BRAInLab (@Lab4BRAIn) profile picture

Basic Research of AI: Optimization in ML, Federated Learning, GenAI, LLM, Maths for AI, AI for Science Head: DrSc Aleksandr Beznosikov [email protected]

Joined September 2025
Photo by BRAInLab (@Lab4BRAIn)Photo by BRAInLab (@Lab4BRAIn)Photo by BRAInLab (@Lab4BRAIn)Photo by BRAInLab (@Lab4BRAIn)Photo by BRAInLab (@Lab4BRAIn)Photo by BRAInLab (@Lab4BRAIn)Photo by BRAInLab (@Lab4BRAIn)Photo by BRAInLab (@Lab4BRAIn)Photo by BRAInLab (@Lab4BRAIn)Photo by BRAInLab (@Lab4BRAIn)
🔥 BRAIn Lab introduces SimplexLoRA: a new method for more efficient LLM fine-tuning. Standard LoRA allocates resources evenly across layers. SimplexLoRA focuses them on task-relevant layers — improving quality under the same compute budget. arxiv.org/abs/2610.37206 #EMNLP2026
2
1
6
332
We evaluated SimplexLoRA on 12 tasks, including sentiment analysis, QA, math reasoning, and summarization. Experiments covered four language models of different sizes, including Llama 3.1 8B and Qwen3-8B. #LLM #LoRA
1
1
40
🌍 The work is being presented at EMNLP 2026 in Budapest. 👏 Co-authors from BRAIn Lab: Andrey Veprikov, Grigory Davydenko, and Alexander Beznosikov. #EMNLP2026 #NLP #AI
102
Hi Budapest! BRAIn Lab is at @emnlpmeeting We are here with 4 accepted works on LLM compression, quantization, refusal directions and LoRA. If you are here - stop by our poster!
8
194
First paper for our young scientists Danil Rudenko and Georgii Bulgakov in collaboration with Yandex Research 🫶 proud of it!
Training neural networks for tabular regression? Our latest work may be of interest And it's not only about which of the 18! regression objectives is best, but touches on the role of noise and global early stopping in tabular DL in general arxiv.org/abs/2610.13909 Thread (1/N)
7
228
Training neural networks for tabular regression? Our latest work may be of interest And it's not only about which of the 18! regression objectives is best, but touches on the role of noise and global early stopping in tabular DL in general arxiv.org/abs/2610.13909 Thread (1/N)
2
5
4
58
7,228
1/10 Muon and Shampoo make optimizer updates matrix-aware. What about weight decay? Our Spectral Weight Decay encourages low-rank weights during training, improving compression and resistance to label noise. We'll present it at NeurIPS OPT 2026. arxiv.org/abs/2610.11730
1
2
1
278
7!!! accepted papers from BRAIn Lab for NeurlPS. So proud of our colleagues! 1. On Communication-Efficient Training of Ensembles in Federated Learning 2. High-probability Convergence of Gradient Methods under Markovian Stochasticity
1
4
641
3. Normalized SGD in the Convex Regime: First High-Probability Guarantees and Momentum Extension 4. ECG Dataset with Multi-Expert Annotations and Delineations
1
1
142
5. Metropolis-Scale Road Network Dasasets for Fine-Grained Urban Traffic Modeling
56
1) Muon works because it updates a matrix as a matrix. LoRA splits that matrix into A and B, then most optimizers update the factors independently. The induced step on BA can point somewhere quite different. We built LoRA-TSD to put Muon's geometry back where it belongs.
1
2
2
3
714
✈️ Last week, BRAIn Lab attended UAI 2026 in Amsterdam, presenting our paper "Markovian Compression: Looking to the Past Helps Accelerate the Future" arxiv.org/abs/2601.05398 – at Amsterdam, The Netherlands
3
5
175
🇳🇱 Amsterdam's canals, bikes 🚲 and unpredictable weather made it even better.
12
🗣️ Andrei Veprikov on the trip: "UAI was a cozy, friendly conference — easy to catch talks and chat with authors afterward. Great for getting a feel for the community, even coming from a related field."
10
We introduce Markovian compressors BanLast and KAWASAKI — they use transmission history to avoid resending the same coordinates, making the same communication budget go further in distributed learning.
59
Happy to share that four papers was accepted to EMNLP 2026!
2
1
24
1,220
🧵 Accepted at UAI 2026, @UncertaintyInAI . Distributed training is bottlenecked by communication, so everyone compresses gradients: RandK, QSGD, PermK. But almost every compressor is memory-less. It redraws coordinates from scratch every step, as if history did not exist.
2
1
3
133
🇰🇷 BRAIn Lab has touched down at KDD 2026 in Korea!
3
5
164
If you're at KDD, swing by poster #109 and chat with us 👋
8
Two more papers also made the cut: • TensorKDD — tensor decompositions for post-training LLM compression • RelKD — ultra-low-bit inference in reasoning models
14
Right now, Alexander Myasnikov is presenting Hierarchical Mixture-of-Experts with Two-Stage Optimization at the poster session. We're rethinking how to optimize MoE architectures — and the results are exciting.
73