Skip to content

Research separate from news

What could become important next

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.

Research only one source so far

Neural networks for modeling hysteretic systems - a survey from Prandtl to neuroplastic models

A survey article on arXiv cs.LG examines neural networks for modeling hysteretic systems. It covers Prandtl neural networks, Preisach neural networks, and their extensions, including the neuro-Madelung model. It focuses on memory representation, rate-dependence, and applications to nonlinear structural and…

arXiv cs.LG (Machine Learning) Original source ↗
Research only one source so far

Machine learning outperforms FLOP estimates in predicting training time

Hugging Face analyzed 272 production fine-tuning jobs (7382 GPU-hours) and found that standard FLOP-based time estimates have an R² of 0.31. A CatBoost regressor achieved an R² of 0.98 (MAE 1.71 h) by incorporating evaluation cadence, checkpoint policy, and I/O factors that simple FLOP calculations ignore.

Hugging Face Blog Original source ↗
Research only one source so far

Harvard study: Half of the electricity used by the largest US data centers comes from gas and coal

A Harvard study published in Nature Sustainability found that approximately 403 of the largest US data centers produce as much CO2 annually as Denmark. They consume 82 TWh of electricity (1.8% of total US consumption), of which 54% comes from fossil fuels. Emissions amount to 25.7 million tons of CO2…

The Decoder (daily AI news) Original source ↗
Research only one source so far

Gallup study: a third of young people use AI companions, feel lonelier

A Gallup survey with the Boston Children's Digital Wellness Lab among 16 000+ respondents (aged 15–22) from 16 countries shows that 33% of young people use AI as a companion, with usage higher among people who have only online friends. Users of AI companions report higher levels of loneliness; the study does not show whether AI causes the loneliness…

404 Media (investigative tech journalism) Original source ↗
Research only one source so far

Test-time adaptation method increases the accuracy of AI-generated text detectors

Researchers from MIT and Cornell have developed a test-time adaptation method that increases the accuracy of AI-generated text detection from 27% to 93%, even for masked text. The method allows the detector to learn from new data without retraining, with applications in education, publishing, HR, and cybersecurity.

MIT News – Artificial intelligence Original source ↗
Research only one source so far

TUSI-Bench: a benchmark for evaluating RAG systems in Persian law

The study presents TUSI-Bench, a benchmark for evaluating RAG systems in Persian law. The benchmark contains 7 261 legal articles and two evaluation tasks: TUSI-QA (1 999 questions) and TUSI-Bar (500 questions). The proposed graph-based retrieval improves recall by up to 8 percent.

arXiv cs.CL (Computation and Language / NLP) Original source ↗
Research only one source so far

Regularized score centering for training off-policy agents

The Regularized Score Centering (RSC) method improves training of off-policy AI agents by combining score centering and KL regularization. Experimentally, it achieves +14.3% on AppWorld and a 9.9x higher score on BabyAI with lower sample consumption.

arXiv cs.LG (Machine Learning) Original source ↗
Research only one source so far

MedGuardBench: a benchmark for personalized safety rules for medical AI

A research paper introduces MedGuardBench, a benchmark with 5 001 scenarios and 87 drugs for personalized safety rules for medical AI. Personalized safety rules with three components are developed, which reduce false alarms and outperform base models in accuracy by taking into account individual…

arXiv cs.AI (Artificial Intelligence) Original source ↗
Research only one source so far

MM-ReportBench: A Benchmark for Evaluating Multimodal Report Generation from Tables

A research team introduced MM-ReportBench with 386 tasks across 185 real-world tables for evaluating multimodal models that generate reports. A dual-judge framework verifies text quality, factual correctness, and visual fidelity. The best of 20 tested models achieved a score of 68.9, with 38.5% of reports containing…

arXiv cs.CL (Computation and Language / NLP) Original source ↗
Research only one source so far

The Shared Word Benchmark: Evaluating Language Models in Dutch

The new Dutch benchmark The Shared Word (TSW-NL) contains 222 word puzzles where three hints describe different senses of the same word. It was created through curation from the Dutch Wikipedia. Humans achieve 49.5–61%, the best open-weight model 39.3%, the GPT-4.1 model 59.2%, the GPT-5.6 Sol model with maximum reasoning 88.8%.…

arXiv cs.CL (Computation and Language / NLP) Original source ↗
Research only one source so far

Framework for multimodal scientific documents with integrated retrieval and reasoning

The Retriever-Perceiver-Reasoner framework combines evidence retrieval, its analysis, and deduction for multimodal scientific documents. A new training process with diagnosis-guided development and targeted feedback mechanisms (RPF, LEF) achieves 69.9–75.0 points on benchmarks, an improvement of 4.0–9.1 points over…

arXiv cs.AI (Artificial Intelligence) Original source ↗
Research only one source so far

Comparison of fine-tuned BERT and open-weight models on Japanese clinical entity extraction

The study compared BERT-CRF with nine open-weight LLMs on entity extraction from Japanese clinical texts (iCorpus, 113 categories). Zero-shot LLMs achieved an F1 of 0.337 compared to BERT's 0.836–0.841. However, after QLoRA fine-tuning, LLMs with 13B+ parameters outperformed BERT: relaxed F1 0.882–0.914. Increasing the output limit improved…

arXiv cs.AI (Artificial Intelligence) Original source ↗
Research only one source so far

PhysScope: Improving the Physical Reasoning of AI Agents with Visual Memory

The PhysScope method improves the physical reasoning of vision-language models without updating their weights, by linking physical concepts with visual observations. With the Gemini-3.8-flash model, it outperforms Canvas-CoT by 9.34 percentage points on the PhysAlign benchmark and improves OOD accuracy on PhysUniBench.

arXiv cs.AI (Artificial Intelligence) Original source ↗
Research only one source so far

DeltaHarness: Self-optimizing tool execution planning in AI agents

The study presents the DeltaHarness method, which adapts tool execution strategies at runtime based on initial measurements. The method consists of a QHI loop that observes runtime behavior, updates dispatch, and selects experiments. In tests on video-editing tasks, it reduced latency by 30% on…

arXiv cs.LG (Machine Learning) Original source ↗
Research only one source so far

How to let AI help in education without losing your judgment

Editorial essay on the use of generative AI in STEM education, referencing the monograph "STEM Education: Shaping Future Learning Practices in the Age of AI". The author emphasizes the need to prepare teachers to use AI as an "interactive sparring partner" according to the seven pillars of effective professional…