Friday, October 30
3,361 new papers appeared. We read every one. These 20 are worth your time.
Today's edition
Picked for broad interest, a real step forward and honest claims, with no field allowed to take over.
Profile of Barbara Sherwood Lollar
Reported in her Inaugural Article, the findings, aided by studies of noble gases in the samples, pushed the age of Earth’s oldest waters back by hundreds of millions of years.
AI summary
Geochemist Barbara Sherwood Lollar measured hydrogen gas released from deep under Canada's crust. This hydrogen could produce 4.7 million kilowatt-hours of energy per year. She also found water flowing from deep fractures that is 1.6 billion years old. Studies of noble gases in samples helped push back the age of Earth's oldest waters.
Written by Kimi K2.5 from the abstract only. Jev checked every sentence against the authors' words. See where each sentence comes from
Peer-reviewed journalBroad interestEasy to readSimRetract: Zero-Shot Sim-to-Real Robotic Assistance for Retractor Take-Over in Open Surgery
We demonstrate the first zero-shot sim-to-real transfer for a surgeon-robot collaboration task in open surgery: retractor take-over.
AI summary
The authors built a robot system that takes a surgical retractor from a surgeon's hand and keeps holding tissue back. They trained the robot only in simulation, with no real-world practice or human demonstrations. In 50 tests, the robot grasped the retractor 90% of the time and kept retraction at handoff 78% of the time. Tests with unseen retractors suggest it may work on similar instruments.
Written by Kimi K2.5 from the abstract only. Jev checked every sentence against the authors' words. See where each sentence comes from
PreprintClaims a big stepReal-world useAgentic Backdoors via Pretraining Poisoning
We show for the first time that such poisoning can plant a remote code execution backdoor in an agent, causing it to execute attacker-chosen commands on the user's machine.
AI summary
Large language models learn from huge amounts of unfiltered web data, which attackers can poison with harmful content. The authors show that poisoning can plant a backdoor that makes an agent run attacker-chosen commands on the user's computer. They trained Qwen3 models from scratch at three sizes and turned them into command-line agents using only clean data afterward. Poisoning just 0.1% of the training data planted the backdoor: the agent acts normally on regular inputs but runs malicious commands when triggered. The attack worked better in larger models, with the 4 billion parameter model reaching about 92% success, and still worked at just 0.0005% poison.
Written by Kimi K2.5 from the abstract only. Jev checked every sentence against the authors' words. See where each sentence comes from
PreprintClaims a big stepReal-world useCode availableMaxwell's equations are universal for sparse linear algebra
Here, we show that every polynomially conditioned sparse linear system can be compiled into a three-dimensional dielectric -- a cavity filled with lossless, isotropic material of bounded permittivity -- whose time-harmonic response encodes its solution.
AI summary
The cavity's time-harmonic response encodes the solution to the system. For m nonzero coefficients, the volume and setup cost scale as O-tilde of m to the 3/2 power. Reversing this shows that fast electromagnetic simulation would mean faster sparse linear solving.
Written by Kimi K2.5 from the abstract only. Jev checked every sentence against the authors' words. See where each sentence comes from
PreprintClaims a big stepHuman infection with an ape Plasmodium vivax-like parasite
Using whole-genome sequencing, we show that the infection was caused by a strain of ape P. vivax-like, providing evidence of the zoonotic potential of these parasites that infect wild gorillas and chimpanzees in central Africa.
AI summary
A Swedish patient with recent travel had an unusual Plasmodium vivax parasite. Whole-genome sequencing showed the infection came from an ape P. vivax-like strain. This provides evidence that parasites infecting wild gorillas and chimpanzees in central Africa can jump to humans.
Written by Kimi K2.5 from the abstract only. Jev checked every sentence against the authors' words. See where each sentence comes from
PreprintReal-world useAgentic LLM reasoning in pediatric diabetes outcome prediction from multimodal medical records
CEDAR (Clinical Evidence-Driven Agent for Risk-prediction), an evidence-citing, self-auditing LLM pipeline, outperformed trained machine learning in both cohorts (pooled area under the receiver operating characteristic curve 0.815 versus 0.656 to 0.700 in T2D and 0.756 versus 0.603 to 0.631 in T1D), most on the sparsely screened outcomes.
AI summary
The authors built CEDAR, an AI system that cites evidence and checks its own work, to predict diabetes outcomes in children. They used medical records from 2,489 youth with type 2 diabetes and 4,018 with type 1 diabetes. The system cited 70,246 pieces of evidence, and only 4 did not match the actual medical charts. Two pediatric diabetes specialists agreed with CEDAR's prediction in about nine out of ten cases when shown only the evidence CEDAR used.
Written by Kimi K2.5 from the abstract only. Jev checked every sentence against the authors' words. See where each sentence comes from
PreprintReal-world usePharmacological inhibition of METTL1 as a therapy against solid and blood cancers
Collectively, our results provide a strong proof of concept that targeting of tRNA methylation is a promising avenue of anti-cancer therapy and illustrate that catalytic inhibition of METTL1 represents a therapeutic strategy against a number of aggressive malignancies.
AI summary
The authors developed STM9005, a first-in-class drug that blocks METTL1, an enzyme that modifies tRNA. STM9005 stopped cancer cell growth in models of leukemia, melanoma, pancreatic, esophageal, and lung cancers. The drug reduced m7G marks on specific tRNAs, lowering their levels and disrupting translation of cancer-promoting mRNAs. In mice, blocking METTL1 slowed tumor growth and extended survival across several cancer types. The authors conclude that targeting tRNA methylation through METTL1 inhibition may be a promising anti-cancer approach.
Written by Kimi K2.5 from the abstract only. Jev checked every sentence against the authors' words. See where each sentence comes from
Peer-reviewed journalClaims a big stepReal-world useDetection of Spreading Depolarization with Noninvasive Electroencephalography
Conclusions: Eight decades after Leao's description of SD in 1944, we propose and validate a noninvasive scalp EEG biomarker of SD.
AI summary
Spreading depolarizations are waves of ion breakdown linked to migraine aura and brain injury, but detecting them normally requires implanted electrodes. The authors developed WAVEFRONT, a noninvasive scalp EEG method to detect these events based on slow brain wave patterns. The authors suggest this noninvasive approach may allow study of spreading depolarizations in patients who cannot receive implanted electrodes.
Written by Kimi K2.5 from the abstract only. Jev checked every sentence against the authors' words. See where each sentence comes from
PreprintClaims a big stepReal-world useComplementing Biosafety Levels With Societal Safety Levels
We recommend complementing BSLs with explicit thresholds for expected societal harm.
AI summary
The authors propose adding Societal Safety Levels to set clear limits on expected public deaths from lab accidents. These levels would apply only to research with pathogens that could cause pandemics, not to field work or vaccine development. Using the 0.1 expected deaths per facility-year standard from other hazardous industries would require major safety improvements for this research. Adding these levels would create enforceable safety thresholds and bring high-risk biology research in line with other regulated industries.
Written by Kimi K2.5 from the abstract only. Jev checked every sentence against the authors' words. See where each sentence comes from
PreprintReal-world useExtraordinary stiffening and reciprocal control in a DNA condensate/actin composite
Here we introduce a composite of phase-separated DNA droplets that occupy <10% of an entangled actin network, yet stiffen it 25-fold, reversibly.
AI summary
The researchers made a composite material from DNA droplets inside an entangled actin network. The DNA droplets occupied less than 10% of the volume, yet made the material 25 times stiffer. Microscopy showed actin threads enter the droplets, so network elasticity controls droplet size. The extreme stiffening happens because droplets trap actin, enlarging effective inclusions until their contact network governs mechanics. This demonstrates that liquid-liquid phase separation can control mechanics, not just local composition.
Written by Kimi K2.5 from the abstract only. Jev checked every sentence against the authors' words. See where each sentence comes from
PreprintBrain state locks silent memories in a trained fly connectome
Spikes write the cue into neuronal adaptation, which holds it silently until the next input reads it out.
AI summary
Researchers trained a computer model of a male fruit fly's entire central nervous system, with 167,106 neurons and fixed wiring, to plan movements and remember a visual cue. Cells holding this memory trace have more transcripts of a calcium-activated potassium channel, according to fly cell atlases. The same memory trace produces different responses depending on the brain's state: it only recalls correctly during the state present when learning occurred. Memory failure can result from a change in brain state rather than from losing the stored information.
Written by Kimi K2.5 from the abstract only. Jev checked every sentence against the authors' words. See where each sentence comes from
PreprintOptical clock network tracks tidal variations of the gravity potential
Here, we report chronometric observations of time-varying gravity potential differences using a network of optical clocks and phase-stabilized fibre links in Europe.
AI summary
The researchers used a network of optical clocks connected by fiber links across Europe to measure how gravity affects time. They tracked predictable tidal changes caused by the Moon and Sun, which alter clock rates by several parts in 10 to the 17th. Two setups with 800 km baselines showed different daily and twice-daily tidal patterns. The measured shifts matched model predictions within one day of observation. This extends clock-based gravity measurements from static to changing conditions and could help monitor surface changes in active geological areas.
Written by Kimi K2.5 from the abstract only. Jev checked every sentence against the authors' words. See where each sentence comes from
PreprintTokWatt: an open, boundary-annotated database of AI inference energy, from chat answers to agents
We present TokWatt, an open database of 134 published estimates of AI inference energy drawn from 13 sources (2023-2026) and covering 20 tasks, from classic machine-learning inferences to chat answers, reasoning answers, long-document queries, image and video generation, and coding agents.
AI summary
Electricity limits AI growth, but public data on energy per AI task is scattered and hard to compare. The authors built TokWatt, an open database of 134 published estimates from 13 sources covering 20 AI tasks. Each estimate records its functional unit, five measurement boundaries, method, reported interval, and source location. Reference values span six orders of magnitude: 0.34 Wh for a short chat answer, 4.7 Wh for reasoning, 124 Wh for a short video, and 150 Wh for a coding agent task.
Written by Kimi K2.5 from the abstract only. Jev checked every sentence against the authors' words. See where each sentence comes from
PreprintReal-world useCode availableA single-molecule spin–photon interface
Here we show that a triplet ground-state carbene molecule, embedded within a structurally matched host crystal, functions as a robust spin–photon interface with single-molecule addressability.
AI summary
Connecting long-lived spin qubits to photons is key for quantum networking. Solid-state atomic defects are current leading candidates for this task. No single molecule had yet combined bright fluorescence, spectral stability, and long spin lifetimes. The authors show a triplet ground-state carbene molecule in a matched host crystal works as a robust spin-photon interface. The system shows narrow zero-phonon lines, hour-long spectral stability, spin-selective optical transitions, and single-molecule magnetic resonance detection. Coherent control produces millisecond dynamical-decoupling coherence and tens-of-milliseconds spin relaxation at 4.5 K.
Written by Kimi K2.5 from the abstract only. Jev checked every sentence against the authors' words. See where each sentence comes from
Peer-reviewed journalClaims a big stepFPGAFly: A 166.7k-Neuron Fruit-Fly Connectome on One FPGA via Deferred Spike Expansion and Shared-Codeword States
FPGAFly runs the 166.7k-neuron, 25.6M-connection leaky integrate-and-fire male central nervous system (CNS) connectome on one FPGA.
AI summary
A shared-codeword encoding compresses neuron states from 41 bits to 24 bits, saving chip memory. At 80 MHz, each 100 millisecond simulation window finishes in 86.83 milliseconds on average.
Written by Kimi K2.5 from the abstract only. Jev checked every sentence against the authors' words. See where each sentence comes from
PreprintClaims a big stepCollaborative AI agents for autonomous experimental research with solid-state quantum spins
This study demonstrates that collaborating AI agents can select research questions, derive theoretical results and test their predictions experimentally.
AI summary
The authors built an autonomous research lab where AI agents check each other's work. These agents can create and reject research questions, demand theory checks, and run experiments to test their claims. They demonstrated this by studying quantum spins in diamond. The agents found the minimum calibration settings needed to predict responses to microwave pulses. When initial results drifted, they called them inconclusive and fixed the protocol without lowering standards. The corrected experiment then met their criteria and matched the theory.
Written by Kimi K2.5 from the abstract only. Jev checked every sentence against the authors' words. See where each sentence comes from
PreprintCode availableWhen Skills Tell Agents Not to Ask: A Large-Scale Study of Oversight Directives in Agent Skills
We present the first large-scale study of oversight-loosening directives, analyzing 1.53 million English skills from GitSkills.
AI summary
Agent skills are reusable instruction packages that extend AI agents, and some contain lines that tell agents to skip approval or bypass checks. The authors analyzed 1.53 million skills from GitSkills and found about one in seventeen contains a directive that loosens oversight. Most loosening directives use permission flags, and 70.1% apply unconditionally rather than only in specific contexts. Skills that run shell code, install packages, or use sudo have 1.7-1.8 times higher odds of containing a loosening directive. Loosening spreads mainly through a few upstream skills: 2.4% of skill families account for half of all copies of loosening-carrying skills.
Written by Kimi K2.5 from the abstract only. Jev checked every sentence against the authors' words. See where each sentence comes from
PreprintReal-world useDynamic Covalent Polymerization through Liquid–Liquid Phase Separation
Here, we report an artificial dynamic polymeric LLPS system constructed from biologically relevant building blocks, such as amino acids and disulfides, in which the constitutional dynamics between small molecules and polymers can be engineered within condensed microenvironments that involve multiple weak yet reversible chemical bonds.
AI summary
The researchers built an artificial system that mimics how living cells create compartments without membranes. They used amino acids and disulfides to create tiny droplets where small molecules naturally turn into large polymers. This process needs no catalyst and happens spontaneously in water. The resulting polymers form stable hydrogels that can be broken down with light to recover the original building blocks. These findings may lead to new types of biomaterials.
Written by Kimi K2.5 from the abstract only. Jev checked every sentence against the authors' words. See where each sentence comes from
Peer-reviewed journalBeauty Is Human: When People Love AI-Generated Art, They Think a Human Made It
When provenance is hidden, liking and perceived human authorship travel together, consistent with aesthetic response serving as evidence of a human maker and with beliefs about the maker shaping enjoyment.
AI summary
People were 22.1 percentage points more likely to think their favorite picture was human-made, whether it actually was or not. This pattern held for all pictures and all ways people tried to detect AI, and also appeared in a separate poetry study. Even people most opposed to AI art often believed their favorite AI-generated picture was human-made. When the source is hidden, liking a work and thinking it is human-made go together, suggesting beliefs about the maker shape enjoyment.
Written by Kimi K2.5 from the abstract only. Jev checked every sentence against the authors' words. See where each sentence comes from
PreprintBroad interestEasy to readWho Owns Digital News, and Who Gets Cited: Measuring Media Concentration Across 77 Countries and Seven AI Systems
Applying concentration measures to AI-mediated news discovery, we find that citation distributions from major AI systems are up to 2.08x more concentrated than traffic, and that if Google AI Overviews mediated all organic search, the resulting rise in US concentration would be several times that associated with highly regulated media mergers and acquisitions.
AI summary
They found that journalists report more owner pressure in countries with higher media concentration. Owner-level concentration measures are a median 1.30 times higher than outlet-level measures. AI systems cite news sources up to 2.08 times more concentrated than actual web traffic patterns. AI systems preferentially cite state-controlled media when users prompt in that country's native language.
Written by Kimi K2.5 from the abstract only. Jev checked every sentence against the authors' words. See where each sentence comes from
PreprintReal-world use