47 stories, 177 Manifests cited.
FrameBlend's selective KV recomputation trims per-frame prefill FLOPs by over 20% and processing time by 14-59% while improving accuracy. A new survey maps how agents should actively sample video instead of watching it all, and MS2CAST runs anomaly detection with peak GPU memory under 1 GB.
Five new preprints attack LLM inference cost from the same direction: move the KV cache smarter. UlaCache reports 2.51x peak decode throughput on B300, HAWK offloads decode to NPUs at 1.72x the cost efficiency, and KVLearn trims time-to-first-token 56% by learning what to keep.
Three preprints refine the map of what drives performance: a formal MoE generalization theory, a law predicting multimodal accuracy from text-only scores, and distillation pushing a 20,000-parameter model past every from-scratch jet tagger.
Four preprints converge on the same lesson for mobile robots: task-aligned pretraining and training strategy matter more than model size, and simulators are becoming portable.
Three preprints refine RL for reasoning at a finer grain: EPIC regulates drift at the prefix level with an exact KL formula, ReAPO assigns credit via shortest paths to reach peak accuracy with 39% fewer tokens, and StepRS-GRPO brings structured exploration to diffusion LMs.
From parallel latent reasoning to diffusion models that fix their own mistakes mid-generation, researchers are building inference machinery that does not think one token at a time.
A systematic PTQ evaluation finds rotation-and-scaling recipes tuned for INT4 transfer poorly to MXFP4 and NVFP4. Meanwhile InfoQuant reports 97% floating-point accuracy retention at W4A4KV4, and QSelect beats single-family quantization by composing quantizers module by module.
New preprints turn agent topology from hand-tuning into optimization: CATS and Bayesian evolution both beat heuristic baselines, while a mapping study shows deployed multi-agent systems run on open models under 8 billion parameters.
One preprint teaches exploration with roughly 100x fewer interactions than RL baselines, another shows in-context RL models fail the tests for Bayesian posterior sampling, and a third turns the ICL-vs-fine-tuning choice into a certified rule.
Four preprints attack the same problem from four sides: agents that stay on track over long trajectories, undo their mistakes, optimize systems rather than files, and write kernels across abstraction levels.
BrainCodes matches state-of-the-art EEG performance with up to 97% fewer trainable parameters. BrainPro uses topographic retrieval to tame montage heterogeneity across nine BCI benchmarks. And MuteBench measures how clinical multimodal models degrade when sensors fail.
Four preprints this week target the economics of agentic RL: a system study pushes training past 4,400 tokens/s/GPU, a new execution layer gets 12.8x spatial efficiency on stateful data tasks, adaptive replay cuts serial test-time passes, and scaling laws rank environments above trajectories.
Three preprints map where agents break trust: verification fails on retrieval rather than reasoning, one in five trajectories exceeds authorization, and persistent memory carries attacks across sessions.
O'Driscoll and Lemkin say only trillion-dollar narratives get funded. Morgan Stanley sees investment-grade issuance up 25% on AI capex. The stress point is neocloud debt.
Snowflake shipped adaptive compute, an agentic marketplace finder, and agent memory patterns; GitHub is moving git to Azure as agent traffic grows exponentially.
DSX targets the power bottleneck behind $300B of backlogged equipment; cloud partners squeeze more GPUs per watt; factory blueprints and confidential computing handle design and trust.
SentinelOne launched agent detection and an autonomous SOC; NVIDIA pushed a governance reference architecture and confidential bare metal.
Pony.ai runs tens of thousands of daily rides and eyes London; Carbon Robotics cut staff; Serve Robotics heads into Q3 results with 2,000+ robots deployed.
AIDP gets private governed database connectivity; AI Vector Search brings semantic search into Oracle Database; a new agentic service workspace keeps humans in the loop.
From Starling's sovereign-cloud agentic model to Siemens' factory agents and Mistral's 3,800-GPU training run, NVIDIA's GTC Berlin blitz put detail behind Europe's AI-independence drive.
Google's Agents CLI now works with LangChain; GitHub moves git to Azure for agent traffic; Oracle adds agentic service apps and workflows that survive restarts.
NVIDIA expanded Cosmos for physical AI, launched StormScope EU for weather, and showed reward-guided protein design with RFD4-Proteina.
Craft backed ASI's $325M Series C on federal contracts; Azoma raised $11.2M for agentic commerce; Optivolt raised $32M for solar brains. Carbon Robotics cut staff.
Intapp is going AI-native, Qualitate automated the expert interview, and Baselayer wants to be the ID layer for agents. The question is who gets trusted, not just who gets automated.
In 20 Lahaina homes, indoor PAHs sat below risk thresholds 5-7 months after the 2023 Maui wildfires; physical partitioning dominates PFAS fate in treatment plants; biochar plus vegetation retains stormwater microplastics.
In preprints, the MEUWE framework extends electronic coherence lifetimes 13.4x; a quasiparticle framework matches CCSD(T) on ozone; and classical DMRG still beats noiseless VQE on transition-metal clusters.
VHH-Bench grades six structure predictors on nanobody complexes; a PPI study shows pooled metrics conflate target effects with mutation ranking; a DTI audit finds standard filters discard 78.4% of negative data.
Airlines, casinos, hotels, and dealerships all printed within a week. The consumer is still buying the experience.
Equipment, analog, sensors, and software all printed within days. The cycle's breadth matters more than any single print.
Four filings from the stressed and the shrinking: a prepackaged bankruptcy, a Dutch auction tender, a voluntary delisting, and another lifeline loan.
Three property REITs raised their 2026 outlooks on leased-up warehouses. The mortgage REIT across the street took a mark-to-market hit.
No drama in the quarterlies, but the corporate actions underneath are loud: an acquisition closing, a merger pending, and a corner-office succession.
Four industrial filings, one pattern: price gains and self-help are carrying earnings while the portfolio reshuffles underneath.
Two central banks sat on their hands this week while a BIS official told them the pricing mechanism they're steering may be changing underneath them.
Five governance filings in one week: three boards got bigger, and two C-suites handed over the legal and finance desks with long transitions.
In preprints: RaPID-derived macrocyclic peptides antagonize CXCR4 with nanomolar affinity; CYP8B1 inhibitors offer a first-in-class MASLD strategy; brain-penetrant DELC2 modulates neuroinflammation in Alzheimer's models.
In preprints, ascorbate depletion drives leukemogenesis via TET2/TET3 and its restoration extends survival in mice; tumor viscoelasticity - not stiffness - tracks breast cancer progression; KRAS-MAPK sustains a CD73 checkpoint in pancreatic cancer.
In preprints, a single proteomic panel classifies Alzheimer's, Parkinson's, and ALS at high AUCs; p-tau217 beats three rival markers over 10 years; but a 46,365-person meta-analysis finds only GrimAge predicts dementia.
Pittsburgh and New Orleans meet at Stade de France on October 25 for a Week 7 regular-season game. For Steelers coach Mike McCarthy, the logistics are new; the reunion with former coordinator Kellen Moore is not.
Chris Ballard is out in Indianapolis six games into a 2-4 season. In baseball, the Athletics hand the keys to the youngest baseball ops chief in MLB while the Astros and Angels search for new architects. Only the Yankees stand pat.
Competition Model 2.0 promises to cut mission selection timelines by up to 45 percent and punish cost and schedule breaches with consequences up to termination. This week, NASA also test-fired an RS-25 engine and watched Perseverance set the record for the longest drive on another world.
In one week, the Fifth Circuit applied Bruen to invalidate the Federal Switchblade Act, affirmed U.S. liability under the FTCA for a hostage killed in an FBI rescue, and ruled a master service agreement is not a maritime contract. The First Circuit added a mortgage lesson.
Deputy Secretary Troy Edgar met eight Angel families at a VOICE roundtable as the program's post-reopening agenda takes shape. The same week, ICE arrested a Mississippi woman accused of threatening to kill federal agents, and the Tenth Circuit refused to let an immigration detention appeal die.
S7236 on ammunition sales and Francesco's Law on safe storage were delivered to the governor on October 21. The same day, committees amended and recommitted bills on license-plate readers, e-scooter insurance, vet-clinic deals, and more.
The Shapiro administration directed $190.3 million to 24 water projects across 19 counties and marked its 500th plugged orphan well. In Montana, Forest Products Week puts numbers on a 10,000-job industry.
In preprints: treated wastewater fertigates lettuce safely in Brazil with a sodicity watch-list; Central Valley soil carbon hides at depth; Utahns' willingness to pay could finance Great Salt Lake conservation.
In preprints: Berkeley Earth reanalysis confirms 0.206 C/decade since 1970 with only marginal acceleration evidence; Southern Ocean freshwater rose 2,000 Gt in the Freezing Zone; marine heatwaves propagate via corridors, cold spells recirculate.