Quando si parla di capacità di ragionamento nei large language model , l’associazione immediata è con modelli addestrati a posteriori tramite reinforcement learning (RL). Il messaggio dominante è: per risolvere compiti complessi, servono training complessi. Ma se fosse solo una questione di come campioniamo, non di cosa alleniamo? È questa la tesi – controintuitiva ma elegantemente dimostrata – del recente studio “ Reasoning with Sampling ” di Aayush Karan e Yilun Du (Harvard, 2025). Il paper mostra che i modelli base , se interrogati con il giusto algoritmo di sampling, possono eguagliare o superare le performance di modelli RL-posttrained su compiti di ragionamento strutturato. Senza alcun training aggiuntivo. Vediamo perché questo è importante, e soprattutto come funziona . Oltre il reinforcement: il problema della “distribution sharpening” Il reinforcement learning con ricompense verificate (RLVR) ha trasformato LLM come ChatGPT in abili risolutori di problemi matem...
Takamama is coming! (ENG) Today we have begun the approach phase and will finally be online in just a few days. To explain Takamaka we have created a series of videos to tell you about technology, protocol, blockchain and Smart Contracts.
We have delegated at Fausto Spoto, professor at the University of Verona and Iris Dimni Ing. at Takamaka the task of illustrating the project with a series of short videos, 5 to 30 minutes. We will talk about: - What is Takamaka - How to complete the Blockchain - Consensus algorithm # T-PoS - How to write a smart contract (practical example). and much more. The content will be in english and published in episodes on www.takamaka.io. Keep in touch and ask your questions on www.takamaka.io talk ethereum publisher project datascience technology business questions series
https://www.linkedin.com/company/takamaka-enterprise-blockchain/?viewAsMember=true
Follow @AndreaBelvede10We have delegated at Fausto Spoto, professor at the University of Verona and Iris Dimni Ing. at Takamaka the task of illustrating the project with a series of short videos, 5 to 30 minutes. We will talk about: - What is Takamaka - How to complete the Blockchain - Consensus algorithm # T-PoS - How to write a smart contract (practical example). and much more. The content will be in english and published in episodes on www.takamaka.io. Keep in touch and ask your questions on www.takamaka.io talk ethereum publisher project datascience technology business questions series
https://www.linkedin.com/company/takamaka-enterprise-blockchain/?viewAsMember=true













