Technion – Israel Institute of Technology

7 related articles

Subscribe to updates

Journal of Mathematical Imaging and Vision Nov 10, 2026
Low-rank adaptation learns one matrix: SingLoRA beats LoRA and LoRA+ on LLaMA-7B at 60% of their parameter budget
Low-rank adaptation (LoRA) writes the weight update as a product of two small matrices, but their norms often differ sharply during training, producing imbalanced gradient flow and unstable optimization. SingLoRA instead writes the update as a single low-rank matrix times its own transpose, removing inter-matrix scale conflicts by construction and roughly halving the trainable parameters. Fine-tuning LLaMA-7B on MNLI, SingLoRA reaches 91.3% accuracy against 89.1% for LoRA and 90.2% for LoRA+, while using only 60% of their parameter budget.
Low-rank adaptation (LoRA) writes the weight update as a product of two small matrices, but their norms often differ sharply during training, producing imbalanced gradient flow and unstable optimization. SingLoRA instead writes the update as a single low-rank matrix times its own transpose, removing inter-matrix scale conflicts by construction and roughly halving the trainable parameters. Fine-tuning LLaMA-7B on MNLI, SingLoRA reaches 91.3% accuracy against 89.1% for LoRA and 90.2% for LoRA+, while using only 60% of their parameter budget.
Technion – Israel Institute of Technology
David Bensaïd（First author, Corresponding author）
SingLoRA Low-Rank Adaptation Parameter-Efficient Fine-Tuning Training Stability Image Generation
bioRxiv Oct 14, 2026
T-cell clonal organization without TCR sequencing: CloneTrast infers clonal niches from transcriptomes alone
CloneTrast uses clonal identities from paired single-cell RNA and TCR sequencing as training labels, so the model learns T-cell clonal-functional states from gene expression alone and needs no TCR data at inference. Training used 26 datasets across 14 cancer types with more than 1.5 million T cells, and external testing used 11 datasets with more than 1 million T cells, including two cancer types unseen in training. On both training and test sets, patient clonal purity improved significantly over a conventional expression embedding, and one NK-like niche enriched for tumor-blood shared TCRs was associated with poor immunotherapy response across studies.
CloneTrast uses clonal identities from paired single-cell RNA and TCR sequencing as training labels, so the model learns T-cell clonal-functional states from gene expression alone and needs no TCR data at inference. Training used 26 datasets across 14 cancer types with more than 1.5 million T cells, and external testing used 11 datasets with more than 1 million T cells, including two cancer types unseen in training. On both training and test sets, patient clonal purity improved significantly over a conventional expression embedding, and one NK-like niche enriched for tumor-blood shared TCRs was associated with poor immunotherapy response across studies.
Technion – Israel Institute of Technology
Ofir Shorer（First author）Keren Yizhak（Corresponding author）
CloneTrast Supervised Contrastive Learning T-cell clonal niches Single-Cell RNA Sequencing cancer immunotherapy
arXiv Oct 2, 2026
SAGO re-evaluates eleven LLMs by per-instance stability: no model generalizes uniformly, and cross-dataset variation can reverse rankings
The work introduces the Stability-Aware Generalization Objective (SAGO), which defines LLM generalization as per-instance behavioral stability of the same input under semantically equivalent variations, and evaluates eleven open- and closed-source models across six datasets on four behavioral axes—generation consistency, internal activations, confidence, and response mirroring—finding that every model exceeds a significance threshold on at least one axis, that the axes capture independent failure modes, and that cross-dataset variation can reverse model rankings.
The work introduces the Stability-Aware Generalization Objective (SAGO), which defines LLM generalization as per-instance behavioral stability of the same input under semantically equivalent variations, and evaluates eleven open- and closed-source models across six datasets on four behavioral axes—generation consistency, internal activations, confidence, and response mirroring—finding that every model exceeds a significance threshold on at least one axis, that the axes capture independent failure modes, and that cross-dataset variation can reverse model rankings.
Technion – Israel Institute of Technology
Nagham Omar（First author）
Stability-Aware Generalization Objective (SAGO)Stability Generalization Score Generalization in large language models (LLMs)
arXiv Oct 1, 2026
MIST stress test: adding an image shifts about a fifth of VLM judge labels, regardless of what the image shows
The authors built MIST (the Misleading-Image Stress Test), 200 English sentences each containing a phrase readable either figuratively or literally and shown with an aligned image, a misleading image, or no image; across thirteen VLM judges an aligned image changed 20.5% of labels and a misleading one 19.4%, versus 11.6% when only the ignore-the-image instruction was deleted with the image left in place, and only 37% of the labels that differ between the two images moved toward the sense shown, while agreement with the human annotators was unchanged whether the image was absent, aligned, or misleading.
The authors built MIST (the Misleading-Image Stress Test), 200 English sentences each containing a phrase readable either figuratively or literally and shown with an aligned image, a misleading image, or no image; across thirteen VLM judges an aligned image changed 20.5% of labels and a misleading one 19.4%, versus 11.6% when only the ignore-the-image instruction was deleted with the image left in place, and only 37% of the labels that differ between the two images moved toward the sense shown, while agreement with the human annotators was unchanged whether the image was absent, aligned, or misleading.
Technion – Israel Institute of Technology
Nagham Omar（First author）
MIST (Misleading-Image Stress Test)LLM-as-a-Judge Figurative Language Annotation Alternative Annotator Test (alt-test)
World Journal of Methodology Sep 20, 2026
Surgical paradox: Pen vs scalpel—an opinion review on reforming academic promotion
This opinion review argues that the current academic promotion system in surgery disproportionately favors quantifiable metrics such as publications and grant funding over the demonstration of clinical skill, creating a fundamental "credentialing paradox"; that AI tools, by streamlining research tasks, amplify this publication-centric culture and widen the disconnect between a surgeon's academic rank and their proficiency in the operating room, placing less-funded faculty and those who dedicate time to clinical care and education at a disadvantage; and that the profession should redefine academic success through a more holistic framework that formally recognizes and rewards excellence in clinical care, education, and mentorship alongside research output.
This opinion review argues that the current academic promotion system in surgery disproportionately favors quantifiable metrics such as publications and grant funding over the demonstration of clinical skill, creating a fundamental "credentialing paradox"; that AI tools, by streamlining research tasks, amplify this publication-centric culture and widen the disconnect between a surgeon's academic rank and their proficiency in the operating room, placing less-funded faculty and those who dedicate time to clinical care and education at a disadvantage; and that the profession should redefine academic success through a more holistic framework that formally recognizes and rewards excellence in clinical care, education, and mentorship alongside research output.
Carmel Medical Center Technion – Israel Institute of Technology
Ahmad Mahamid（First author, Corresponding author）
Credentialing paradox in surgical academic promotion
Journal of the American Medical Informatics Association Sep 17, 2026
DiagnosticXchange: An Open-Source Framework for Evaluating Safety, Efficiency, and Diagnostic Reasoning in Clinical AI Systems
This study developed and validated DiagnosticXchange, an open-source clinical simulation framework in which AI systems diagnose cases by ordering tests, requesting imaging, and performing procedures, with each action mapped to CPT codes capturing cost, time, work relative value units, and invasiveness; using 8 large language models on 216 peer-reviewed cases across 19 specialties (1728 sessions), it found that three systems with near-identical accuracy (93.5%–94.0%) differed significantly in cost (P < .001; 1.75-fold between the most and least expensive) and 2.
This study developed and validated DiagnosticXchange, an open-source clinical simulation framework in which AI systems diagnose cases by ordering tests, requesting imaging, and performing procedures, with each action mapped to CPT codes capturing cost, time, work relative value units, and invasiveness; using 8 large language models on 216 peer-reviewed cases across 19 specialties (1728 sessions), it found that three systems with near-identical accuracy (93.5%–94.0%) differed significantly in cost (P < .001; 1.75-fold between the most and least expensive) and 2.
Rambam Health Care Campus Mahidol University Siriraj Hospital Technion – Israel Institute of Technology
Moran Sorka（First author）
DiagnosticXchange Clinical AI diagnostic system evaluation Large language models
arXiv Sep 17, 2026
Social Laws for Multi-agent Coordination in Stochastic Environments: Formalizing and Verifying α-Robustness
This work extends social laws from deterministic, goal-based settings to stochastic, reward-based multi-agent environments by introducing α-robustness, a measure of the fraction of guaranteed utility each agent retains while following its optimal single-agent policy under the assumption that all agents obey the social law, and by presenting a verification approach that reduces to solving a series of Markov decision processes; experiments on grid toy environments show that with action success probability 1, parallel lanes, opposite lanes, and switch corners with a clockwise social law reach α=1 robustness, while a slow-speed social law does not always improve guaranteed utility.
This work extends social laws from deterministic, goal-based settings to stochastic, reward-based multi-agent environments by introducing α-robustness, a measure of the fraction of guaranteed utility each agent retains while following its optimal single-agent policy under the assumption that all agents obey the social law, and by presenting a verification approach that reduces to solving a series of Markov decision processes; experiments on grid toy environments show that with action success probability 1, parallel lanes, opposite lanes, and switch corners with a clockwise social law reach α=1 robustness, while a slow-speed social law does not always improve guaranteed utility.
The University of Texas at Austin Technion – Israel Institute of Technology
Rolando Fernandez（First author）
Social Laws Robustness Verification Markov Decision Processes Stochastic Game

You’re all caught up

Low-rank adaptation learns one matrix: SingLoRA beats LoRA and LoRA+ on LLaMA-7B at 60% of their parameter budget

T-cell clonal organization without TCR sequencing: CloneTrast infers clonal niches from transcriptomes alone

SAGO re-evaluates eleven LLMs by per-instance stability: no model generalizes uniformly, and cross-dataset variation can reverse rankings

MIST stress test: adding an image shifts about a fifth of VLM judge labels, regardless of what the image shows

Surgical paradox: Pen vs scalpel—an opinion review on reforming academic promotion

DiagnosticXchange: An Open-Source Framework for Evaluating Safety, Efficiency, and Diagnostic Reasoning in Clinical AI Systems

Social Laws for Multi-agent Coordination in Stochastic Environments: Formalizing and Verifying α-Robustness