luo, Zangwei Zheng, Zirui Zhu, Yang You

EMNLP 2024 paper code

TL;DRRetrievers for multimodal in-context examples are surprisingly sensitive to which modality they see; MSIER, a supervised retriever trained on the MLLM’s own confidence, picks better examples across three tasks.

              @inproceedings{luo2024textual,
                title     = {How Does the Textual Information Affect the Retrieval of Multimodal In-Context Learning?},
                author    = {Luo, Yang and Zheng, Zangwei and Zhu, Zirui and You, Yang},
                booktitle = {Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing (EMNLP)},
                pages     = {5321--5335},
                year      = {2024},
                doi       = {10.18653/v1/2024.emnlp-main.305}
              }

Helen: Optimizing CTR Prediction Models with Frequency-wise Hessian Eigenvalue Regularization

Zirui Zhu, Yong Liu, Zangwei Zheng, Huifeng Guo, Yang You

WWW 2024 paper code

TL;DRIn CTR models the top Hessian eigenvalue tracks feature frequency, so frequent features land in sharp minima: Helen perturbs each feature in proportion to its frequency and beats standard optimizers on seven models and three datasets.

              @inproceedings{zhu2024helen,
                title     = {Helen: Optimizing {CTR} Prediction Models with Frequency-wise Hessian Eigenvalue Regularization},
                author    = {Zhu, Zirui and Liu, Yong and Zheng, Zangwei and Guo, Huifeng and You, Yang},
                booktitle = {Proceedings of the ACM Web Conference 2024 (WWW)},
                pages     = {3485--3496},
                year      = {2024},
                doi       = {10.1145/3589334.3645463}
              }

Grasp Stability Prediction with Sim-to-Real Transfer from Tactile Sensing

Zilin Si, Zirui Zhu, Arpit Agarwal, Stuart Anderson, Wenzhen Yuan

IROS 2022 paper code

TL;DRCouples a robot dynamics simulator with a physics-based tactile contact model, calibrates both on real data, and predicts grasp stability zero-shot on real objects with 90.7% accuracy.

              @inproceedings{si2022grasp,
                title     = {Grasp Stability Prediction with Sim-to-Real Transfer from Tactile Sensing},
                author    = {Si, Zilin and Zhu, Zirui and Agarwal, Arpit and Anderson, Stuart and Yuan, Wenzhen},
                booktitle = {2022 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS)},
                pages     = {7809--7816},
                year      = {2022},
                doi       = {10.1109/IROS47612.2022.9981863}
              }

Inhomogeneous Social Recommendation with Hypergraph Convolutional Networks

Zirui Zhu, Chen Gao, Xu Chen, Nian Li, Depeng Jin, Yong Li

ICDE 2022 paper code

TL;DRSocial influence is item-specific: friends may agree on some products and not others, so SHGCN models user–user–item triplets as hyperedges to learn finer-grained social signals for recommendation.

              @inproceedings{zhu2022inhomogeneous,
                title     = {Inhomogeneous Social Recommendation with Hypergraph Convolutional Networks},
                author    = {Zhu, Zirui and Gao, Chen and Chen, Xu and Li, Nian and Jin, Depeng and Li, Yong},
                booktitle = {2022 IEEE 38th International Conference on Data Engineering (ICDE)},
                year      = {2022},
                url       = {https://arxiv.org/abs/2111.03344}
              }

Beyond Research

Off the clock, I am an avid tennis player 🎾 with 12+ years on court. At Tsinghua, I captained the EE department's tennis team and led it to the Ma Yuehan Cup team championship, the university's top campus sports title. Always up for some rallies if you are around!