Graduate Researcher · University of Waterloo

Liliana Hotsko
LLM systems & research

I'm a machine-learning researcher building hypernetworks that compile knowledge straight into a model's weights — turning a whole code repository, or a plain-English description, into a lightweight LoRA adapter for a frozen language model, with zero added cost at inference time.

My work sits at the intersection of LLMs for software engineering, parameter-efficient fine-tuning, and privacy-preserving NLP — and I care most about systems that are efficient enough to run on a laptop, fully offline.

Waterloo, ON, Canada · open to research collaborations
Selected work Email GitHub LinkedIn Scholar X / Twitter
Portrait of Liliana Hotsko
01

About

I'm a graduate Computer Science researcher at the University of Waterloo, advised by Yuntian Deng and Pengyu Nie. I earned my BSc at the Ukrainian Catholic University, and along the way have held research roles at NTU, NYU Tandon (with the University of Oxford), and industry R&D at ELEKS.

My research begins with one question: rather than feeding models ever-longer prompts, can we compile the knowledge they need directly into their weights? In Code2LoRA, a hypernetwork reads a repository's commit history and generates a task-specific LoRA adapter for a frozen code model. In Program-as-Weights, a compiler turns a natural-language specification into a small neural program that a lightweight interpreter runs locally.

Across both, I aim for work that is reproducible, efficient, and private — models a developer can actually download and run on their own machine, fully offline.

Hypernetworks LoRA / PEFT LLMs for code Privacy-preserving NLP On-device inference Agent systems

MSc, University of Waterloo

Computer Science · since Sep 2025

Research Fellow, Waterloo

Hypernetworks & PEFT · since May 2025

2 papers on Hugging Face

Code2LoRA & PAW — Paper of the Day

02

News

Soon
Code2LoRA package release coming up
pip install code2lora — the c2l-terminal on-device assistant lands on PyPI.
Sep 2026
Program-as-Weights accepted at NeurIPS 2026 NeurIPS
PAW accepted to the 40th Conference on Neural Information Processing Systems · programasweights.com
Jun–Aug 2026
Research Fellow at NTU
Global Connect Fellowship · testing LLM API gateways · poster.
Jul 2026
PAW — Paper of the Day on Hugging Face 🤗 Paper of the Day
Program-as-Weights featured on the HF daily papers · huggingface.co/papers/2607.02512
Jul 2026
Program-as-Weights archived
Preprint, code, and public demo released at programasweights.com.
Jun 2026
Code2LoRA — Paper of the Day on Hugging Face 🤗 Paper of the Day
Featured on the HF daily papers · huggingface.co/papers/2606.06492
Jun 2026
Code2LoRA archived
Preprint, checkpoints, and RepoPeftBench released · project page.
Sep 2025
Started MSc at the University of Waterloo
Graduate studies in Computer Science, advised by Yuntian Deng.
Jun 2025
Graduated from Ukrainian Catholic University
BSc in Computer Science, Lviv, Ukraine.
May 2025
Started research fellowship at the University of Waterloo
Joined the group working on hypernetwork-generated adapters.
Aug 2023 – Apr 2025
R&D Engineer at ELEKS
Research & development in applied machine learning and NLP · eleks.com
Sep–Dec 2023
Research Fellow at NYU (Tandon School of Engineering)
In collaboration with the University of Oxford.
03

Selected work

commit history GRU state O(1) per commit diff Hypernetwork head emits LoRA · rank 16 Qwen2.5-Coder frozen · 1.5B ❄ + LoRA injected CPU-only generation — the base LLM is never loaded to build the adapter.
◆ Featured · First author

Code2LoRA

Hypernetwork-generated, repository-conditioned & commit-streaming LoRA adapters · University of Waterloo

A hypernetwork trained once across hundreds of repositories that maps a repo's commit history straight to a LoRA adapter for a frozen code LLM — zero inference-time token overhead, updated in O(1) per commit, and deployable on a laptop, fully offline. Ships as c2l-terminal, a Claude-Code-style on-device assistant.

63.8%
cross-repo EM
(vs 45.7% base)
78.9%+33.3
post-cutoff OOD
~37 MB
generated adapter
HypernetworkGRULoRA Qwen2.5-Coder-1.5BRepoPeftBench
Project page 🤗 Paper arXiv Models & data PyPI · soon
◆ Featured · Co-first author · NeurIPS 2026

Program-as-Weights (PAW)

A programming paradigm for fuzzy functions · Waterloo · Cornell · Harvard

Compile everyday "fuzzy" tasks — triaging logs, repairing JSON, ranking by intent — from a natural-language spec into a compact neural program. A 4B compiler emits a PEFT adapter for a frozen 0.6B interpreter that matches Qwen3-32B prompting at ~1/50th the memory, running at 30 tok/s on a MacBook M3, fully offline.

73.8%vs 68.7%
EM · 0.6B vs 32B prompt
~50×
less inference memory
10M
FuzzyBench examples
NeurIPS 2026Neural compilerText-to-LoRA Qwen3FuzzyBenchOn-device
Live demo 🤗 Paper arXiv GitHub
compile once · cloud "classify if this is urgent" 4B compiler spec → neural program program.lora · 23 MB run locally · offline 0.6B interpreter frozen ❄ · 30 tok/s + program injected → "urgent"
client request OpenAI schema Gateway translate ↔ Provider API metamorphic oracles conserve streamcheck paramaudit errormap Evidence + confidence gate fuzz seeds · roundtrip pairs · GitHub-calibrated bug taxonomy translation bug found Source-derived oracles target request/response rewrite mistakes in production gateways.
◆ NTU · Global Connect Fellowship

Finding Translation Bugs in LLM API Gateways

Metamorphic testing for unified LLM gateways · Nanyang Technological University, Singapore

LLM gateways rewrite requests and responses between provider schemas — and silent translation bugs can change model behavior downstream. We built an oracle pipeline with fuzz seeds, captured before/after pairs, and metamorphic invariants (conserve messages, streamcheck, paramaudit, errormap) calibrated to real GitHub issue distributions across major open-source gateways.

4
gateways
in study
6+
sub-oracle checks
GH
bug-calibrated
test mix
LiteLLMMetamorphic testingDifferential oracles FuzzingLLM APIs
Poster
◆ BSc thesis · Ukrainian Catholic University

LLM Anonymization Framework

Privacy-preserving RAG with pre-LLM entity masking · English & Ukrainian

A pipeline that anonymizes text before it reaches an LLM: transformer-based NER, LLM-assisted masking, encrypted vector storage, and retrieval-augmented generation — so sensitive documents can be queried without leaking identifiers in prompts or embeddings.

2
languages
EN + UK
NER
+ LLM masking
RAG
encrypted
vector search
PrivacyNERRAG EncryptionUkrainian NLP
GitHub Thesis
raw document Contact John Doe at [email protected] · Kyiv office… NER + LLM mask entities → tokens Encrypted index vector store Safe RAG query LLM sees redacted text Identifiers stripped upstream — retrieval and generation stay privacy-aware.
Natural language "open Spotify & dim lights" MacTell CLI Web Code Media One interpreter layer routes intent to macOS actions across modalities.
◆ UCU preprint · Open Interpreter lineage

MacTell — Open Interpreter

Natural-language control plane for macOS · Ukrainian Catholic University

A unified natural-language interface for macOS that plans and executes user intent across the command line, browser automation, code execution, and media control — one conversational layer instead of siloed tools.

4
action
surfaces
NL
→ plans → exec
macOS
native
automation
AgentsmacOSCLI Web automationCode execution
GitHub Preprint
◆ Explainable AI · UCU collaboration

Benchmarking Monotonic GNNs

Rule extraction & architecture comparison for monotonic graph networks

Monotonic GNNs promise interpretable structure — we extracted symbolic rules from trained models and benchmarked how graph architecture and extraction hyperparameters affect rule fidelity and coverage, connecting neural predictions to human-readable logic.

M-GNN
architectures
compared
XAI
rule
extraction
bench
quality vs
params
GNNMonotonicityRule extraction XAIBenchmarking
GitHub Report
monotonic GNN Rule extractor symbolic logic IF x₁ ↑ THEN y ↑ IF neighbor ≥ t … coverage · fidelity Compare architectures by the rules they admit — not just accuracy alone.
04

Get in touch

Let's build models that run anywhere.

I'm always happy to talk about hypernetworks, parameter-efficient adaptation, LLMs for code, or on-device inference. Reach out for collaborations, feedback, or a chat.

[email protected] GitHub Google Scholar X / Twitter LinkedIn Download CV