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.
The research paper proposes a continual learning method that consolidates knowledge without an offline phase using biologically inspired mechanisms (isolation rule, refractory rotation). On the split-MNIST benchmark, it achieves 91.6 % accuracy, comparable to or better than experience replay, ER-ACE and A-GEM.
Research examines whether reducing model deceptiveness via compensatory feature injection improves refusal of harmful requests. On Qwen 3.5 models, they found that a decrease in deceptiveness does not guarantee stronger direct refusal, but under user pressure some versions partially recover (up to 95%).
An Apple Machine Learning Research study tests how well language models serialize tree expressions into natural language and recover them. Results: the channel is asymmetric (up to a 60.4 point difference in accuracy), the best pair achieves 92.9%, 73.6% of failures come from generation. Training on ~3600 examples…
Microsoft Research Asia – Singapore, opened on 24 July 2025, has published an annual review of its activities in advanced AI research. The institution collaborates with local partners on multimodal healthcare AI, AI agents, and applications for financial services and education. The focus is on translating fundamental research…
The article explains three main challenges of deploying AI vision from the lab into the real world: models trained on high-quality, controlled data fail under unpredictable conditions; specifically, pose estimation performs excellently in light but fails in darkness, because real-world data with low…
Researchers including Yann LeCun are developing world models — AI models for physical reality — trained on a combination of video and action data. Due to a lack of such data, they are using video game material with millions of games containing control inputs. The startup Worldmodeldata is attempting to be…
MIT researchers used an AI algorithm to optimize the lipid nanoparticle formulation in RNA vaccines. The resulting vaccines remain stable at room temperature for up to a year, or at around 38°C for two months. When tested in mice, they elicited the same immune response as a Moderna-like vaccine.
A study criticizes the measurement of interpretability during model quantization. Cosine similarity is reported without the parameters (n, ρ) needed for interpretation. On Qwen2.5-1.5B-Instruct: for INT4 the direction rotates beyond noise, for INT8 no change is detected. Released with code and data.
A research preprint presents the READ method for composing multiple LoRA adapters into a single model without interference. New adapters can read the inputs of old ones, but not write their outputs. On SuperGLUE, an improvement of more than 20 points; on the domain suite, more than 7 points.
Research reveals that outdated documents in RAG cause errors in 30-37% of Llama and Qwen model responses without instructions; with an explicit instruction, up to 66-75%. The study tests 12 models on a benchmark with 317 verified knowledge reversals in medicine, law, and software. The solution is information about temporal validity.
Research proposes autoencoder-based mixing with iterative refinement as an alternative to the attention mechanism in transformers. The new approach uses 1.9× fewer FLOPs than attention while maintaining performance at the level of BERT and TinyBERT on the C4 dataset.
The paper presents Self-Play Search Distillation (SPSD), a method for generating high-quality synthetic training data through self-play simulations on board games. SPSD leverages MuZero-like networks and produces structured chains of thought. On the Qwen3-4B-Base model, it increases performance in…
DeepEdu-v1 is an AI tutoring system for Vietnamese education, built on the SCALE framework with optimizations for long context and a self-improving agentic layer. It increases accuracy from 70% to 79.5% on complex tasks, achieves a 2x TTFT speedup compared to vLLM on consumer GPUs, and ensures on-premise…
Introduction of the KNOWS benchmark for evaluating web agents on complex, open-ended tasks. Agents achieve <3% success rate on full tasks; failures in visual steps prevent outputs, even when agents complete >50% of the other steps. The benchmark identifies limitations in tool use, visual understanding and…
The scientific review covers 211 studies (2018–2026) on fake review detection. It tracks the evolution from traditional machine learning through PLM models to LLM-based approaches. It analyzes eight types of evidence sources (text, sentiment, behavior, metadata, graphs, multimodal content). It identifies open challenges in adversarial…
The study presents the ViSTA adapter (0.516M parameters) for integrating clinical measurements into pretrained vision-language models. On MIMIC-IV, it achieves an AUC of 0.7376 for predicting acute kidney injury with 2B parameters (comparable to GPT-5.6 Sol) and 69.27% accuracy on temporal questions with 90% fewer…
A research team presented LUMO, an offline voice assistant running on a Raspberry Pi 5 with 8 GB RAM. The system combines local ASR, a 4-bit GGUF quantized LLM, and TTS. It achieves a WER of 6.8%, latency of 2–4 s, and power consumption of ~9 W. It supports English and Bengali without a cloud connection.
A study on arXiv analyzes where LLM-generated medical answers verified against authoritative sources fail. The authors induce a taxonomy of retrieval errors (5 dimensions) and reasoning errors (6 steps) using LLM-as-Judge, tested 4 retrieval methods and 6 frontier models, and found that scaling…
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
