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.
1381
published research events
Latest work
A significant claim from a single source is published only after further confirmation.
The study introduces CoHARL, which jointly adjusts agent strategies and the code for their interfaces during reinforcement learning. According to the authors, it solved 27.6 % of tasks on SWE-rebench, 6.4 percentage points more than interface search alone.
The study introduces the DocSkill framework and the MMHop-DocBench dataset with 60 thousand question-answer pairs in 10 languages. Using Qwen3-VL-8B, the framework achieved an accuracy of 41.0 and an F1 score of 49.2, according to the authors, exceeding the best-performing baseline evaluated by 5.5 and 6.3 points, respectively.
The Fin-SD method increased accuracy for GPT-OSS-20B on FinanceReasoning Hard by 3.03 percentage points. Generic feedback achieved the same improvement. According to the study, what mattered was where in the process the intervention occurred, rather than the content of the hint.
The study introduces the AutoVend benchmark: one agent manages 20 vending machines over 180 simulated days. Four of the five agents tested underperformed rules-based inventory management. Accounting for inference costs changed the ranking of models by economic performance.
The OA-RAG method combines formal reasoning over ontologies with LLMs and provides evidence alongside a model-based assessment of how well answers are supported. According to the authors, it achieved the best overall results among the methods compared in tests involving three ontologies and six medical tasks.
The study proposes weighted uncertainty calibration when the control policy changes, without modifying the model. In an experiment using trajectories from a Franka robot, prediction set coverage at a nominal level of 90 % reached 90.9 %, according to the authors, while standard calibration achieved 55 %.
The study compared 15 pairs of base and instruction-tuned models. In every case, instruction tuning made number selection more uneven. The tuned models almost never selected values at either end of the range or multiples of ten.
The authors trained Icelandic encoders with 12 million to a billion parameters on 9 billion tokens. According to the study, the billion-parameter model outperformed IceBERT on all 11 tasks; initialization from Ettin yielded the best results on all evaluated tasks.
The study introduces CAMR, a method that selects both new and past examples for continual learning. In tests on Qwen3-8B-Base, the authors report an Original Match score of 93.66 % on the WikiBigEdit dataset after 3 000 facts and the best final results on four selected TRACE tasks.
The authors introduce OmniBrain-0, a model that combines planning, task tracking and error recovery in robotic manipulation and navigation. It uses distillation from multiple teacher models; according to the authors, it has lower operational overhead than the closed teacher model.
The authors present a speech language model for Basque based on the LLaMA-Omni architecture and the Latxa model. They used translated data and synthetic speech for training. They report a content rating of 3.86/5, a transcription word error rate (WER) of 9.94 % and a UTMOS audio quality score of 3.07.
MIT researcher Paris Smaragdis works on automated sound source separation, modifications to acoustic environments and tools for working with music. The available excerpt introduces his research focus and career path, rather than the results of a specific new study.
A September survey of 972 New Zealand residents found that 57 % expect to use AI in the future. The misuse of generated images, video or audio is very or extremely concerning to 74 %. The results were weighted to reflect the demographic makeup of the population.
Researchers at Google Research described alternating between interaction and offline weight updates using their own training examples. According to the authors, the sleep framework reduces forgetting, retains knowledge from contexts of up to 10 million tokens and improves reasoning.
Ai2 has released a study, a benchmark, data and code for evaluating the originality of research ideas using LLMs. According to the study, changing a few words in the prompt shifted one model's accuracy by more than 50 percentage points; literature searches did not consistently help.
Thore Graepel presented a proposal from Metis Reasoning: store knowledge outside the model weights and check assumptions with a separate module. According to the company, the approach is intended to reduce hallucinations, but so far there is only a theoretical framework with no working system.
According to the available excerpt, the model GPT-6 Astra deciphered a military letter from 1809 about army positions before the Austrian attack on Bavaria. The operator of the Cryptiana website reportedly confirmed that the solution was correct.
The study introduces Recursive Semantic Divergence to estimate discrepancies between runs of the same agent. According to the authors, using the metric as a reward during fine-tuning reduced variability in task execution from 132 % to 91 % in tau2-bench and from 90 % to 44 % in toolsandbox.
AI Radar monitors Czech and international sources every day, looking for changes that truly deserve attention.
MonitorsOfficial AI company blogs, specialist media, and research sources.
Selects and combinesFilters out information noise and combines articles about the same change into a single event.
Summarizes and explainsExplains significant events in English: what happened, why it matters and where the information comes from.
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.
Everything you need to navigate the AI world
Today’s briefingAn overview of news from the past 24 hours and a main story when available. Beside them is Live · AI Flash, followed by research and links to other Radar sections. On mobile, the blocks appear one below another.
AI FlashAn ongoing feed of brief updates with an evidence status. Links lead to a Radar detail page when one is ready, otherwise to the original source. You can also find reset and outage histories here.
Practical applicationsWhat new tools and features can do, what you can try and what their actual impact could be.
Model selectionModel comparison by type of work, capabilities, price and speed.
Research and archiveA separate research overview, topic search and older events by date.
One event, everything that matters
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.
Importance: ▮▮▮ majorOpenAIModels✓ 6
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.
▮▮▮ major · ▮▮ important · ▮ we're tracking = how significant the change is✓ 6 = six independent publishers, not the number of articles or feeds✓ official = a clear release, law or incident is substantiated by the relevant authority1 source = no independent confirmation yetbold = who is behind the changegray text = a brief summary of what happened
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.
The same news, two practical uses
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.
For individualsWhat you can use or try, how the change can help you at work and what to watch out for.
For businessesWhat impact the change could have on processes, costs, risks and other business decisions.
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.