breaking papers · 73 analyzed
AI-powered analysis of breakthrough research from arXiv and beyond. We surface the work that matters before it hits the news cycle.
Anthropic's Claude helped Nobel laureate Giorgio Parisi prove a decade-old identity in the math of jamming, where disordered systems like grains or glass freeze into rigidity.
The open dataset matches 1.3 million code files with records of how they actually run, so AI can be trained on what code does at runtime, not just how it reads.
A convenience-store robot restocking cans is not a labor substitute. It is a training rig for a foundation model that does not exist yet, and the human pilot on the line is the data engine that makes the math work.
MIT Media Lab built a tool that scans a custom prompt before the model replies, reading how it pushes the chatbot along dimensions like empathy, honesty, and flattery. The millions now customizing AI companions will never run it.
A commercial low-Earth-orbit (LEO) positioning, navigation, and timing (PNT) constellation launching in October 2026 turns GPS from a single point of failure into a buyer's market.
DAMO Academy, Alibaba's research arm, Renmin University and the University of Chinese Academy of Sciences used an LLM-wrapped materials agent to screen 2.
The Infinigence-Tsinghua project moves from a research reinforcement learning system to a unified data-to-deployment stack with new models and real-robot platforms. Its performance claims are project-reported, not independently verified.
An academic study of 2,991 GitHub projects finds the real shift after a bot joins lives in coordination, recognition, and conflict, not in what the bot can do.
A drop-one-step test gives agents a self-generated reward signal for which searches actually move the answer forward.
USC's Ψ₀ teaches vision and motor control as two separate skills, lifting success rates by more than 40% on eight long, multi-step manipulation tasks at the Robotics: Science and Systems (RSS 2026) conference.
The same model in a multi-agent structure beats itself as a single agent on 96% of papers. The paper's own data says calibration, not comprehension, is the part humans still own.
The classic antitrust question — did the algorithms settle on a collusive price — misses the real mechanism.
A new framework turns quantum advantage claims into auditable numbers. One current benchmark already fails the test.
A new arXiv preprint borrows the predict-verify-recover pattern from modern CPU branch prediction, and a blast-radius bound, machine-checked in the Lean4 proof assistant, makes a wrong guess cheap to undo.
A new benchmark puts a number on what flagship robot foundation models can do with a person in the workspace, and the answer is closer to zero than to half.
An AI test reading the same slides pathologists already examine matched the standard 21-gene genomic recurrence test at how well it separated patients who recurred from those who didn't and ran in days, not weeks.
KAIST's HOUND robot dog crossed 0.7 miles of campus and 0.2 miles of forest by picking its own gait on the fly — a visible hint that legged robotics is starting to ask "can it pick the right walk" instead of "can it walk."
OpenAI's GPT-Red admits the bottleneck human safety testing cannot clear, and the arms race that follows.