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
Submissions to the Ramanujan Challenge, research-level number theory problems, yielded an open proof of conjecture. Judging them reliably proved to be a second challenge.
Compressed in 2 and 4 bits without retrain, the Kroko transducer processes 1.28 s of audio in 0.90 s on a chip without a neural accelerator.
Rather than judging AI on its most publicized breakthroughs, researchers tested it on a random sample of recent conjectures. Trivial counter-examples and unoriginal evidence are alongside real novelties.
An autonomous laboratory brings together agents who choose the question, derive theory, design experience... and reject a first measurement distorted by a drift of the signal.
Agents swarms translated into Lean a demonstration of 15,000 pages, or 43.1 million lines, and identified more than 2,000 errors, including a gap that had to be filled by new mathematics.
The LongChang test bench drops language models in simulations where, for example, heavy objects fall slower. GPT-6-Astra identifies anomaly in 16.7% of cases; The other eleven, never.
In six days and fourteen problems, the system produced unprecedented results in quantum information and chip design. His author also details his failures, including too lax evaluators exploited by the agent.
By organizing data, tools and procedures in reusable bricks that improve in use, MatOS exceeds Codex CLI and Claude Code with an equal basic model on several scientific test benches.
Tested blindly on three articles before their official corrections, the open source tool located each error at the very stage designated by the erratum, and found others.
Refutation of the verified ETH in Lean, proof of local smoothing conjecture, existence of Yang-Mills theories in 4D: major announcements that now await an independent review.
Pre-entry models cite AI in 27% of their study, career or product recommendations; their versions published after instruction do so in 59% of cases.
A new measure, the response coverage gap, shows that six LLMs leave behind entire profiles of true respondents, to the point of doing less than a random response draw.
Folding, reverse design, understanding: one model covers six tasks and reaches a TM-score greater than 0.8 with one sample.
A team identifies 14 types of drifts and detects 3,964 on 300 tracks, and then reduces those remaining at the end of the task by 36.8%.
Eko agents, based on Claude Opus 4.7, invented games, erected a memorial for a lost ring and discovered only techniques that helped them in the following tasks.
A comparison of human-AI and human-human teams shows that iterative review, not AI, improves research proposals.
Thought for celestial navigation when GNSS is unavailable, the method is more resistant than its predecessors to the disturbances of a terrestrial observation, at a comparable time of calculation.
By placing up to 100,000 agents from four LLMs on different networks, physicists show that topology dictates the speed of consensus and vulnerability to manipulation.
Researchers have blindly read articles published in 2026 in major journals: models demand massive revisions, contradict each other and often miss errors that have been deliberately slipped.
Xinzhi Zhang and Zhao Song demonstrate a constant terminal for Kannan-Lovász-Simonovits conjecture. AI was used as a computational and verification assistant, with humans keeping their hands on the strategy.