YerevaNN

27/04/2026

On the last day of ICLR workshops we have four posters.

1. At GEMBio Workshop Tatevik Abrahamyan will present how we benchmarked molecular optimization algorithms against three categories molecular design tasks, including a novel one (for this line of research) that tests whether AI models can generate molecules with selective binding to two proteins (or pockets).

PMO-Dock: Benchmarking Docking, Specificity, and Generalization in Molecular Optimization

2. At FM4Science Workshop Filya Geikyan will present our new tokenizer for 3D molecular structures. It is an attempt to make autoregressive models competitive with diffusion or flow-match-based methods for 3D conformer generation.

CoordToken: Accurate Tokenization of 3D Small Organic Molecules

Then we have two more posters at Machine Learning for Remote Sensing workshop:

3. Ani Vanyan will present our new benchmark for cross-satellite generalization of remote sensing foundation models:

GeoCrossBench: Cross-Band Generalization for Remote Sensing

4. Anna Khosrovyan will present how we should evaluate "face recognition" for vessels in satellite imagery.

HOSS-Bench: Open-Set Cross-Modal Vessel Re-Identification Benchmark

See you there!

24/04/2026

We are happy to announce that six of our team members are in Rio de Janeiro this week for ICLR 2026.

Our lab will be presenting seven posters across various workshops, highlighting our latest research in foundation models, molecular optimization, and remote sensing. If you’re attending the conference, we’d love to connect and discuss potential collaborations!

📅 Sunday, April 26
1. PMO-Dock: Benchmarking Docking, Specificity, and Generalization in Molecular Optimization | GEM Workshop
2. Beyond Data Size: Impact of Dataset Diversity and Density in Self-Distillation | DATA-FM Workshop
3. Bridging the Sim-to-Real Gap: RF Localization with Large-Scale Synthetic Pretraining | DATA-FM Workshop
4. Aerial Vision-Language Navigation: Enhancing Navigation with Map Grounding | ES-Reasoning Workshop

📅 Monday, April 27
5. CoordToken: Accurate Tokenization of 3D Small Organic Molecules | FM4Science Workshop
6. GeoCrossBench: Cross-Band Generalization for Remote Sensing | ML4RS Workshop
7. HOSS-Bench: Open-Set Cross-Modal Vessel Re-Identification Benchmark | ML4RS Workshop

Anna Khosrovyan, Ani Vanyan, Narek Nurijanyan, Armen Manukyan, Filya Geikyan, Tatevik Abrahamyan

23/04/2026

🔊 We are happy to announce the third edition of the Armenia LLM Summer School!
📅 We are returning to Yerevan from August 3–7, 2026, and we’re excited to be hosted by our partner, the AI9 Startup Campus.
What’s on the Agenda?
▫️Day 1: Pre-training, Scaling Laws & MoE (with Andre Martins, IST / Unbabel)
▫️Day 2: SFT & Reinforcement Learning (Speaker TBC)
▫️Day 3: Test-time Scaling & Compute-optimality (with Ivan Mashkov, NVIDIA)
▫️Day 4: Agents, Reasoning & Workflows (Speaker TBC)
▫️Day 5: Advanced Topics: VLA, Robotics & World Models (Speaker TBC)

We cap off the week on August 8th with an intensive 24-hour Hackathon! This will be organized by AI9, with Armenia LLM Summer School organizing committee serving as the content partner. The Hackathon will have separate application form and will be open to both the school attendees and other applicants.

Join us in person in Yerevan or via our online track for theoretical sessions.
✅ Apply now: armllm.github.io/2026

Special thanks to the AI9 Startup Campus for being a partner for this year's school as well!

25/03/2026

ԵՊՀ - Yerevan State University, YerevaNN and Eleveight AI invite AI engineers, students and researchers from Armenia to participate in OpenAI Parameter Golf!

The goal is to train the best language model that fits in 16MB and runs in under 10 minutes on 8×H100 GPUs.

Yerevan State University will provide compute on its cluster for participants from Armenia, sponsored by Eleveight AI. Eleveight AI will also award a 500,000 AMD prize for the best submission from Armenia.

Because of how Parameter Golf is designed, it is not trivial to determine the best submission among participants from Armenia. We believe we found a fair approach, please see https://github.com/YerevaNN/Parameter-Golf-Armenia/ for details.

To apply for 20 GPU-hours on the YSU cluster for your Parameter Golf submission, use this form: https://forms.gle/RGGjcAhCsBGY4NjY9

We wish you every success in making it onto the OpenAI leaderboard!

24/03/2026

Lots of news coming out of the lab over the next days. Stay tuned!

09/03/2026

We are happy to co-organize the 4th edition of the AI Conf Armenia along with the ԵՊՀ - Yerevan State University, Union of Advanced Technology Enterprises-UATE and Krisp.

Stay tuned for more details on agenda, speakers, participants, and more!

🤖🇦🇲 AI Conf Armenia-ն՝ արհեստական բանականությանը նվիրված հայկական խոշորագույն համաժողովը, վերադառնում է։

Գալիք AI Conf-ը՝ ոչ թե պարզապես 4-րդ, ոչ թե քառապատկված,
այլ 4-րդ աստիճանում` ճիշտ համահունչ ոլորտի էքսպոնենցիալ զարգացմանը։ 📈

📅 Ապրիլի 18-ին հավաքվում ենք Մայր ԲՈւՀ-ի հարկի տակ՝ վերաիմաստավորելու ԱԲ ոլորտի աճը և նախանշելու ապագայի ուրվագիծը։

🎯 Առանցքում՝ GPU-ներով մասշտաբավարվելու և երկիրը մասշտաբավորելու հեռանկարը։

🤝 Կազմակերպիչներ՝

ԵՊՀ - Yerevan State University, Krisp, YerevaNN, Union of Advanced Technology Enterprises-UATE

02/01/2026

We finished 2025 with a fresh publication on AI applications in the RF domain.

A mobile device can be localized in an urban environment by measuring signal strength from various base stations in the city. The most straightforward approach is to collect fingerprints on the streets and then, for a given device, compute a similar fingerprint and match it to the nearest neighbor.

In this paper, we ask whether this approach can be extended to streets with no prior fingerprints. Deep learning methods can fuse urban maps, radio signals, and base station locations to predict the device location. However, the error remains astonishingly high, given the scarcity of real-world datasets.

We used NVIDIA Sionna to simulate a large number of RF signals in an environment modeled after the city of Rome and showed that pretraining on synthetic data can reduce the localization error by 50%.

Rafayel Mkrtchyan Armen Manukyan Ararat Saribekyan Theofanis Raptis Hrant Khachatrian

We wish you a productive new year!

Link to the paper: https://www.sciencedirect.com/science/article/pii/S1566253525011662

Link to the code: https://github.com/YerevaNN/RF-Loc-Sim2Real

06/12/2025

Join us at the NeurIPS 2025 Workshop on Embodied World Models for Decision Making to discuss the challenges of automatic aerial navigation!

We fine-tune small VLMs to solve the CityNav benchmark. We show how certain design choices affect the results. We show synthetic data is needed. Our results are SOTA but still a long way to go to reach acceptable performance.

Hakob Tamazyan, Narek Nurijanyan, Boris Martirosyan, Hrant Khachatrian

26/11/2025

The cost of launching new satellites is rapidly decreasing. In the coming decade, hundreds of new Earth observation satellites will be deployed, many featuring novel sensors and spectral band combinations. Can AI models trained on existing labeled data generalize to imagery from these new sensors before specific labeled datasets become available?

We systematically address this question by introducing a new benchmark, GeoCrossBench. It is designed to test modern foundation models with a simple recipe: train on existing datasets on one subset of bands (e.g., RGB of Sentinel-2) and then evaluate on other subsets, such as RGB+NIR and NIR+SWIR from Sentinel-2, or even Sentinel-1. We show that DINOv3 (the regular version, not satellite) outperforms all remote-sensing-specific foundation models in the ViT-B category.

Furthermore, we extend ChannelViT to use iBOT/DINO-like pretraining with channel subsampling. This model, called χViT, is pretrained on a large corpus of satellite imagery and outperforms DINOv3 in transfer learning settings.

We publicly release the code and the model weights. https://github.com/YerevaNN/rs_foundation_models

Read the preprint on arxiv: https://arxiv.org/abs/2511.02831

👏 The models were trained on H100s at ԵՊՀ - Yerevan State University and on a node kindly donated by Nebius.

👏 This work was supported by the Higher Education and Science Committee of Armenia of RA (Research project No 24RL-1B049) (Բարձրագույն կրթության և գիտության կոմիտե)

👏 This work was also supported by the Sastic -- Strategic Armenian Science and Technology Investment Community.

Hakob Tamazyan, Ani Vanyan, Alla Barseghyan, Anna Khosrovyan, Evan Shelhamer, Hrant Khachatrian

24/10/2025

We would like to take a moment to sincerely thank our valued partner, 🤍 Mayro company.

Their consistent and thoughtful support is key to our research, allowing our team to continue its work.

It is through partnerships that the research community in Armenia can grow and develop.

We are deeply grateful for their commitment to our work.

To learn more about Mayro: https://www.mayro.am/

11/09/2025

YerevaNN’s Alla Barseghyan presents at DataFest Yerevan! Details in the post ⤵️

Closing out the DataFest Yerevan 2025 speaker lineup with Alla Barseghyan

Alla is a Machine Learning Researcher at YerevaNN. Her work focuses on data-centric AI, exploring how dataset quality and structure impact the performance of machine learning models across different domains.

📚 In this talk, Alla will explore the idea that sometimes less data can actually be more. She will present data specialization techniques for curating compact yet highly effective datasets tailored to specific architectures, objectives, and compute budgets. By benchmarking data filtering algorithms across multiple modalities, from natural images (ImageNet) to remote sensing imagery and X-ray scans, her work sheds light on strategies for building higher-quality datasets with fewer resources.

🧠 Talk Title: Less is More? Data Specialization for Various Modalities

📅 September 12–13
📍 Woods Center
✍️ Register now: https://datafest.am/