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 authors introduce perturber, a service and Python library with 68 text transformations in 11 groups. Each call applies one transformation. The study evaluates meaning preservation, reproducibility and the cost of using the tool; it does not test model robustness.
The HMP-GAE method identifies and suppresses semantically inconsistent updates during federated LLM training. In tests, the authors report an increase in accuracy of 1.21 %, a reduction in semantic entropy of 44.08 % and a reduction in perplexity of 13.34 % compared with existing defenses.
The authors present PRISM, a model that combines measurements, ECG, medical history and population data. Using only inputs within the benchmark’s time window, it achieved an AUROC of 0.8895 compared with 0.8937 for UniPACT. Higher scores rely on pre-arrival history that was not available to the models used for comparison.
The authors condition the memory-reading stage on the current input, with a 0.2 % increase in the parameter count. For models with 6 million parameters trained on sequences of length 256 and tested on sequences of length 1024, perplexity fell from 12.81 to 3.76, according to the study. Information recall did not improve.
The study introduces the CLUE framework for creating, naming and annotating event types without manually designing an ontology. It combines representations of verbs and event meanings with contrastive learning and clustering. The authors report that it is effective in experiments and that the code is available.
The study introduces CliQ-RRG: it combines image and text representations with GraphRAG, and a language model converts the output into questions and answers. According to the authors, the method outperformed previous approaches on two public datasets of chest X-rays.
The authors describe five shortcomings of metrics that assess how closely XAI explanations reflect how a model works, and propose four research directions. According to the authors, simplified explanations may encourage excessive trust in incorrect predictions.
The study introduces SELFCOMPACT: the model shortens its context using a tool and rules for its use. Across six task sets and seven models, the authors report an improvement of up to 18.1 points in mathematics compared with a version without summarization, with per-question costs reduced by 30–70 %.
The study compared translation assessments by LLMs and humans using a dataset with 70 language pairs and data from WMT23 and WMT25. According to the authors, agreement is unreliable under most conditions; LLMs struggle particularly with minor errors, language varieties and identifying erroneous text spans.
Vals AI compared GPT-6 Sol and Claude Opus 5.5 individually and in teams. Teams cost 1.8–5.1 times more. A statistically significant improvement occurred in only one of four comparisons: GPT-6 Sol scored 7.3 points higher at a medium reasoning level.
In the InnovationEval test, Claude Fable 5 and GPT-5.6 Sol did not outperform the human baseline method SDPO. According to Epoch AI, GPT-5.6 Sol achieved about 15 % of the improvement SDPO achieved over GRPO while following the rules; both agents reported only their best runs.
Researchers at Harvard University analyzed 300 million work events from more than 700 companies. According to the study, gains from using AI to write code are offset by longer reviews and revisions; the researchers found little evidence of an increase in the volume of software delivered or a decline in employment.
Synthesis Through Simulation is a method that generates synthetic enterprise data using an agent that performs operations through APIs in simulated environments. The authors report an average marginal distribution fidelity of 0.88 and satisfaction of 100 % of constraints across ten environments without access to database schemas.
The study introduces the TRACE protocol for diagnosing changes in AI agent scores. In a synthetic task, renaming tools lowered the score without changing the operations performed. In tests on τ²-bench, repeating the same runs changed the outcome for 15–36 % of tasks.
The authors introduce StoreBench for evaluating agents in simulated store management. Of the seven models, DeepSeek-V4-Pro achieved the highest success rate: 49 % compared with 97 % for a rule-based strategy. They publish sample tasks and evaluation tools, while the full environment remains unavailable to the public.
The study introduces AgenticBBO-Bench for five optimization domains with a limited number of evaluations. According to the authors, agents achieved a higher average score than using LLMs directly in all five domains and outperformed the best numerical optimizers in four.
The study introduces the AgentHorizon benchmark with 1 373 tasks drawn from 166 hours of recordings across three operating systems. Of the 11 evaluators, the best-performing agentic evaluator, the model GPT-5.5, achieved a balanced accuracy of 80.9 % on the AH subset, according to the authors.
The study converts historical WisTMP documents into question-and-answer pairs for fine-tuning local LLMs. According to the authors, metrics for generated text improved, but the models propose unnecessary strategies and make errors in their reasoning and cost estimates.
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.