How they use Nebius AI Cloud

Télécom SudParis’s SAMOVAR researchers used Nebius GPU resources for Task-in-Prompt adversarial attack experiments. The study reports Nvidia H100 inference with Unsloth and 383.61 GPU hours. Its PHRYGE benchmark comprises 40 adversarial tasks involving methods such as ciphers, riddles and programming challenges. It tested six language models and found vulnerabilities across all six. Reza Farahbakhsh acknowledged Nebius compute support when announcing the paper’s acceptance to ACL 2025.

A separate multilingual fake-news study used Nebius H100 cloud resources for synthetic news generation with a quantized 4-bit Llama 3 8B instruction-tuned model. Researchers compared zero-shot and few-shot prompting to augment the TALLIP and MM-COVID datasets for BERT-based classifiers. The study reported gains of up to 7.7 F1 points in some languages. Results varied by language and augmentation strategy, and excessive augmentation could reduce performance.

Nebius Academy also supports the team’s research into contextual toxicity, which proposes measuring contextual harm rather than treating toxicity as an intrinsic property of text. A May 2026 feature lists a collaboration portfolio spanning ASCII-art moderation attacks, Task-in-Prompt attacks, fake-news augmentation and contextual harm measurement. It identifies MultiSent-RAG, a multilingual sentiment-analysis system, as upcoming. In August 2026, Nebius described funding Télécom SudParis’s LLM toxicity and AI safety research through its Academy and Research Grants Program.