PhD @ UC Berkeley EECS  ·  Research Scientist Intern @ Meta

Zhiyu Ni 倪志宇

I teach neural networks to reason with logic — and make them harder to fool.

⊢
scroll
01

About

Portrait of Zhiyu Ni
Berkeley, CA 37.8716° N, 122.2727° W

Hi! I'm a CS PhD student at UC Berkeley, advised by Prof. Pierluigi Nuzzo. I work where neural intuition meets symbolic rigor: teaching language models to reason in formal logic, so their conclusions are proven, not guessed. On the security side, I probe the safeguards we build around these models until they break.

Before Berkeley I spent two and a half years as a PhD student at the University of Southern California (2022–2024), and earned my B.S. in Physics from the University of Science and Technology of China as an Outstanding Graduate (top 3%). This year I'm a Research Scientist Intern at Meta.

If you're interested in reasoning, formal verification, or AI security, reach out — I'm always happy to collaborate.

Now
PhD, CS · UC Berkeley
Focus
Reasoning · Formal methods · Security
Stack
PyTorch · vLLM · SGLang · Z3
Before
PhD @ USC · B.S. Physics @ USTC
02

Research

Three threads, one question: can we trust what a model concludes?

∀∃

Auto-formalization & Logical Reasoning

Turning natural-language arguments into executable first-order logic, then letting a solver do the reasoning. Drafting many formalizations, pruning the contradictory ones, and keeping what verifies.

  • FOL
  • Z3
  • FOLIO
  • AR-LSAT
⊢

Agents with Guarantees

Agentic systems that come with something to check — assessor agents with structured failure tracking, and self-evolving contract frameworks for verified controller synthesis in cyber-physical systems.

  • Agents
  • Contracts
  • CPS
  • Verification
¬◊

AI Security & Robustness

Showing where defenses learn shortcuts instead of intent, building adversarially robust sparse GNNs, and secure cloud–device learning that keeps the informative features encrypted.

  • Prompt attacks
  • Adversarial GNNs
  • Privacy
03

Publications

* equal contribution  ·  ♦ equal advising  ·  click a paper for highlights

  1. An agentic framework where contracts evolve alongside the synthesized controller, so every generated controller comes with a specification it can be verified against.

  2. Controlled diagnostics reveal position, token-trigger, and topic-generalization shortcuts in fine-tuned prompt-attack defenses. High rejection rates often come from surface correlations, not from understanding harmful intent.

    • ↑50% false refusals
    • ↓40% accuracy under shift
  3. An inference-time framework that drafts diverse plans, prunes contradictory or ambiguous formalizations, and aggregates verified paths, with no extra supervision.

    • 78.43% AR-LSAT
    • 100% PrOntoQA
    • 100% LogicalDeduction
  4. A reproducible assessor-agent framework with structured failure tracking and standardized agent-to-agent interfaces, plus an auto-formalization agent that compiles natural-language logic into executable Z3Py under auditable budgets.

    • 86.70% cleaned FOLIO
    • vs 73.89% chain-of-thought
  5. A secure cloud–device framework that combines hypernetwork-based personalization with selective feature encryption. It matches full fine-tuning without backpropagation on the device, and dynamic feature sampling encrypts only the informative channels.

  6. GrokTransfer initializes a target model with embeddings learned by a weaker model. It provably generalizes directly on synthetic XOR and removes delayed generalization in MLPs and Transformers across tasks.

  7. ARGS jointly prunes perturbed graph edges and GNN weights with label- and homophily-aware objectives. The result is highly sparse GNNs with much stronger adversarial robustness and lower inference cost.

  8. The first systematic look at how sparse graph lottery tickets hold up under structural attacks, which motivated the ARGS defense.

04

Experience

  1. 2026.05 — now

    Research Scientist Intern · Meta

    Sunnyvale, CA · Meta AI

    Building domain foundation models for ads ranking.

  2. 2025.01 — now

    PhD Researcher · UC Berkeley EECS

    Berkeley, CA · Advisor: Prof. Pierluigi Nuzzo

    Neuro-symbolic reasoning, auto-formalization, and verifiable agents.

  3. 2025.05 — 2025.08

    AI Research Intern · Synopsys

    Sunnyvale, CA · DSO.AI

    Built an Analog Agent that identifies the function and structure of circuits automatically. A hybrid pipeline of GNNs and subgraph-isomorphism tests recovered 100% of non-overlapping blocks.

  4. 2022.08 — 2024.12

    PhD Researcher · USC DesCyPhy Lab

    Los Angeles, CA · Electrical & Computer Engineering

    Adversarial robustness of sparse graph neural networks (ICASSP 2024 Oral, TMLR 2024).

  5. 2022.02 — 2022.06

    NLP Research Intern · iFLYTEK

    Hefei, China

    Trained large multi-layer Transformers for bidirectional machine translation and beat Google Translate by 15% BLEU.

  6. 2020.09 — 2021.01

    Research Intern · University of Michigan

    Remote · Prof. Junjie Zhu

  7. 2018 — 2022

    B.S. Physics · USTC

    Hefei, China · Outstanding Graduate (top 3%)

05

News

06

Honors & Service

2025

DAC Young Student Fellow

Design Automation Conference

2024

ICASSP Oral Presentation

Adversarial vulnerabilities of graph lottery tickets

2022

Outstanding Graduate

USTC · top 3%

Reviewer
  • NeurIPS 2026
  • ICLR 2026
  • ICASSP 2026
  • ICLR 2025
  • ICASSP 2025
  • IJCNN 2025
07

Around the World

Every glowing dot is a city someone has visited this page from. Thank you for stopping by.

drag to spin

Open to collaborations

Let's build reasoning
that holds up.