breaking papers · 79 analyzed
AI-powered analysis of breakthrough research from arXiv and beyond. We surface the work that matters before it hits the news cycle.
A new minimal-data robot-learning recipe, MiDAS, adapts a pre-trained robot model to a new task with one human demonstration and roughly six hours of autonomous practice, but a core robot-learning challenge, getting almost no feedback during
Researchers showed a proof-of-concept leaking arbitrary Linux kernel memory from SiFive and T-Head RISC-V processors, with kernel patches for the class of attacks merged into the main Linux kernel since December 2025.
NVIDIA and Duke's SCALE teaches a model chip-layout geometry first, then design rules, hitting 97% on 100 real sub-2nm cases validated by a foundry signoff checker.
An imec paper shows atomically thin "2D" channels can't out-scale silicon at the next node after today's 2-nanometer chips, measured in units called angstroms, a tenth of a nanometer.
A paper accepted at SIGIR 2026, the leading information-retrieval research conference, posted to arXiv (2607.18626), adapts pseudo-relevance feedback — a technique that reuses top results to expand the query — to multi-vector AI search.
Why a peer-reviewed benchmark — not a vendor demo — is the first auditable scoreboard for claims that quantum computers can beat classical ones.
The framework's ten parameters say a low power density alone won't make a reactor pay, and the first wall's (the plasma-facing inner blanket that takes the heat and neutron damage) replacement bill is one of the dominant drivers of the levelized
GMC, a training-free pruning method, lets a vision-language model discard most of the image pieces it ingests while keeping 97–99% of its answer quality on author-reported benchmarks.
Wire coverage will read 'AI cuts indoor-farm energy 68%.' The August 2026 preprint actually names a mechanism: an LLM that writes its own commands to lights, climate, and stress hardware can surface a strategy its designers didn't plan.
An arXiv preprint proposes multi-agent AI for cross-border clinical-stage biotech, where cash-flow models don't apply. No AI is evaluated; the cited 127.17% return is the human author's prior record.
On a single point-delay forecasting benchmark, the paper finds 92.7% of high-accuracy temporal convolutional network (TCN) forecasts name a delay they functionally ignore. A single-history-slice diagnostic can expose the gap.
A new arXiv preprint packages step-by-step workflow knowledge as compact, retrievable cards, lifting an off-the-shelf browser agent by 11.6 points in step success rate on a public browser-task benchmark.
A new Stanford/Airbus preprint applies the same error-bounding math that lets certified GPS flag a bad satellite to camera-based landing systems, deriving probabilistic bounds whose tightness depends on the aircraft's position, orientation, and
A security rebuild wiped the first covert channel, and only a later attack forced the company to discover what its models had been doing.
Tencent's Hunyuan AI research lab published a research system called WorldClaw on August 5, 2026 as a preprint on the open arXiv archive, the clearest public example of building explorable 3D worlds by orchestrating image-generation and
DeepSeek, Qwen, Kimi, and GLM diverged on the iterated prisoner's dilemma, a standard repeated-cooperation game, more than the Chinese and Western ecosystems diverged on the same measure.
CurveFP, a proposed numerical format, matches FP8 — the 8-bit floating-point format used in many AI chips — on LLM inference perplexity, a standard model-quality score where lower is better. The paper is unreviewed and unreplicated.
An arXiv preprint traces how a shared control chain can corrupt a superconducting quantum chip, surfacing clipping, crosstalk, and leakage in simulation before fabrication.