Updated daily396 verified stories
Material AI papers, studies, datasets, evaluations, and scientific advances with direct links to the underlying research.
Every story links to the strongest available evidence: original sources when available, otherwise clearly attributed reporting.
What happened, why it matters, and what to watch — no jargon tax.
When the signal is thin, we publish nothing rather than padding the feed.
A growing stream of verified perspectives for people who need to understand AI without chasing hype.
An arXiv paper introduces Relative Transfer, a measure of how training examples at different difficulty levels affect one another, and proposes Transfer-aware Dynamic Curriculum Sampling for adapting training schedules during post-training.
An arXiv preprint describes Co-RL, a multi-agent reinforcement-learning framework in which separate models reward one another. The authors report gains across text-only and multimodal benchmarks without ground-truth labels, while acknowledging the risks of self-reinforcing errors and training collapse.
A new arXiv paper proposes Task Specialization Fine-Tuning to allocate limited fine-tuning across related task regions. It reports improved task coverage in optimization, control, and LLM experiments, but gives no numerical results or independent validation.
Accepted to the 2026 Asilomar Conference, a paper proposes SURE-Ridge, a closed-form method for causal-graph estimation in equal-variance linear Gaussian models. Authors report lower small-sample structural error and fastest tested-baseline runtime; the source does not establish performance beyond these assumptions.
Researchers describe J-Miner, a method that converts internal signals from fine-tuned language-model classifiers into inspectable rules, reporting up to 98.3% reproduction of source decisions and nearly the same average task accuracy in much smaller student models.
A single-author arXiv paper reports that giving an agent a tool to switch between specialized LoRA adapters helped it solve two synthetic coding tasks and reduced the reported capability tax by up to 18 times.
An ECML PKDD 2026 paper presents MultiSigBERT, combining narrative medical reports, structured records and temporal patient data to estimate individualized oncology risk scores. It reports a 0.743 concordance index on an independent test set; the source has no comparator results or clinical-deployment evidence.
An arXiv paper proposes optimizing multi-layer model edits through the complete forward pass, reporting the strongest combined results among tested methods in five of six model-dataset settings.
An arXiv paper reports that adversarial examples generated through publicly released EEG encoders can affect inaccessible downstream models without queries to their parameters, outputs, or gradients.
An arXiv paper introduces Mr.Dec, a model that combines daily electronic health-record updates with intermittent chest X-ray findings in chronological order. The authors report state-of-the-art results on two MIMIC datasets, but the supplied record provides no scores or evidence of clinical deployment.
A new benchmark argues that document-sensitivity classifiers can appear more accurate than they are when residual labels leak into the text. On its cleaned 16,000-document corpus, BERT reached 89.14% accuracy, while logistic regression offered a cheaper alternative.
An arXiv paper introduces Data-DPO, which selects supervised fine-tuning examples using a target model’s responses during brief training probes. Authors report better results than existing selection methods and full-data training on Vision-Flan and LLaVA-CoT, but the record lacks effect sizes or independent validation.
One useful briefing each week
Get the week’s verified AI news, original data, useful tools, learning picks, and fresh AI jobs.
Hiring an AI professional or launching a useful AI product? Put it in front of people who came here to learn and act.
Post an AI job Submit an AI tool