Breaking Papers

Factories are training robots in simulation and deploying them to real floors based on a single unverified claim: 99% accuracy. ABB made the number. Nobody else has checked it. Cadence and Nvidia just announced they are building tools to close the gap anyway.

HETDEX catalogued 33,612 hydrogen halos around early galaxies. Nearly half required a two-component model the field had never systematically applied — and researchers say better instruments would likely reveal the same structure in every single one.

Two top AI models hit the same success rate while acting completely differently. That contradiction just explained why your agent fails.

ParityQC just ran a quantum algorithm on 52 qubits — nearly double the old record. But the result worked roughly once in every 100 attempts, and the real story is the compilation architecture that made it possible.

The Stanford HAI AI Index shows AI agents at 77.3% on real-world computing tasks — but the benchmark testing genuine scientific reasoning puts the same systems at 38.78% against an 83.5% PhD expert baseline. A 45-point gap nobody is reporting.

Your future encrypted data might be secured by an 1836 optics trick. That's not a metaphor.

Compliance with the EU Product Liability Directive costs the same for a startup as for Microsoft. That is the problem.

JWST found a sulfur paradox in four gas giants 129 light-years away — a star depleted in sulfur hosts planets enriched in it, and nobody can explain why. The data is real, the gradient is tentative, and it matters for whether our solar system is typical.

A model trained to sort numbers started preferring owls 60 percent instead of 12 percent. The data had nothing to do with birds. The fine-tuning economy customizing AI models for thousands of businesses may have been silently propagating hidden preferences all along.

IBM Research built a benchmark to test whether AI agents follow the rules. They found that models violate constraints systematically. The real problem: the industry and regulators have no standard way to measure it.

DESI finished its 5-year cosmic map 13 months early — 47M galaxies vs. 34M planned. The press release calls it a paradigm shift. The analysis that could produce one arrives in 2027. Until then, LambdaCDM — 30 years of standard cosmology — is still standing.

The paper proposing to find life via planetary statistics is peer-reviewed. Its catch: astronomers first need a catalog of what dead planets look like, and that baseline does not exist yet. JWST is working on it — three to five years.

MIT built a quantum sensor that measures three things at once at room temperature on a chip — and the same lab already has the upgrade path on arXiv

Google and Cloudflare just set 2029 deadlines to migrate away from current encryption after two papers showed breaking elliptic curve crypto needs far fewer qubits than previously thought. Enterprise migration takes a decade. Cryptocurrency has no switch to flip.

In agentic AI, the bottleneck is CPU tool processing — causing up to 90.6% of total latency, not GPU inference. Both Arm and Nvidia launched CPUs for this orchestration layer in the same month. The GPU-centric infrastructure model may be structurally incomplete.

Researchers at the University of Virginia gave a team of robots distinct LLM-driven personalities and ran 105 people through a study. The results suggest robot individuality is not a UX feature — it is infrastructure.

The paper's key insight: judge feedback quality by whether the author revised the paper. Train on 19,534 ICLR submissions using that signal, and a small open-weight model beats Gemini-3-flash on precision: not by being smarter, but by learning from what actually worked.