undefined

undefined

Over the past year, we've seen a lot of open-source model releases: DeepSeek V3, Kimi K2, Qwen3-Next, and others. These models are now competitive with GPT-5 and Claude on many benchmarks, and the teams behind them have been openly sharing their methods. The papers describe several breakthroughs that change how we think about pre-training.

In this talk, Vladislav will cover the main ideas that emerged in 2025: FP8 training at 600B-parameter scale, new optimizers like Muon that are finally here to challenge AdamW, and other advances in training efficiency and MoE architecture.

Large Language Models (LLMs) are increasingly used to generate synthetic textual data for training smaller specialized models. However, a comparison of various generation strategies for low-resource language settings is lacking. While various prompting strategies have been proposed—such as demonstrations, label-based summaries, and self-revision—their comparative effectiveness remains unclear, especially for low-resource languages. In this paper, we systematically evaluate the performance of these generation strategies and their combinations across 11 typologically diverse languages, including several extremely low-resource ones. Using three NLP tasks and four open-source LLMs, we assess downstream model performance on generated versus gold-standard data. Our results show that strategic combinations of generation methods—particularly target-language demonstrations with LLM-based revisions—yield strong performance, narrowing the gap with real data to as little as 5% in some settings. We also find that smart prompting techniques can reduce the advantage of larger LLMs, highlighting efficient generation strategies for synthetic data generation in low-resource scenarios with smaller models.

This paper presents AI Super Agent, a self-organizing multi-agent system designed to autonomously decompose, plan, and execute complex tasks across multimodal domains.

At its core lies a Cognitive Core — a unifying control architecture that integrates perception, reasoning, memory, and goal management within a continuous Plan–Execute–Control (PEC) loop. This core dynamically orchestrates Model Context Protocol (MCP) servers, maintaining coherence between reasoning processes, action execution, and long-term memory.

The framework incorporates a Graph-based Memory (GraphRAG) enhanced with Deep Research Algorithms, enabling contextual retrieval, semantic graph reasoning, and iterative knowledge synthesis. An Action Graph Engine represents and manages causal task dependencies, allowing agents to construct, evaluate, and refine strategies in real time.

Through this architecture, AI Super Agent demonstrates the capability to self-organize, spawn specialized sub-agents, and adaptively learn from multimodal feedback. Experimental evaluations in domains such as business process automation, financial analytics, and research intelligence reveal substantial improvements in reasoning depth, task completion rate, and coordination efficiency compared to conventional multi-agent baselines.

Beyond its technical contributions, AI Super Agent establishes a foundation for autonomous cognitive ecosystems — systems capable of co-evolving with human collaborators, enabling scalable problem-solving, continuous discovery, and the expansion of collective intelligence.

This paper presents a multi-agent adversarial framework for solving complex business problems in which single AI models exhibit instability, overconfidence, and sensitivity to query formulation. The approach is based on an architecture of interaction between multiple independent agents using roles, constraints, and evolutionary argument refinement strategies to generate alternative positions. The full system pipeline is considered: role initialization, parallel position generation, peer review and criticism mechanisms through schema-guided reasoning, iterative argument refinement, and final result aggregation. Special attention is given to adaptive computational budget management, preventing opinion collapse, and extracting reproducible consensus based on structured preferences. The framework implements a stateful-by-design multi-agent architecture: a moderator agent solves these problems by orchestrating rounds and managing the budget without interfering with the content of positions, while debater agents maintain context between rounds and can use external tools to justify their arguments. This approach differentiates itself from stateless subagents in existing SDKs. It demonstrates how such architectures can be used to support strategic, investment, and management decisions.

Human intelligence is ”by default" multimodal: we see, hear, and act at the same time. In 2025, the industry took a significant step towards making LLMs the same: vision + voice + action ceased to be a “binding around the text” and began to turn into full-fledged modalities inside/around the model.
In the report, we will analyze the engineering and research patterns that will be used to add new modalities to LLM in 2025: (1) input adapters (encoder → projection → LLM tokens), (2) end‑to‑end omni architectures with audio‑video timing and streaming speech generation (using the example of Thinker‑Talker/TMRoPE‑approaches), (3) “action‑modality" — from tool‑calling to computer‑use models that perceive the GUI as an image and perform mouse/keyboard actions in an iterative loop.
A separate emphasis is on the fact that in 2025, public releases began to systematically record metrics for voice (WER/noise tolerance, speech stability), for vision—reasoning (MMMU, etc.) and for action/computer‑use (success rate on OSWorld/WebArena/WebVoyager and analogues), which allows you to compare approaches and build a reproducible R&D cycle “measure → improve → ship”

Ore mining is where metal industry begins. Metal products quality directly depends on the ore quality. Ore mining is the complex process still relying on a manual control. In Severstal Digital we are aiming toward the complete manufacturing digitalization. In this work we are presenting a solution for the intellectual ore passes monitoring. Such a technology gives us a more automated control over the mining, prevents production downtimes and increase the ore quality. Developed system has proved economical efficiency and currently in use on Severstal factories.

This talk presents an end-to-end system for real-time, multi-camera incident analytics that operates under strict latency and resource constraints typical of industrial edge deployments. A key feature is the system’s ability to process 100 live RTSP camera streams simultaneously on a single A100-80GB GPU and to detect a specific incident: cargo thrown by workers or forklifts during loading/unloading operations. The system continuously scans the video stream, locates a 40-frame segment containing a cargo-drop incident, forwards it to an operator for verification, and logs it into a dedicated report for statistical analysis and subsequent decision-making.


We first articulate the practical pain points encountered when ingesting and analysing such streams:
A) A flood of independent camera channels that must be synchronized in real time while respecting network-bandwidth;
B) GPU-memory limits;
C) Discussion of sparse, non-stationary and biased data that preclude known metrics at project kick-off;
D) Camera-drift phenomena that require auxiliary neural networks to re-align sensors;
E) Problem of deffirintiation a strict border between “good” from “bad” frames video quality;
F) A sub-second human-in-the-loop interaction model whose UI wireframes and feedback pipeline are described in detail.

Second, we propose a neural pipeline that slices incoming frames, classifies them for action presence, extracts regions of interest at tile-level granularity and applies channel separated neural network, as alternative fod 3D convolutions for final tile classification, achieving end-to-end latency below 200 ms on a 10G MIG instance on single A100-80G GPU.

Third, we detail the data-engineering workflow: training a segmentator on full frames, labelling tile-level objects via a custom web tool and closing an active-learning loop in closed corparate contour.

Finally, we define and continuously track both business metrics (incident-to-damage regression, false-negative cost, operator NPS) and technical metrics (like frame-drop percentage), while monitoring model drift.

The talk will be interesting as for CV engeneers and for bussiness who wants to prove economic effect from such system.

Large Language Models (LLMs) are an essential part of modern artificial intelligence systems in the fields of natural language processing, computer vision, and reinforcement learning. However, the rapidly growing complexity of LLMs, characterized by a 450x increase, leads to a widening gap between model size and hardware capacity and increased communication overhead during distributed training, fine-tuning, and inference. To tackle these issues, leading hardware vendors are implementing support for low-bit formats in the new generation of hardware. However, such lossy approaches as quantization and sparsification reduce communication but lead to a significant quality drop. Lossless compression is a promising alternative for reducing data volume without degrading quality, but it adds substantial computational resources. Compression of low-bit values is a challenging task due to the high entropy of these data and lack of simple repeating patterns. To address this limitation and increase low-bit data compressibility, we focus on entropy-based data transformations. The grouping of blocks of data with similar entropy properties together and lossless entropy data clustering are perspective approaches in this field. These transformations are computationally inexpensive and consistently improve lossless compression ratios for low-bit LLM data across multiple state-of-the-art compressors, thereby reducing end-to-end communication overhead in distributed workflows.

AI agents show remarkable success at various short tasks, and are rapidly improving at longer-horizon tasks, creating a need to evaluate AI capabilities on dangerous tasks which require high autonomy. Evaluations (evals) comprising long-running "real-world" tasks may be the best proxies for predicting general performance, but they are expensive to create, run, and compare to human baselines. Furthermore, these tasks often rely on a large, interwoven set of agent skills, which makes predicting capabilities development difficult. We hypothesize that precursor capabilities including 'persistence', 'dexterity', and 'adaptability' are upstream of robust autonomous performance on long-horizon tasks, and design simple procedurally-generated 'proxy' evals to target these precursors. We then use agent performance on our proxy evals to calibrate a preliminary method of capability prediction on a more complex task: SWE-Bench. Our preliminary results show that performance on certain proxy evals can be unusually predictive of performance on other evals. We find that a simple adaptability proxy based on developmental psychology correlates with SWE-bench with, and three other proxies correlate with SWE-bench at r > 0,8. A proxy eval which only takes 10 steps is strongly correlated with the performance of many other evals, which otherwise take much longer to terminate (100s of steps). For our predictive model, our initial results correctly predict agent scores on SWE-bench, but have large error bars, suggesting that -- testing more models on more synthetic evals -- we can quickly and cheaply predict performance on important long-horizon tasks.

We propose a novel explainable document classification framework that integrates Concept Whitening (CW) with graph concepts that are derived from stable graph patterns, and extracted via methods based on Formal Concept Analysis (FCA) and pattern structures. Document graphs are constructed using Abstract Meaning Representation (AMR) graphs, from which graph concepts are extracted and aligned with the latent space axes of Graph Neural Networks (GNNs) using CW. We investigate four types of graph concepts for their effect on concept alignment: frequent subgraphs, graph pattern concepts, filtered equivalence classes, and closed subgraphs. A novel filtration mechanism based on support, along with a custom penalty metric, is proposed to refine graph concepts for maximizing concept alignment. Experiments on the 10 Newsgroups and BBC Sport datasets show that our document graphs effectively capture both structural and semantic information, thereby supporting competitive classification performance across multiple GNN model architectures and configurations. For the 10 Newsgroups dataset, GNN models equipped with a CW module show an average increase of 0.7599 in the macro-averaged F1 score of the Concept Alignment Performance (CAP) metric, with an average drop of only 0.0025 in the document classification macro-averaged F1 score. Similarly, on the BBC Sport dataset, the average CAP improvement is 0.6998, with an average drop of 0.0894 in document classification performance. Additionally, concept gradient importance analyses and concept similarity heatmaps provide insights into the interpretability and structural separability of the GNN's latent representations, achieved using CW.

Learning diffusion bridge models is easy; making them fast and practical is an art. Diffusion bridge models (DBMs) are a promising extension of diffusion models for applications in image-toimage translation. However, like many modern diffusion and flow models, DBMs suffer from the problem of slow inference. To address it, we propose a novel distillation technique based on the inverse bridge matching formulation and derive the tractable objective to solve it in practice. Unlike previously developed DBM distillation techniques, the proposed method can distill both conditional and unconditional types of DBMs, distill models in a one-step generator, and use only the corrupted images for training. We evaluate our approach for both conditional and unconditional types of bridge matching on a wide set of setups, including super-resolution, JPEG restoration, sketch-to-image, and other tasks, and show that our distillation technique allows us to accelerate the inference of DBMs from 4x to 100x and even provide better generation quality than used teacher model depending on particular setup.

Generative AI is already transforming how designers work — accelerating ideation and opening a way to creating with fewer limitations. Yet today's tools still address only part of the creative process. This talk explores how new design workflows are emerging, what AI already does well, and where it still falls short for creative professionals.

At Recraft, we focus on building models and other tech that give creative professionals full control over their vision. Achieving this means solving some hard technological challenges, which we will also discuss during this talk.

Extreme weather events such as heavy rains, thunderstorms, and hail play a huge role for different parts of human life: aviation, agriculture, everyday life, etc. We decided to focus on the task of thunderstorm nowcasting because of two reasons: (1) thunderstorms are often accompanied with heavy rain and hail, so they have a big impact to industry and everyday life, and (2) thunderstorms are often rapidly developed, so it is useful and necessary to forecast them with a small time discretisation which is 10min in our case. We prepared a visual transformer based model which uses real-time data from weather radars and geostationary satellites to forecast areas with high probability of lightnings along to precipitation nowcasting. The resulted quality of our model is better than classical approaches for that task such as numerical weather forecast and optical flow in terms of F1-measure and IoU, as well as in terms of visual assessment. The proposed model is integrated into Yandex Weather service as a production model.

Text-to-image (T2I) diffusion models are popular for introducing image manipulation methods, such as editing, image fusion, inpainting, etc. At the same time, image-to-video (I2V) and text-to-video (T2V) models are also built on top of T2I models. We present Kandinsky 3, a novel T2I model based on latent diffusion, achieving a high level of quality and photorealism. The key feature of the new architecture is the simplicity and efficiency of its adaptation for many types of generation tasks. We extend the base T2I model for various applications and create a multifunctional generation system that includes text-guided inpainting/outpainting, image fusion, text-image fusion, image variations generation, I2V and T2V generation. We also present a distilled version of the T2I model, evaluating inference in 4 steps of the reverse process without reducing image quality and 3 times faster than the base model. We deployed a user-friendly demo system in which all the features can be tested in the public domain. Additionally, we released the source code and checkpoints for the Kandinsky 3 and extended models. Human evaluations show that Kandinsky 3 demonstrates one of the highest quality scores among open source generation systems.

Text-to-image generation models have gained
popularity among users around the world. However,
many of these models exhibit a strong bias
toward English-speaking cultures, ignoring or
misrepresenting the unique characteristics of
other language groups, countries, and nationalities.
The lack of cultural awareness can reduce
the generation quality and lead to undesirable
consequences such as unintentional insult, and
the spread of prejudice. In contrast to the field
of natural language processing, cultural awareness
in computer vision has not been explored
as extensively. In this paper, we strive to reduce
this gap. We propose a RusCode benchmark for
evaluating the quality of text-to-image generation
containing elements of the Russian cultural
code. To do this, we form a list of 19 categories
that best represent the features of Russian visual
culture. Our final dataset consists of 1250
text prompts in Russian and their translations
into English. The prompts cover a wide range
of topics, including complex concepts from art,
popular culture, folk traditions, famous people’s
names, natural objects, scientific achievements,
etc. We present the results of a human
evaluation of the side-by-side comparison of
Russian visual concepts representations using
popular generative models.