breaking papers · 68 analyzed
AI-powered analysis of breakthrough research from arXiv and beyond. We surface the work that matters before it hits the news cycle.
IBM's new arXiv paper proves shallow (constant-depth) quantum circuits can solve two narrow problems no theoretical model of an AI language model can: a proof, not a benchmark.
Graphcore and Arm compressed Meta's 11B vision-language model to 3.7 GB for a Pixel 8a, with average visual question-answering accuracy dropping from 74.4% to 66.1%.
A multi-institution digital twin of a quantum sensor shows that optimizing for one performance metric can quietly blow the other two, by up to two orders of magnitude.
The model-quality race is over. The next chip fight is over who can run the answers cheapest, fastest, and longest, and the alliance map is already shifting.
Average accuracy looks great. Repeat-run reliability is the metric production actually depends on, and almost no public scoreboard surfaces it.
An agent-based model is a market, city, or epidemic simulation built from many interacting actors. A new preprint speeds calibration by 61% on one benchmark, but the accuracy gain depends on the model.
A new preprint matches human and LLM group chats on reasoning tasks and finds AI agents hit the right answer by copying majorities, converging early, and surfacing less unique information than people.
Everything ran in MuJoCo, a robotics simulator. The team says sparse human takeovers at the contact point did the heavy lifting.
An arXiv pre-print reports a four-party quantum key run at 750 ± 10 bits per second, with a security proof that holds against a cheating participant and composes with other cryptographic tools.
A fully open 7-billion-parameter model is matching frontier systems on math reasoning by redesigning training, not scaling compute.
The 2024 SPAC (a public-listing merger) brought Zapata almost no cash and over $20M in debt, and the question is whether a journal cover and applications revenue measure the same thing.
Companies now use one AI to grade another's outputs, and those AI judges often share training data. A new method that accounts for that correlation beats naive majority vote by 9 to 14 percent.
Researchers ran a peer-reviewed quantum optimization technique on Rigetti's 9-qubit chip that uses fewer physical qubits than a problem's variables normally require, with a tunable trade-off in sequential operations the chip must perform.
Physicists at the Raman Research Institute have shown that a single flip on a two-photon system can delay, prevent, or speed up 'entanglement sudden death', the rapid loss of quantum correlation.
Shanghai's Jubrain Panshi (具脑磐石) released Cog-WM 1.0, a brain-inspired "cognitive world model" for robots — an internal predictive model of how its environment will change as it acts, rather than a pixel generator.
The hard question for AI mental health is whether the app is built to move you forward or to keep you subscribed.
By scoring AI physics predictions against the underlying equations, researchers turn the violation into spatially adaptive confidence bands that hold across six physics problems.
HardFlow, published this week in IEEE Transactions on Pattern Analysis and Machine Intelligence, wraps pretrained diffusion and flow-matching models — a newer class of generative AI — so their final output respects robotics, control, and vision