Research separate from news
A preprint is a signal, not a finished product or an independently confirmed result. We therefore track research papers separately and show the actual state of supporting evidence for each one.
A significant claim from a single source is published only after further confirmation.
A research preprint describes CT-Merging, an algorithm for efficiently merging multiple LoRA adapters into a single universal adapter. On the DC-Merge CLIP benchmark, it improves by 2.56 points on ViT-B/32 and 1.51 points on ViT-L/14 over existing methods.
A study introduces JAXBench, a benchmark suite with 50 JAX workloads for optimizing TPU kernels using AI. It includes 17 production operators from models such as Llama-3.1 and DeepSeek-V3. With Gemini 3 Flash, correctness increased from 5.8% to 37.3%, with speedups of 1.28x to 1.36x.
A research paper introduces optimized Triton kernels for accelerating LLM inference. SonicSampler combines logit processing and token selection into a single kernel, supporting dynamic behavior per request. It achieves speedups of up to 10× for the top-k algorithm and up to 16× for speculative decoding while preserving…
A scientific preprint describes an approach to runtime verification of machine learning model decisions in optical networks. The system checks the coherence of model explanations and the physical correctness of decisions before executing them in automated control loops. Experiments on optical link quality classification…
A research paper introduces EvoSQL, a method for improving SQL generation from natural language. It combines a generator and an LLM critic with candidate memory, verifies SQL through execution and LLM analysis, and improves through SDPO fine-tuning. Tests on the Spider and BIRD benchmarks show improvements; on BIRD-Dev, performance…
A scientific paper introduces Codec-Gauge, a technique using learned orthogonal channel transformations for more efficient KV-cache compression in Transformers. Across six models with 3–6 bits/value, it achieved a 44% reduction in KL divergence compared with raw coordinates and outperformed DCT, PCA and Hadamard controls.
Scientists developed a deep learning pipeline with a Swin3D Transformer to predict the spatiotemporal behavior of electrodes in Li-ion batteries. Gaussian Positional Encoding and Temporal Encoding innovations improve accuracy over point-cloud methods. The method reduces computational demands by orders of magnitude and enables faster…
CLOE is a new method that combines an autoencoder with a Christoffel-function-based detector for anomaly detection in high-dimensional data. It preserves the simplicity and low calibration requirements of Christoffel methods while improving their scalability.
A new benchmark evaluates AI agents optimizing LLM inference on an H100 GPU within a 2-hour limit. Agents achieve speedups of up to 8× over PyTorch and approach vLLM, but fall short of hyperparameter optimization (11.53×). The identified problem is not domain knowledge but the ability to propose and…
A research paper on arXiv describes a new cache eviction strategy in multi-agent systems. The strategy combines recomputation costs, the number of graph dependencies and agent invocation frequency. On three benchmarks, it reduces latency by up to 64.7 % compared with no caching and by 31.1 %…
A preprint introduces DC-Leap, a method for accelerating diffusion LLM inference without retraining. It uses a dynamic verification strategy to address Joint Probability Dependence Error and draft-guided decoding. Tests show speedups of up to 53× on MBPP, or 105× with KV-Cache, with…
A research preprint introduces a framework for continual learning in forecasting models that combines attention mechanisms with experience replay. The approach enables adaptation to new contexts without catastrophic forgetting and reduces retraining costs. Evaluated on benchmarks and the dataset…
A study introduces ConfidenceBench, a benchmark measuring how well models assess their own uncertainty. Across 200 questions, Claude Opus 4.6 and Gemini 3.1 Pro Preview achieved the lowest Brier score (0.103), while Gemini 3.1 Flash-Lite shows severe miscalibration (0.367). The finding shows…
Research shows how information is organized in active inference agents. Causal emergence (measured by Integrated Information Decomposition) concentrates in a slow global latent layer separated from reward processing. Its quantities are architectural and, during learning without external reward…
A study examines whether regex filters improve the security of LLM applications using Gemini-2.5-flash. Across 45 adversarial attempts, the filter never blocked dangerous content (0 %), although an LLM judge detected refusals in 56–100 % of cases. Conclusion: the effectiveness of model alignment depends on the measurement metric.
A research paper on arXiv introduces an algorithm for more efficient processing of text-attributed graphs with LLMs. The method combines data distillation with Wasserstein Distance and generates textual summaries. It achieves a better performance-compression trade-off in downstream tasks.
Researchers identified characteristic spectral instability (Spectral Drift) in neural networks' internal activations when they fail. They introduced Self-Detecting Neural Networks (SDNN), a lightweight framework with 5 % additional parameters that detects failures using Fourier transforms and wavelets.…
STeMP is a new protocol for standardized documentation of spatiotemporal machine learning models in environmental research. It is available as a GitHub project with an R package and a web application. It has three sections: Overview, Model and Prediction, with automatic warnings about common errors.
AI Radar monitors Czech and international sources every day, looking for changes that truly deserve attention.
The result is a quick overview of what has actually changed in the AI world, rather than another stream of articles.
Use the CS/EN switch to read the same Radar in Czech or English. English content is published after its translation has been checked, so new and older items may appear later.
Each row represents one event — not one article. At a glance, you can see its significance, credibility and main point.
Illustrative example, not a current news item.
Agent mode is available to all paying users
Until now, the mode was available only on the highest plan; it is now available on all paid tiers without a waitlist.
The detail page contains a fuller summary, its significance and original sources. Practical impact appears in the detail and the For individuals and For businesses views. An AI Flash item reaches the main selection only after it has been expanded and meets the publication rules.
We first summarize each event in the same way for everyone. Based on those same facts, we then explain what the change means for your own use and what it could mean for how a company operates.
Today’s briefing is the same for everyone. Pages For individuals and For businesses can be found in the main navigation — they select only events relevant to the given use case.
Reset history · Outage history · Subscribe in a feed reader (RSS) · Methodology and label explanations
