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.
BEAM, peer-reviewed at ICLR 2026, scores whether an agent still trusts the right fact across very long conversations, tickets, and documents. past.dev's run is the first on the open harness.
Three AI coding models on 188 real concurrency bugs (race conditions, deadlocks) in a no-network container, graded by each project's own tests. Scores look close, until the hard half.
10,000 OpenAI agents finished a famous math problem in days. A scaling law is the idea that more compute reliably buys more capability; this one trades model size for agent count, and new failure modes.
Both are still computational predictions for low-power magnetic memory chips, the class behind MRAM and hard-drive read heads: one was designed from the laws of physics, not empirical tuning, the other is a 1999 material being re-checked
Their pitch: real-time text-to-speech running on Tenstorrent's RISC-V servers — an open-source chip architecture competing with Nvidia — with usage-based pricing and no upfront commitment.
A 7-day smart-speaker A/B test delivered +4.53% active users and +6.30% total listening time, both statistically significant (p<0.
ELMS, short for Evidence-based LLM-guided Monte Carlo Search, from a Georgia Tech team reuses failed protein structures as repair instructions and solved 26.7 of 30 tasks on MotifBench, a protein-scaffolding benchmark, up from 16.
A new technique lets the AI that forecasts crops, water, and land-surface conditions selectively forget bias-causing data from real-world shocks (new dams, sensor swaps, irrigation booms) in hours on a single GPU, cutting a key bias measure by
The training-free method steers pretrained generative models (a family used to turn noise into images or simulations) toward physics or measurement constraints with a single closed-form solve, one algebraic calculation rather than an iterative
A paired study of two small classifiers that decide which model, tool, or action an LLM agent uses next, and the three errors the authors found in their own analysis.
ProAct, a vision-language-action model, anchors each action chunk in the robot's last motion and a predicted future frame. Reported as an arXiv preprint, not yet peer-reviewed.
A thin capacitive skin under $10 in materials maps 202 touch points identically on human and robot hands, so a person's demonstrations can train a robot for contact-rich tasks.
Counterdiabatic driving — a shortcut that lets a quantum system change state without the usual slowdown — has dropped its Hamiltonian requirement (the math describing how a system evolves) and now extends to gate-based circuits (the gate-by-gate
Numerical experiments above one million qubits split quantum error-correction thresholds into two distinct failure patterns — one set by the lattice geometry, one set by the code's topology — in the standard error-correction design pattern used
TypeSafe's Jev returns a single decision instead of a paragraph, beating bigger open-source models Qwen and Gemma at typed structured-output tasks for under $10 in compute.
Researchers extracted a 'pain direction' from 25 open-weight language models, and a public reproduction of the experiment was petitioned, briefly removed, and reinstated without explanation.
A new measurement study shows why one Adam run (the default neural-network training optimizer), in a neural network under ill-conditioning (a sharply uneven loss surface), can drift a thousandfold from the theoretical ideal and still reach the
An arXiv preprint proposes sampling tailored modifications to a model's own settings, rather than just longer streams of answers, as a way to buy more capability per fixed compute budget.