AM-IA brings together news, research papers and applications relating to artificial intelligence published around the world into a single feed — and makes them available in your European language at the click of a button.
Scroll to browse · Click to filter
Nature looks back at an experiment in which an AI reconstructed existing research concepts, and the harsh judgment made by the researchers behind this work.
The EpicStar framework combines working memory with a bank of winning episodes; on StarCraft II, it wins more often by consuming an order of magnitude less tokens.
By replacing the single judge with a moderate deliberation between several models, the method goes beyond frontier models for 8 to 15% of their cost.
Out of 1,661 attack cases and 65.69% average success, the authors uncovered a gap between recognizing a security constraint and sticking to it.
Researchers are exploiting the interchangeability of encrypted blocks within the same provider to force a less protected model to render the reasoning of an advanced model in plain text.
By replaying the experiments of 168 papers presented at a machine learning conference, AI agents were only able to confirm a minority of the claimed results.
An arXiv paper tackles the accumulation of errors in world models that make step-by-step predictions, by allowing the length of each prediction step to adapt.
A study estimates between 731 and 1,125 million m3 the annual water requirement of the IA server fleet in the United States, and 24 to 44 Mt CO2-equivalent associated emissions.
A study of eight languages shows four regimes of linguistic coherence, with a brutal collapse on non-Latin languages and the least represented.
Placed in an examination situation under common time constraints, they behave like gluttonous sequential solvers: they exhaust themselves on the first questions and ignore their value.
A Stanford study published in Nature Human Behaviour crosses questionnaires and real conversation histories: it is not intensive use that harms, but confidential use in people who are already poorly supported.
Over 26 scenarios and 204 episodes, the benchmark turns memory on and off to verify that the progress observed is going through the intended channels — which is not always the case.
Across five domains and a hundred subtasks linked together, the libraries of skills contribute little compared to the simple context: the gains come from adaptation, not from robust abstraction.
The benchmark evaluates the customization of agents from long interaction histories rather than frozen profiles, and reveals a "gaps between knowledge and action".
Marta Garnelo and Wojciech Czarnecki rule out four common explanations and show that a leading model is the only one, on nine methods, whose accuracy drops when the number of dimensions increases.
The F-ICL test bench compares the induction in the context of the models to an exact computable standard: the accuracy of the responses masks an inductive bias that the scale does not correct.
The benchmark covers three types of graphs, four task families and 84 executable tools. Authors' conclusion: the quality of tool calls counts more than their number.
A study explores what children understand — and don't understand — about the true environmental cost of the artificial intelligence systems they use every day.