publications

A selected list with short digests. An asterisk (*) denotes equal contribution; see my Google Scholar for the full list.

2026

  1. Preprint
    WiSP: A Working-Set View of Mixture-of-Experts Serving on Extremely Low-Resource Hardware
    Jiamu Zhang, Liang Wu, Mayank Darbari, and Liangjie Hong
    Preprint (work in progress), 2026
    TL;DREfficiencySystemsMoE

    Mixture-of-Experts models put most of their parameters in experts, yet only a few fire per token – and on commodity GPUs the usual layer-level CPU offload streams every expert across PCIe on each forward pass, throwing away MoE’s sparsity. WiSP treats low-resource MoE serving as a GPU working-set problem, paging in only the experts a workload actually reuses, and reaches up to 1.95x the decode throughput of static offload at the same memory budget with byte-identical outputs. A companion allocator, MV-WSA, splits VRAM between experts and the KV cache by equalizing marginal latency benefit per byte.

  2. Under Review
    Rethinking the Evaluation of Efficiency Methods for Multi-Agent Systems
    Jiamu Zhang*, Lingxi Zhang*, Pengjun Lu*, Qiyue Zhang, and 5 more authors
    Under review, 2026
    TL;DREvaluationMulti-AgentEfficiency

    Many methods claim to make LLM multi-agent systems (MAS) cheaper by pruning agents or communication edges – but do the gains really come from the method? Reported improvements are often measured under method-specific prompts and starting topologies, and many appear in settings where a single agent or even random pruning already does well. We introduce a controlled, MAS-demanding diagnostic benchmark that evaluates representative efficiency methods under shared model, agent, and runtime settings across topology, scale, depth, and tool use, and find that many reported gains are setup-dependent – arising from structural collapse, disabled tool pathways, or easy pruning rather than genuine efficiency improvements.

  3. Under Review
    Recognizing and Restructuring Latent Experts for Large Language Model Compression
    Jiamu Zhang, Alessandro Mason, Ning Xie, Aarav Swami, and 4 more authors
    Under review, 2026
    TL;DREfficiencyCompressionMoE

    Static compression like pruning permanently removes parameters deemed globally unimportant – but in LLMs, importance is highly input- and task-dependent, so static removal is brittle under distribution shift. Inspired by conditional computation in Mixture-of-Experts, MoRE (Mixture-of-Recognized-Experts) is a training-free paradigm that turns static redundancy in pretrained Transformers into input-dependent sparse computation. We find that pretrained MLP layers have a hybrid structure – a shared backbone that is always active plus input-dependent residual components that enable specialization – and MoRE exploits it through training-free routing, activating different subsets of pretrained parameters per input without modifying any weights. Across model families and benchmarks, MoRE matches or beats structured pruning on accuracy–efficiency trade-offs while staying robust under distribution shift and knowledge-intensive QA.

2025

  1. Flexible Group Count Enables Hassle-Free Structured Pruning
    Jiamu Zhang*, Shaochen Zhong*, Andrew Ye, Zirui Liu, and 8 more authors
    In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2025
    TL;DREfficiencyPruning

    Structured (grouped-kernel) pruning normally forces you to pre-commit to a rigid group count, making it painful to reuse across architectures. We make the group count flexible and hassle-free, so a single procedure prunes convolutional models cleanly while preserving the dense, hardware-friendly structure that delivers real speedups.

  2. Sweeping Promptable Spoofs under the DirtyRAG: A Practical, Query-Blind RAG Attack Done Right
    Shaochen Zhong*, Jiamu Zhang*, Hoang Anh Duy Le, Wenya Xie, and 15 more authors
    In ICLR 2026 Workshop on AI for Wild (AIWILD), Spotlight (under review), 2025
    TL;DRSecurityEvaluation

    RAG is a powerful attack surface – often one of the few places to reliably perform third-party prompt injection – yet existing attacks are query-dependent and impractical. DirtyRAG is a query-blind attack that works under realistic threat models: given an arbitrary benign passage and a 128-passage injection budget, it can, on average, fill the top-50+ retrieval results for any reasonable query with adversarial content. We also build an evaluation suite that better reflects realistic RAG threat models.

  3. Stop Overthinking: A Survey on Efficient Reasoning for Large Language Models
    Yang Sui, Yu-Neng Chuang, Guanchu Wang, Jiamu Zhang, and 7 more authors
    Transactions on Machine Learning Research (TMLR), 2025
    TL;DRSurveyEfficiencyLLM

    Long chain-of-thought reasoning boosts accuracy but burns a lot of test-time compute – the “overthinking” phenomenon, where a model may spend hundreds of tokens on a trivial question. This is the first structured survey of efficient reasoning for LLMs, organizing the fast-growing literature into model-, output-, and prompt-level directions. It has been widely covered by AI media (e.g., 新智元, 科技行者, and Towards AI) and reshared across X and RED, with citations climbing fast.

  4. ACL
    Redefining Language Model Evaluation with Ground-Truth-Free Ranking Frameworks
    Jiamu Zhang, Jiayi Yuan, Andrew Wen, Hoang Anh Duy Le, and 4 more authors
    In Findings of the Association for Computational Linguistics (ACL Findings), 2025
    TL;DREvaluationLLM

    Evaluating LLMs usually leans on ground-truth labels that are expensive and often unavailable. We rank models without any ground truth, using a reasoning-consistency framework that yields reliable, label-free rankings well correlated with traditional benchmarks.

  5. In-Progress: Structured Pruning in the Wild: Benchmarking Practical Robustness Under Real-World Corruptions
    Jiamu Zhang, Shaochen Zhong, Hoang Anh Duy Le, and Xia Hu
    In IEEE Security and Privacy Workshops (SPW), 2025
    TL;DRRobustnessEvaluation

    Structured-pruned models often look fine on clean data but degrade under real-world corruptions. We benchmark the practical robustness of pruned models under realistic corruptions, surfacing when and how pruning quietly trades away reliability.

  6. arXiv
    The Science of Evaluating Foundation Models
    Jiayi Yuan, Jiamu Zhang, Andrew Wen, and Xia Hu
    arXiv preprint arXiv:2502.09670, 2025
    TL;DREvaluation

    Foundation-model evaluation is fragmented and frequently ad hoc. This work lays out principles and a structured framework for evaluating foundation models rigorously – what to measure, how, and why – toward assessments that are trustworthy and comparable.

  7. Context Matching is not Reasoning: Assessing Generalized Evaluation of Generative Language Models in Clinical Settings
    Andrew Wen, Qiuhao Lu, Yu-Neng Chuang, Guanchu Wang, and 11 more authors
    npj Digital Medicine, 2025
    TL;DREvaluationHealthcare

    Generative LLMs can appear competent in clinical QA by matching surface context rather than genuinely reasoning. Through a generalized clinical evaluation, we show that strong benchmark numbers can mask a real lack of reasoning in medical settings.

2024

  1. GNNs Also Deserve Editing, and They Need It More Than Once
    Shaochen Zhong*, Hoang Anh Duy Le*, Zirui Liu, Zhimeng Jiang, and 9 more authors
    In International Conference on Machine Learning (ICML), 2024
    TL;DRRobustnessGraphs

    Model editing – fixing mistakes postmortem – is common everywhere except graphs, which are theoretically unfriendly to edits. We trace the root cause to over-accommodation of editing targets and present the first graph-editing method that lives up to real-life scrutiny of sequential edits.

2023

  1. One Less Reason for Filter Pruning: Gaining Free Adversarial Robustness with Structured Grouped Kernel Pruning
    Shaochen Zhong, Jiamu Zhang*, Zaichuan You*, Sebastian Zhao*, and 5 more authors
    In Advances in Neural Information Processing Systems (NeurIPS), 2023
    TL;DREfficiencyRobustness

    Pruned convolutional models can match clean accuracy yet collapse under perturbed inputs. We fix this essentially for free with a theoretically motivated grouped-kernel-pruning procedure that leverages kernel smoothness – a property known to promote frequency robustness – at no extra cost.