Searcharxiv⌕ Search

arXiv subjects

Parsa Ghazvinian

Publications and source records attributed to Parsa Ghazvinian.

3 recordsLinked to original sources

Answer or Defer? Testing VLMs' Stated Confidence on Brain MRI

Vision-language models (VLMs) could support medical-image interpretation, but their use requires deciding which interpretations to accept and which images to refer to a clinician. We test whether thresholding stated confidence can retain useful coverage while limiting errors among accepted answers, and how this depends on the population. Eight open-weight VLMs, including two reasoning models, assess 3,888 brain MRI slices from post-treatment glioma patients (BraTS 2024) and healthy volunteers (IXI) for tumor or gross abnormality. BraTS labels are mask-derived; IXI slices are treated as negatives. On BraTS, no threshold selects a nonempty set with at most 5% empirical error for any model, even when chosen retrospectively. This also holds for four models evaluated on a large-lesion, distant-negative subset. At a 10% error budget, Gemma-4-12B retains 24.4% of slices and every other model at most 2.5%. To investigate this limited selection ability, we separate confidence differences between predicted answers from differences within each answer. For four models with available outcome-level summaries, pooled mean confidence differs more between answers than between correct and incorrect responses within each answer. A covariance decomposition explains how this pattern can produce opposite confidence-correctness associations across populations. At a 20% error budget, accepting every "tumor" answer covers more slices than any confidence threshold for two models. These findings caution against using unvalidated stated confidence as a standalone deferral signal and support population-specific validation before deployment. Code is available at https://github.com/amir-sbg/VLM-Answer-or-Defer.

cs.CV↗

Authority Bias in Conversational Search Engines for Academic Paper Recommendation

Large Language Models (LLMs) are increasingly used as conversational search engines for academic literature, yet whether they judge papers on content or on authority signals has not been tested causally. We investigate authority bias: systematic preference for papers based on author prestige, venue, and citations rather than content. Holding title and abstract constant, we vary authority metadata across three counterfactual conditions (original, flipped, boosted) over eight LLMs (five open-weight and three frontier closed-weight) in an in-context, single-turn, top-1 recommendation setting. Our experiments show that authority bias is substantial and directional, varies markedly across models, and is only partially addressable through prompt-level debiasing. We further document a say-do gap: debiasing instructions suppress authority mentions far faster than authority-driven flips, so surface auditing systematically underestimates behavioral bias.

cs.AI↗

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

Due to the extensive application of machine learning (ML) in a wide range of fields and the necessity of data privacy, privacy-preserving machine learning (PPML) solutions have recently gained significant traction. One group of approaches relies on Homomorphic Encryption (HE), which enables us to perform ML tasks over encrypted data. However, even with state-of-the-art HE schemes, HE operations are still significantly slower compared to their plaintext counterparts and require a considerable amount of memory. Therefore, we propose MOFHEI, a framework that optimizes the model to make HE-based neural network inference, referred to as private inference (PI), fast and efficient. First, our proposed learning-based method automatically transforms a pre-trained ML model into its compatible version with HE operations, called the HE-friendly version. Then, our iterative block pruning method prunes the model's parameters in configurable block shapes in alignment with the data packing method. This allows us to drop a significant number of costly HE operations, thereby reducing the latency and memory consumption while maintaining the model's performance. We evaluate our framework through extensive experiments on different models using various datasets. Our method achieves up to 98% pruning ratio on LeNet, eliminating up to 93% of the required HE operations for performing PI, reducing latency and the required memory by factors of 9.63 and 4.04, respectively, with negligible accuracy loss.

cs.CR↗