About

I study how sequence models remember — and how to make that memory cheaper without changing what the model does.

I'm an M.S. student in Artificial Intelligence Engineering in the Department of Electrical and Computer Engineering at Carnegie Mellon University. Before CMU, I received a B.A. in Computer Science, with minors in Mathematics and Economics, from New York University.

My research sits between architecture and inference: efficient attention, linear recurrent and state-space models, and the key–value caches that long-context language models depend on — when they can be reused, compressed or restructured while keeping the model's behavior intact. I bring a background in AI theory to applied, measurable systems work. I write my own code end to end and release the code and data behind my papers.

Four of my sole-author papers were accepted in 2026, including AuthorityKV at the EACL 2027 main conference and ProofKV in Findings of EACL 2027.

Alongside language models, I have also collaborated on machine-learning-enhanced multi-criteria decision models for transport safety, published in Scientific Reports, Results in Engineering and Knowledge-Based Systems.

Seeking Summer 2027 internships in ML engineering and research (May–Aug 2027). Get in touch →

Efficient long-context inference

KV-cache reuse, compression and repair, and compact external memory for frozen models, in long-context and retrieval-augmented generation, measured against full recomputation.

Recurrent memory architectures

Gated linear attention, DeltaNet-style models and Mamba: how a fixed-size state allocates capacity, and where recall breaks down.

ML for decision support

Hybrid weighting, aggregation and clustering models for ranking and grouping countries on road-safety indicators.

News

  1. AuthorityKV was accepted to the EACL 2027 main conference and ProofKV to Findings of EACL 2027; code, data and camera-ready papers are public.

  2. SAFE-KV was published in Findings of IJCNLP-AACL 2026 (ACL Anthology).

  3. ASCENT, on compact external memory for frozen long-context language models, was published in Array.

  4. SAFE-KV was accepted to Findings of AACL-IJCNLP 2026.

  5. Released open reproduction packages for four transport-safety papers; each rebuilds the published results with a single command.

  6. Started the M.S. in Artificial Intelligence Engineering at Carnegie Mellon University.

  7. Graduated from New York University with a B.A. in Computer Science.

  8. First-author paper on a machine-learning-nested MCDM model published in Results in Engineering.

  9. Two papers on hybrid decision models for transport safety published in Scientific Reports.

Publications

Language models

EACL2027

AuthorityKV: Preventing Authority Reversals in KV-Cache Compression

Xingjian Zhang

Proceedings of EACL 2027 (main conference), to appear

Shows that KV-cache compression can silently break a prompt's instruction hierarchy: across three instruction-tuned models, three widely used compressors reverse 2,055 of 7,200 decisions that the uncompressed models get right. AuthorityKV, a training-free rule for any score-based compressor at an unchanged budget, removes all of these reversals with annotated span boundaries and 91% of them using only the chat roles and source provenance an orchestrator already holds; it raises system-prompt compliance on System-IFEval by 36–46 points and cuts successful prompt injections from 7 of 64 to 0.

EACL Findings2027

ProofKV: Compiling Complete Evidence Chains into Reusable KV Memory for Structured-Evidence QA

Xingjian Zhang

Findings of the Association for Computational Linguistics: EACL 2027, to appear

A multi-hop answer needs every link of an evidence chain, which token-by-token eviction and prompt compression do not keep together. ProofKV reads a document once into a reusable evidence index, composes evidence chains for each question with a small scorer and prefills the decoder only over the chains that fit a KV-byte budget. It is the best compact method in every setting across five structured-evidence QA tasks, two decoders and three token caps — +19.33 F1 over LLMLingua-2 and +13.32 F1 over same-scorer independent ranking at identical token counts — and serves 1.16–1.84× faster than full context with a reused index.

AACL Findings2026

SAFE-KV: Causal Suffix Repair for Behavior-Preserving Non-Prefix KV Cache Reuse in RAG

Xingjian Zhang

Findings of the Association for Computational Linguistics: IJCNLP-AACL 2026, pages 3063–3098

Treats non-prefix KV-cache reuse in retrieval-augmented generation as a behavior-preservation problem — full prefill is the reference and answers are compared by normalized short-answer equivalence — and repairs the causal suffix so that reused caches keep the model's outputs.

Array2026

ASCENT: Scale-Complementary External State for Frozen Long-Context Language Models

Xingjian Zhang

Array, 32, 101399 (2026)

A target-blind writer compresses a long input into a compact external state before the question arrives, and a frozen reader answers from that state. The state lifts five 7–8B readers by 67–86 points on 16K-token BABILong (Qwen2.5-7B: 33% → 100%) and needs 78–84× fewer reader FLOPs than full context (SmolLM2).

arXiv2025

Tracing LLM Reasoning Processes with Strategic Games: A Framework for Planning, Revision, and Resource-Constrained Decision Making

Xiaopeng Yuan, Xingjian Zhang, Ke Xu, Yifan Xu, Lijun Yu, Jindong Wang, Yushun Dong, Haohan Wang

Preprint, arXiv:2506.12012 (2025)

AdvGameBench: strategic games (tower defense, auto-battler, turn-based) that expose how LLMs plan, revise and act under resource budgets, with process metrics such as over-correction risk, correction success and over-budget rate related to win rate.

Machine learning for transport safety

Results Eng.2026

Machine learning nested MCDM model to enhance decision reliability for transport safety engineering

Xingjian Zhang, Nanbo (Aaron) Zhang, Jialin Li, …, Faan Chen

Results in Engineering, 29, 108543 (2026)

Nests PSI weighting, AROMAN aggregation and a Gaussian mixture model with t-SNE to rank and group the G20 countries on nine road-safety performance indicators across four waves from 2010 to 2023, with multilevel robustness analyses.

Sci. Rep.2025

A hybrid machine learning-enhanced MCDM model for transport safety engineering

Xingjian Zhang, Haowen Chen, Jingxuan Chen, …, Faan Chen

Scientific Reports, 15, 36467 (2025)

Combines DCRITIC weighting, WASPAS aggregation and k-means with a graph-based technique to assess transport safety in 35 member states of the Organization of American States on 15 indicators for 2010 and 2020.

Sci. Rep.2025

Road safety measurement with reliability using an advanced hybrid decision model

Jiaxu Jin, Hanrui Feng, Haojing Gao, …, Xingjian Zhang, …, Faan Chen

Scientific Reports, 15, 35099 (2025)

An HDVP–BeVarMax decision model that measures road safety with reliability across the 13 East Asia Summit countries on 18 indicators and four waves from 2012 to 2023.

KBS2024

An integrated MCDM model with enhanced decision support in transport safety using machine learning optimization

Zeyu Guo, Jiani Liu, Xiangchen Liu, …, Xingjian Zhang, …, Faan Chen

Knowledge-Based Systems, 301, 112286 (2024)

A LOPCOW–MULTIMOORA–DBSCAN model with grid search and k-means that evaluates transport safety in 30 European countries on 11 indicators for 2010 and 2020.

Research collaborations

  1. Efficient Sequence Modeling

    May 2025 – Jan 2026

    With Prof. Eunsol Choi and Yilun Kuang (NYU)

    • Designed SubMLR, a multi-level low-rank attention architecture between quadratic attention and constant-memory linear RNNs that beat strong baselines on Zoology MQAR.
    • Built chunk-compression models on linear attention (~10% higher accuracy at matched FLOPs) and a FLOPs-matched training and evaluation pipeline on Slurm with PyTorch and Triton.
  2. LLM Reasoning Benchmarks

    Feb 2025 – Jun 2025

    Collaboration with Prof. Haohan Wang (UIUC)

    • Co-developed AdvGameBench, a strategic-game benchmark of how LLMs plan and revise under resource budgets, and defined its process metrics.
  3. Chart Question Answering

    May 2025 – Jan 2026

    Collaboration with Prof. Qingyun Wang (William & Mary)

    • Added nine chart-type experts to the ChartMoE vision-language model for chart reasoning and built verified Program-of-Thought training data for SFT and DPO/GRPO.
  4. Transport Safety Modeling

    Jan 2024 – Sep 2025

    Collaboration with Dr. Faan Chen (Harvard)

    • Built hybrid clustering and multi-criteria decision models for multi-country road-safety assessment, leading to four first- or co-first-author journal papers and four open reproduction packages.

Code

Each repository contains the code, the data it runs on, and scripts that regenerate the published results.

AuthorityKV Training-free authority-tier wrapper for NVIDIA KVPress, the IH-CacheBench benchmark and the records behind every reported number; a CPU quickstart shows the effect in one command. Dataset on the Hugging Face Hub. PythonApache-2.0EACL 2027 ProofKV Evidence index, chain scorer, compact prefill and vLLM paged-KV serving route, with per-example result records, scorer checkpoints and a table-by-table reproduction guide. PythonMITEACL 2027 Findings safe-kv Package, configs, processed data and per-example results for SAFE-KV. Regenerates all 33 tables and 5 figures from shipped results on CPU; 135 unit tests. PythonMITAACL 2026 Findings ascent Code, frozen configurations and per-row records for ASCENT. One command re-verifies the result manifests, re-renders every manuscript table and prints every reported number on CPU in about ten seconds. PythonApache-2.0Array 2026 transport-safety-psi-aroman-gmm Single-command pipeline for the G20 case study: decision matrices for four waves, PSI–AROMAN rankings, robustness analyses and GMM + t-SNE grouping. PythonMITResults Eng. 2026 transport-safety-mcdm DCRITIC–WASPAS–k-means for 35 OAS countries. Parses the WHO source reports through to all 9 tables and 13 figures, offline in under two minutes. PythonMITSci. Rep. 2025 road-safety-hdvp-bevarmax HDVP–BeVarMax for 13 East Asia Summit countries on 18 indicators and four waves, with the robustness and benchmarking analyses. PythonMITSci. Rep. 2025 road-safety-lopcow-multimoora-dbscan LOPCOW–MULTIMOORA–DBSCAN with grid search and k-means for 30 European countries on 11 indicators, 2010 and 2020. PythonMITKBS 2024

Education

  1. Carnegie Mellon University

    M.S. in Artificial Intelligence Engineering, Electrical and Computer Engineering

    Pittsburgh, PA

    2026 – 2027 (expected)
  2. New York University

    B.A. in Computer Science · Minors in Mathematics and Economics

    New York, NY

    May 2026

Contact

I'm seeking Summer 2027 internships (May–Aug 2027) in ML engineering, applied science and research. Email is the fastest way to reach me.