1 paper · 1 filter
Michal Štefánik, Marek Kadlčík, Josef Kuchař +1
Beyond scaling their parameters and data, large language models currently gain versatility on new problems along a single axis: the tokens they spend on chain-of-thought (CoT). We…