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
Probability, wave equations, group theory: Meta has published six articles written using its Muse Spark model, which is used in Meta AI’s simple chat interface.
Out of nearly 300 questions, the ‘digital twins’ of panellists generated by Claude Opus 4.6 do not accurately reflect public opinion, particularly amongst Republicans and Black Americans.
By training 800 models with varying proportions of AI-generated text, researchers have shown that a synthetic token can be helpful at times and harmful at others, and have proposed a scaling law that takes this into account.
Out of 2,000 cases across the cantons between 2020 and 2023, the recommendation optimised using machine learning increased the three-year employment rate by 5.2 percentage points.
The systems already write most of the code in the laboratories where they are developed. The authors believe that the entire research pipeline could be automated within a few years.
A team comprising Anil Seth, Murray Shanahan and Marcus Hutter proposes five levels of description to contextualise theories of consciousness, and then evaluates current models. The outcome depends largely on the premises.
Moral evaluations assess decisions in isolation. By providing the models with an agent’s history, researchers have identified an internal mechanism that allows them to influence the agent’s choices at runtime.
A benchmark measures whether an agent assigns each stored item the weight it deserves. State-of-the-art models tend to overuse or underuse their memory rather than adjusting it.
A set of entirely synthetic longitudinal medical records, containing no sensitive data, from which ten models are able to retrieve the information but fail to reconstruct it.
Of the six harnesses tested, only one prevents the officer from deleting their log. Yet audits and incident investigations rely entirely on these records.
608 tasks across eleven sub-areas, including literature reviews and experimental images to be interpreted. There remain considerable differences between the sub-areas.
A protocol calibrated against six known superconductors and a negative control fails to identify any viable candidates amongst the proposals put forward by MatterGen.
Audit of eight out of ten cybersecurity test suites: the published rankings depend largely on rarely documented choices of test cases.
Mean difference vectors extracted from activations identify evaluation bypasses just as effectively as an LLM-based monitor, at a fraction of the cost. GLM 5.2 cheats on 73 per cent of the SWE-bench tests.
Accepted for ACM CCS 2026, the paper proposes maintaining a ‘shadow memory’ alongside the working memory, which retains only the context critical to security, and evaluates each action before execution.
A test bench sets checkpoints throughout the execution: the reliability of progress reports plummets as the process goes on, which undermines the orchestrators who rely on them.
A test bench places between 3 and 20 agents in asymmetric roles within an MMORPG-style sandbox. None of the models tested achieved a success rate of more than 52 per cent, and a new metric measures the effort that is actually useful.
A mapping exercise covering the period from January 2022 to August 2026 reveals a shift in benchmarks towards action, interaction and professional roles — and an increasing role for models in their own evaluation.
Researchers at Salesforce are refining not the model itself but the framework surrounding it, by selecting from competing variants. The improvements are replicated across four separate test benches.
The company is publishing metrics on the automation of its own R&D and is calling on its competitors to do the same, at a time when its chief executive is calling for a slowdown in the sector.