Martin Vechev

Data contamination in language models poses serious trust issues for evaluations.

SafeCoder improves the safety of code generated by language models.

Examining the effectiveness and vulnerabilities of watermarking in AI-generated content.

A novel approach enhances data recovery while addressing privacy concerns in federated learning.

Research shows vulnerabilities in federated learning's approach to text privacy.

A new benchmark for evaluating fairness in representation learning methods.

Examining the dangers of quantized language models and their potential misuse.

Research reveals the challenges of watermark detection in large language models.

A unified library enhances fairness in comparing neural network training methods.

New methods enhance uncomputation efficiency in complex quantum programs.

A method to adapt language models while reducing skill loss.

Analyzing vulnerabilities in popular code completion tools and their implications for developers.

This article discusses a new rating system for evaluating language models more fairly.