Mis à jour quotidiennement1796 histoires vérifiées
Couverture par l'IA vérifiée à la source des lancements de produits, des changements de politique, de la recherche sur la sécurité et des évolutions de l'industrie, expliquée dans un anglais simple par une équipe éducative à but non lucratif.
Chaque article est lié aux preuves disponibles les plus solides : sources originales lorsqu'elles sont disponibles, sinon reportages clairement attribués.
Que s’est-il passé, pourquoi c’est important et que regarder – sans le jargon.
Lorsque le signal est faible, nous ne publions rien plutôt que de compléter le flux.
Histoires d'IA vérifiées à la source, les plus récentes en premier, pour les personnes qui ont besoin de comprendre l'IA sans courir après le battage médiatique.
Routed Graph Handoff uses a lightweight language-model router to choose between structured dependency graphs and natural-language messages when AI agents delegate tasks.
A new benchmark evaluates whether language models can generate and improve GPU code for database-style queries, with the paper reporting speedups of up to 2.11× on one H100 GPU and 2.54× across four H100 GPUs.
A paper accepted at EMNLP 2026 reports that flipping routing-layer bits in mixture-of-experts language models can sharply inflate output length, potentially creating an availability and inference-cost risk.
An arXiv preprint proposes fixing a report’s evidence, numbers, directions and permitted wording before an LLM generates connective prose. The authors report higher cross-run reproducibility than a hybrid template in fMRI and randomized-trial reporting, alongside lower observed token use and median latency in one…
A newly submitted paper presents ClueWeaver, a reward-trained framework that divides long-narrative question answering between evidence retrieval and interpretation in compact, locally deployable language models.
A new arXiv paper proposes Golden-GRPO Injection, a three-stage self-learning framework intended to help large language models absorb newly injected knowledge and apply it across paraphrases, combined documents and reasoning tasks.
A new preprint reports that BanglaMamba, a state space model for Bangla fake-news detection, uses less GPU memory and processes data faster than tested BERT-based systems, while trailing BanglaBERT on accuracy and external-dataset performance.
A new preprint reports that direct messages and peer relays can shift the stances of language-model agents, and argues that evaluations should track exposure histories and actions, not text alone.
A new preprint reports that several normalized metrics used to compare language models across languages can introduce biases linked to tokenization, encoding and orthographic differences. It argues that sentence-level negative log-likelihood over semantically equivalent sequences produces more consistent comparisons.
A new dataset evaluates large language models as diagnostic agents in interactive clinical encounters rather than through isolated questions. MTDiag combines cases from emergency medicine, clinical records and published case reports, and proposes knowledge-grounded metrics beyond diagnostic accuracy.
Researchers created HealthBench-Psych, a 610-conversation mental-health subset of OpenAI’s HealthBench, and released its evaluation pipeline, model responses, grades and analysis code.
A survey accepted to Findings of EMNLP 2026 catalogs 80 methods for adapting foundation models without human labels, preference data, stronger teachers or executable verifiers. It warns that internal learning signals can either improve a model or recursively amplify its errors.
Un briefing utile chaque semaine
Recevez les actualités vérifiées de la semaine sur l'IA, les données originales, les outils utiles, les choix d'apprentissage et les nouveaux travaux d'IA.
Embaucher un professionnel de l’IA ou lancer un produit d’IA utile ? Présentez-le aux gens qui sont venus ici pour apprendre et agir.
Publier une offre d'emploi IA Soumettre un outil d'IA