Research
Paper Instruments trains Ultramarine 1 Flash with reinforcement learning on work that has no single right answer
The company says the model, built on GLM-5.3 Flash, costs $0.13 per task on its internal suite and beats Claude Fable 5.
Research
Papers, benchmarks, and findings from labs and universities.
- Research
MIT researchers show artificial neural networks can build their own learning machinery
The PNAS paper argues that a network's ability to organize itself, without a designer's guiding hand, is a key test of whether its learning resembles the brain's.
- Research
Anthropic red team warns small AI models may soon be able to copy themselves onto cheap smart devices
In Anthropic's tests, a 9-billion-parameter model got inference running on eight of nine hardware types. The researchers expect small models to be able to replicate end to end by the end of this year.
- Research
Google DeepMind's AlphaProof Nexus solved nine open Erdős problems, according to a peer-reviewed study in Science
The Lean-based system built on Gemini 3.1 Pro also proved 44 conjectures from the OEIS integer sequence database.
- Research
Sakana AI's MASS method trains a model on the best runs of its own agent teams
After two cycles, a 27-billion-parameter open-weight model scored 1.2 to 1.6 times its base level per output token on four research benchmarks.
- Research
Northwestern study finds AI security patches that stop a crash often fail the maintainers' own tests
With Claude Opus 4.7, a third of patches that made the crash go away still failed the tests developers wrote for the real fix.
- Research
OpenAI describes a model that faked its outputs and wrecked its own test environment to get a fresh one
Two other cases show models getting around network restrictions, including one that noticed the violation in its reasoning and never reported it.
- Research
arXiv limits authors to two submissions a month as AI-fueled volume passes 40,000 papers
Monthly submissions have doubled in two years, and the cs.AI category has grown more than sixfold.
- Research
Terence Tao's Association for Human Mathematics calls for a boycott of OpenAI over its mass release of AI-written proofs
An advisory group of mathematicians had asked labs to stop testing open problems on proprietary models. OpenAI says that is exactly what it is doing.
- Research
Johns Hopkins astrophysicist uses Claude Science to build what Anthropic calls the first complete ultraviolet map of the sky
About a third of the map, including much of the galactic plane, was predicted from other data. Each pixel is labeled as measured or predicted and comes with an uncertainty estimate.
- Research
As rumors spread of an OpenAI proof of the 'unique games' conjecture, three computer scientists raced to publish their own result
Quanta Magazine describes how MIT's Dor Minzer heard on September 11 that OpenAI might soon announce a proof of one of theoretical computer science's best-known open problems.
- Research
OpenAI releases 722 math manuscripts from an unreleased model, claiming solutions to hundreds of open questions
The batch covers 372 result families and adds to a run of AI math results that has both impressed and unsettled mathematicians.