[Bibliothèque numérique]
Les petits enfans de Marceline Desbordes-Valmore en ligne sur Wikisource et la bibliothèque numérique !
Retour sur l'histoire d'une collaboration en 2024 avec @Wikimedia_Fr et la wikipédienne @dmontagne 👉 https://tinyurl.com/peyf54ed
Consulter l'ouvrage sur la BibNum de la BDL : https://tinyurl.com/489x2eh4
Large language models exhibit a puzzling inconsistency: they solve complex problems yet frequently fail on seemingly simpler ones. We investigate whether LLMs internally encode problem difficulty in a way that aligns with human judgment, and whether this representation tracks generalization during reinforcement learning post-training. We train linear probes across layers and token positions on 60 models, evaluating on mathematical and coding subsets of Easy2HardBench. We find that human-labeled difficulty is strongly linearly decodable (AMC: $ρ\approx 0.88$) and exhibits clear model-size scaling, whereas LLM-derived difficulty is substantially weaker and scales poorly. Steering along the difficulty direction reveals that pushing models toward "easier" representations reduces hallucination and improves accuracy. During GRPO training on Qwen2.5-Math-1.5B, the human-difficulty probe strengthens and positively correlates with test accuracy across training steps, while the LLM-difficulty probe degrades and negatively correlates with performance. These results suggest that human annotations provide a stable difficulty signal that RL amplifies, while automated difficulty estimates derived from model performance become misaligned precisely as models improve. We release probe code and evaluation scripts to facilitate replication.
November und Dezember sind Pingbackwochen. Schreiben und linken, um zu vermitteln.
https://osl.hypotheses.org/15111
#OpenAccessWeeks #OhneAccessWeak #LinkedOpenStorytelling #OpenAccessWeek
Did you miss our #OpenAccessWeek webinar? Our team shared how you can use KCWorks to disseminate your scholarship and publications...AND gave a sneak peek at a major system upgrade that'll be launching very soon. You're not going to want to miss it!
Le compte Mastodon animé par la mission Services aux chercheurs de la DiBSO, créé en février 2025, fête son 100e abonné (et même 101e dans la foulée) après huit mois d’existence. Le compte d'Université Côte d'Azur est en contact avec ceux d’autres institutions, comme Recherche Data Gouv, les comptes Science ouverte de Saclay et de Rennes, l’INIST, le CoSO ainsi que de nombreuses bibliothèques.
Merci à @teinturs d'être la centième à nous rejoindre, et également d'avoir participé au café virtuel de l'Open Access Week !
La scienza deve decentralizzare. Il post di Rory Mir per @eff in occasione della #OpenAccessWeek
La produzione di conoscenza non avviene nel vuoto. Ogni grande scoperta scientifica si basa su lavori precedenti e su uno scambio continuo con i colleghi del settore. Ecco perché dobbiamo affrontare la minaccia che i principali editori e piattaforme possano influenzare in modo improprio le modalità di accesso alla conoscenza scientifica, o addirittura sopprimerla.
Quando l'infrastruttura è troppo centralizzata, i gatekeeper acquisiscono nuovi poteri per catturare, manipolare e censurare. Il risultato è un sistema che diventa meno utile, meno stabile e con maggiori costi di accesso. La scienza prospera grazie alla condivisione e all'equità di accesso, e il suo futuro dipende da una rivolta globale e democratica contro le piattaforme centralizzate predatorie.
https://www.eff.org/deeplinks/2025/10/science-must-decentralize
Knowledge production doesn’t happen in a vacuum. Every great scientific breakthrough is built on prior work, and an ongoing exchange with peers in the field. That’s why we need to address the threat of major publishers and platforms having an improper influence on how scientific knowledge is accessed—or outright suppressed. Infrastructure we rely on must be built in the open and on interoperable standards, and hostile to corporate (or governmental) takeovers. Universities and the science community are well situated to lead this fight.
Heute geht die #OpenAccessWeek zu Ende.
Die #OpenAccess Bewegung hat viel erreicht über die Jahre.
Denn es ist zum Standard geworden, dass die Ergebnisse mit öffentlichen Geldern finanzierter #Forschung natürlich kostenlos der #Gesellschaft zugänglich sein müssen:
👉 https://de.wikipedia.org/wiki/Open_Access
Wir haben aber noch einen Stück des Weges 🛤️ vor uns.
Viele #Hochschulen und #Forschungsinstitute berichten (für die Öffentlichkeit!) auf Kanälen, die nur zu nutzen sind, wenn man umfangreichen #AGBs und #Nutzungsbedingungen zustimmt.
Hier auf #Mastodon, #PeerTube und #Pixelfed ist das ganz anders, vorbildlich im Sinne von OpenAccess:
denn jede Meldung hier ist über das Internet frei abrufbar – ohne langatmige AGBs, die nur kaschieren, dass die Anbieter der Dienste alle Daten vom Nutzer abgreifen, derer sie habhaft werden.
Schon deshalb gehören die #UnisInsFediverse, wie in unserem Appell formuliert
👉 https://openpetition.de/UnisInsFediverse
This is the best thing I have read about #OpenAccess in years. It's a far-ranging journey into so much of what is wrong and could be better in academia and in our world. Plus Hothead Paisan.
Seriously, read it this weekend. https://matthewcheney.net/blog/into-the-fugitive-open/
The theme for this year’s International Open Access Week is “Who Owns Our Knowledge?”, a question that ought to be nonsensical, but which is becoming ever more urgent. I cannot help but think about Open Access (OA) within the current context both of our world generally and also my own world personally. The general world’s disasters […]
As we close out #openaccessweek read about our work to strengthen collaborations between the Open Science & Wikimedia communities. This year we've had the privilege of working with Wikimedia affiliates to conduct a survey to map interest in Open Science. The survey found that there is strong interest in collaboration which inspired us to begin to develop an Open Science Wiki Community, based upon the long-standing and successful GLAM Wiki Community.
🔓 Deuxième formation doctorale aujourd'hui dans le cadre de l' #openaccessweek avec un atelier pour accompagner les doctorant.e.s à la publication dans HAL. En tout cette semaine, plus de cinquante étudiant.e.s auront été familiarisés avec l'écosystème de la #scienceouverte les modèles de publications en #openaccess, les revues prédatrices, les accords transformants et la prise en charge des APC, à la création d'IdHAL et à la création de CV #HAL
La semaine prochaine, place aux cafés virtuels, ouverts à tous, autour de la publication en open sans frais et les gains de visibilité grâce à l'open access. Les liens sont disponibles ici : https://bu.univ-cotedazur.fr/services-aux-chercheurs/science-ouverte/programme-de-lopen-access-week
Bibliothèque Diderot de Lyon
N-gated Hacker News
Jens Bemme
Knowledge Commons
Science Ouverte UniCA
macfranc
Aktionsb. neue soziale Medien
Robin DeRosa
BOAI