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.
FrED enables training data attribution without access to model weights. It combines feature similarity with domain-specific knowledge graphs to improve identification of influential training samples. Validated on image generation and weather forecasting.
A study (6 LLMs, 1103+1195 ideas) compares creativity evaluation: models agree with humans on novelty, but differ on context (social and market factors). Individual models apply standards of varying breadth; LLMs are less sensitive to contextual information.
The research team presents a framework for more accurate latency prediction when deploying LLMs on mobile and edge devices. On Pixel 8, it improved decoding accuracy from R² 0.957 to 0.973, and on Pixel 8 Pro, prefill latency from -1.383 to 0.966. The method combines static parameters with dynamic hardware telemetry.
A research paper proposes a new framework for evaluating personal LLM agents. The authors identify four conditions for effective testing: explicit temporal interventions, persistent user state, cross-component effects, and configuration variation. After auditing published benchmarks, they found no protocol…
Introduces CEL, a library with 18 datasets and 14 counterfactual explanation methods. It includes a unified evaluation protocol with metrics for validity, coverage, sparsity, and plausibility. The first comprehensive benchmark for systematically comparing methods within a unified framework.
The research paper introduces TILT, a method for improving image generation from text descriptions. It addresses diffusion model failures on complex prompts involving multiple concepts. The method uses the model's intrinsic reward without external supervision. Results on the T2ICompBench benchmark show improvement without reducing quality…
RED-PIM research optimizes transformer processing through algorithm–architecture co-design. The method reduces inter-bank data movement from O(N²) to O(N) and shrinks cache matrices from N×N to d×d. On real data, it improves inference by 99.60 % for long documents and 13.44 % for shorter…
Researchers introduced TLM for processing long, temporally structured texts using LLM agents. Traditional RAG fragments chronological context; TLM preserves it through LGCM and SHAP feedback. It was tested on medical questions and financial event prediction, with better results.
An arXiv preprint examines how the weight of multi-horizon latent consistency affects video predictor geometry. On Moving-MNIST, increasing lambda from 0 to 0.8 reduced latent-space expansivity (L20 from 4.96 to 1.01) and prediction error. The same effect does not hold in other domains (Pendulum…
A research contribution introduces latent PDE mapping to improve generalization of physics-informed neural networks across geometries with limited data. Tests in cardiac electrophysics achieved up to a 4–6× reduction in L2 error.
The research paper presents the concept of Industrial Tokenization — a way of structuring heterogeneous industrial data from sensors, maintenance and diagnostic systems into standardized units for interpretation using large language models. The pilot implementation includes vibration diagnostics.
A study examines how four language models represent self-harm content. Using linear probes on two datasets, it found that relevant information crystallizes in the final 3–7 % of network layers; the most accurate probes are not necessarily the most linearly separable.
The research analyzes 8 records from deployments of gesture interaction in kiosks. It identifies 20 failures arising from the difference between individual-frame recognition and the stability of interaction events. It defines 6 failure classes and proposes a runtime abstraction for event confirmation.
Researchers studied embedding documents into Gemma model weights through LoRA adapters. They found that training data quality strongly affects accuracy — data curation increased accuracy from 57.7 % to 85.7 % on a 15-document corpus. The LoRA adapter outperformed a BM25-RAG baseline.
A research preprint describes a diffusion model for translating images from simulation to reality using a wavelet transform. On CARLA video, it improved automated metrics (ADE, FDE) by 5.4 % and 5.1 %.
A mathematical proof that the maximum of n real numbers can be represented by a ReLU network with two hidden layers for n≤10. More generally, the maximum can be represented with log₅(n)+1.57 hidden layers; the result also applies to continuous piecewise-linear functions.
A scientific paper from arXiv uses machine learning to examine how toxic behavior develops and spreads in the decentralized social network Mastodon, which has no uniform moderation standards. The goal is to understand toxicity trends and their impact on community health.
The research paper improves the theoretical analysis of convergence for LoRA (Low-Rank Adaptation). It shows that finding a stationary point requires only O(ε⁻⁴) gradient evaluations instead of an exponential number. It proposes the LoRA-NSGDM and LoRA-STORM algorithms with improved oracle complexity for the stochastic setting.
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
