breaking papers · 86 analyzed
AI-powered analysis of breakthrough research from arXiv and beyond. We surface the work that matters before it hits the news cycle.
Pursuit-evasion is the 3D, multi-agent version of tag, where matching who chases whom decides the outcome, and a new GitHub release gives the field the shared benchmark it has been improvising without.
DataBrain's Text2Dashboard prototype passed 4 of 4 single-panel dashboard tests and 10 of 10 fault scenarios, but its authors call the result a small prototype, not a production claim.
A preprint steers deep neural network navigation testing toward regions where an explainability method called Integrated Gradients flags the model as least confident, finding more transferable failures than blind search on a LeoRover small wheeled
A new systems paper, FluidPD (Prefill-Decode), splits each request into two phases and lets a server borrow GPU time between them, or swap worker roles in place, when one side goes idle.
TEMPEST, a 2.8MB AI model, hits 91.71% accuracy on a 45-driver dataset and only loses 4.3 points as the pool grows. The privacy question is what the paper does not ask.
GA-GRPO (Guidance-Augmented GRPO, a training-recipe variant for reasoning AIs) turns expert hints, self-explanations, and retrieved reasoning patterns into a signal-to-noise dial, with a closed-form optimum and a matching lower bound on the bias
The arXiv preprint merges equality and inequality constraints into one geometry. The paper reports 100% on its own test suite, with hard collision checks still separate.
A new preprint trains a neural network to read spatial correlations from a grid of Rydberg atoms (atoms held in a highly excited state that interact strongly with neighbors) under their natural interaction dynamics, numerically beating the standard
CASTLE, a technique that gives each agent a frozen offline world model (arXiv 2610.
A new arxiv preprint on mixed multi-agent AI simulations finds base model choice predicts engagement more than persona assignment, a result that reorders the design hierarchy for builders.
Google's engineering post claims a 1.69× speedup on its TPU v6e AI chip at 1440p. The underlying idea is older, and the speedup is smaller than the academic paper behind it.
A simulation study finds 2-3 replanning rounds deliver nearly the same coverage as a fully adaptive drone, 15x faster, at a 12% cost premium.
ETH Zürich, Germany's CISPA Helmholtz Center for Information Security, and NYU researchers propose Terracotta, a programmable DRAM memory controller and composable primitives that let new memory techniques ship by configuration, not a chip redesign.
A new paper proposes 'grafting': learn the update on an earlier saved version of the model, then scale it onto the latest version of the model, with an optional filter (a mask) on the most sensitive update directions.
A 300-million-parameter transformer trained only on randomly generated sequences can compress real Wikipedia in six languages when given new data.
A new preprint argues video models break physics because they see every frame at once. Their fix: rough sketch first, then a parallel cleanup pass — with a LoRA adapter builders can drop into existing video models.
An AI turns 176 thin-film-coating papers (atomic-layer deposition) into a queryable database, but its F1 (precision-recall) score drops to 0.344 out of 1.0 on indium-gallium-zinc oxide (IGZO), the display-backplane chemistry that still needs a human.
The veteran investor is backing ex-DeepMind engineer Shivani Poddar's Wajo, whose assistant Fo calls businesses, hires people, and pays on the user's behalf.