Parsa Ghazvinian

Research

My research sits at the intersection of efficient and trustworthy AI. I'm particularly interested in work that connects algorithmic ideas with systems-level considerations, so that methods hold up under real compute, data and privacy constraints.

Efficiency & optimization

I study methods that make modern AI systems cheaper to adapt and run, including parameter-efficient fine-tuning, pruning and sparsity, and efficient inference for large language and vision-language models. Much of this work asks a simple question: does a popular efficiency idea actually deliver what it promises once it is tested fairly?

Does learned rank allocation help LoRA?

Under review, ACL Rolling Review

Question
Methods such as AdaLoRA let importance scores decide which layers get more rank, on the assumption that this spends a fixed parameter budget better than giving every layer the same rank. That assumption had rarely been tested with matched controls.
Approach
A controlled study covering both standard and shared-parameter low-rank adapters, with every run paired with fixed-rank and random controls.
Finding
Learned rank allocation was never detectably better than a fixed rank at the same parameter budget. The choice of importance score mattered mainly when a trained adapter was compressed in one step.

How should we judge visual-token pruning?

Paper in preparation

Question
Pruning methods for vision-language models are usually compared against an “upper bound”: the best subset of image tokens found by searching for a known answer. What does that bound really measure?
Approach
Aim the same search at a wrong answer, then check the bounds with a held-out evaluation.
Finding
The search makes models confidently report objects that are not in the image, so these bounds reflect how easily a model can be steered rather than what it actually sees. They overstate the remaining headroom by about 30 points.

Papers

Does Learned Rank Allocation Help Shared and Standard Low-Rank Adapters? A Controlled Study at Equal Final Rank Budget

P. Ghazvinian, M. Dhakal, U. Jinadu, Y. Ding, R. Sunderraman

ACL Rolling Review Under review

Robust & trustworthy AI

My work in NLP and LLMs looks at how models behave under noisy, subjective or potentially biased conditions: learning from annotator disagreement, modeling subjective human judgments, and studying biases that emerge when AI systems recommend things to people.

Authority bias in LLM-based search engines

EMNLP 2026 (Main Conference)

Question
When LLMs act as conversational search engines for academic papers, do they judge a paper on its content or on who wrote it?
Approach
Keep each paper’s title and abstract fixed and change only its authority signals: author prestige, venue and citation counts.
Finding
Authority bias is substantial and varies widely across models. Prompt-level debiasing only partly helps: models stop mentioning authority much faster than they stop acting on it, so auditing their explanations underestimates the bias.

Learning from disagreement under noise

Findings of EMNLP 2026

Question
In subjective tasks such as hate-speech detection, annotators genuinely disagree, and majority voting erases minority perspectives. How well do models that learn from each annotator hold up when the labels are also noisy?
Approach
NRA-Embed adds a noise-robust contrastive objective to annotator embeddings.
Finding
Its performance is comparable to the state of the art on clean data, and it is more robust as noise increases.

Papers

Authority Bias in Conversational Search Engines for Academic Paper Recommendation

U. Jinadu, P. Ghazvinian, A. Budathoki, B. M. Ampel, R. Sunderraman, Y. Ding

EMNLP 2026 (Main Conference)

When Disagreement Meets Noise: Noise Robust Annotator Embeddings for Subjective NLP

K. Jacob, P. Ghazvinian, U. Jinadu, N. T. Yang, Y. Ding

Findings of EMNLP 2026

Privacy-preserving ML

Homomorphic encryption lets a model run on data it never sees in the clear, but encrypted inference is still far slower and more memory-hungry than ordinary inference. My work reduces that overhead so encrypted inference becomes practical.

MOFHEI

IEEE TPS 2024

Approach
A framework that converts pretrained models into a form suited to homomorphic encryption and prunes them in a way that matches how encrypted data is packed.
Result
Encrypted-inference latency cut by 9.6× and memory by 4× at negligible accuracy loss. The framework is open source.

Memory-efficient encrypted inference

ACNS 2024

Approach
A caching and dynamic weight-loading scheme that reuses recurring values within each layer.
Result
Peak memory cut by up to 50× for a 13% runtime cost, so encrypted inference can run on low-cost hardware.

Packing-aware pruning

IEEE S&P 2023 (extended abstract)

Approach
Pruning designed around how fully homomorphic encryption packs data, for faster encrypted inference.

Papers

MOFHEI: Model Optimization Framework for Fast and Efficient Homomorphically Encrypted Neural Network Inference

P. Ghazvinian, R. Podschwadt, M. GhasemiGol, D. Takabi

IEEE TPS 2024

Memory-Efficient Privacy-Preserving Machine Learning Based on Homomorphic Encryption

R. Podschwadt, P. Ghazvinian, M. GhasemiGol, D. Takabi

ACNS 2024

Packing-Aware Pruning for Efficient Private Inference Based on Fully Homomorphic Encryption

P. Ghazvinian, R. Podschwadt, M. GhasemiGol, D. Takabi

IEEE S&P 2023 (Extended Abstract)