Skip to main content

Daily report

AI and science frontiers · 2026-11-17

Only content delivered through the publication boundary on this date is included.

Sealed edition v1

bioRxiv

Heme hydroxylases no longer need natural scaffolds: computational design steers estradiol oxidation to C6 or C16

A computational design pipeline builds heme peroxygenases from scratch, taking only a cysteine-ligated heme, a glutamate-histidine dyad that activates hydrogen peroxide, and a substrate positioned against the iron-oxo intermediate, so a hydroxylase can be tailored to a chosen C–H bond. On estradiol, two parallel campaigns directed oxidation to the benzylic C6 or the unactivated C16, giving initial yields below 1% but diastereomeric excess of 93% and 99%; targeted mutations on the C16 design raised estriol yield to 37% while keeping high selectivity.

arXiv

Encoding channels separately wins at scale: Meta FAIR's DINO-BoC beats joint channel encoding on microscopy and satellite imagery

Scientific images differ in channel count and semantics, and joint channel encoding was long assumed necessary to learn cross-channel relations. Meta FAIR and collaborators compared strategies on an extended training set of about 7.75 million images and found that DINO-BoC, which encodes each channel independently and concatenates features at inference, outperforms the joint-encoding Channel-ViT on most microscopy and geospatial tasks and exceeds the supervised HyperNet baseline on CHAMMI's mean linear-probe out-of-distribution score. The finding covers in-domain, cross-dataset and unseen channel combinations, though protein-localization tasks still need a linear probe to catch up under nearest-neighbor evaluation.

arXiv

Fixed-gain PD forces a trade-off between motion and contact force: ImpedanceMimic lets the policy set joint impedance, tracking force better across 129 participants

Physics-based motion imitation usually converts joint-position targets into torques with a fixed-gain proportional–derivative (PD) controller, so one motion can correspond to different contact forces, and fixed gains struggle to serve stance and swing at once. ImpedanceMimic has a reinforcement-learning policy output per-joint stiffness, damping and position setpoints, supervised in training by force-plate ground reaction forces and center of pressure, while inference uses motion alone. Across 129 participants and six activities it beats fixed-gain PD, direct torque, muscle actuation and existing imitation controllers on joint-angle error and force fidelity, and transfers zero-shot to unseen humans and a simulated humanoid robot.

arXiv

SPARC lets experimental evidence rewrite model weights: 8 of 11 recommended conditions fall below 300 mV overpotential

Closed-loop experiments usually fix the model family and representation first, then let new data update only predictions. SPARC builds an evidence matrix from the out-of-sample predictions of competing regression models, assigns each model's influence through a sparse posterior, and combines candidate-dependent predictive dispersion with a residual scale into a trust score that modulates exploration. In an automated alkaline oxygen-evolution campaign, 8 of 11 recommendations fixed in advance reached overpotentials below 300 mV, versus 18 of 125 initial experiments, with a best value of 242 mV.

arXiv

Latent feedback makes each FLOP go further: up to 48% less compute, or 47% less unique data, than repeating epochs

A latent-feedback Transformer feeds each token's top-layer state into the next token's input; earlier comparisons held training tokens fixed and never repeated data. Across 60M to 778M parameters, 12.5 to 400 unique tokens per non-embedding parameter and one to four epochs, with both models tuned at every size and data amount, latent feedback reaches a Transformer's loss with less compute at fixed unique data and with less unique data at fixed epochs, and both savings grow with unique data per parameter and with repetition. A fitted scaling law predicts held-out 778M losses to within 0.3% and the gap between the two models to within 0.05%.

arXiv

Bulk transcriptomes no longer need a model per cohort: PRTS reads a labelled cohort as context and edges TabPFN v3.5 on mean rank correlation

PRTS is a 220-million-parameter model that reads a labelled cohort together with new samples and returns binary, multiclass, continuous or censored-survival predictions in one forward pass, with no per-cohort training. It reached the highest mean rank correlation on 175 held-out tasks and 44 TCGA tasks, but the lead is small and comes mainly from prediction within a cohort; across cohorts and at unseen sites it is level with the strongest tabular foundation model, TabPFN v3.5, while running more than twice as fast.

arXiv

SpinAware reads one mean-field solution per spin state and flags 99% of correlated orbitals, carrying transition-metal spin gaps onto IBM hardware

Multireference calculations on transition-metal complexes depend on choosing an active space, yet measuring which orbitals are strongly entangled requires the very correlated calculation the space is meant to make affordable. SpinAware predicts DMRG single-orbital entropies and orbital-pair mutual information from the unrestricted Hartree–Fock solution of each spin state, using 60 orbital descriptors and a transformer. Over 11 spin states it reproduces the entropy profiles with a mean absolute error of 0.027, and a candidate pool covers 99% of the DMRG-active orbitals while keeping on average 15 of 40, supporting sample-based quantum diagonalization on IBM processors up to 32 qubits.

arXiv

CompCert4Lean moves a verified C compiler into Lean: one expert directing parallel agents completes the port in six weeks

The verified compiler core of CompCert 3.17 existed only in Rocq, and its definitions and proofs could not be imported directly into Lean. CompCert4Lean partitions the development along its dependencies into 166 units, translates definitions and theorem statements first under independent audit, then reconstructs proofs in parallel; one expert directing parallel LLM agents completed the port in about six weeks, discharging all 4,908 recorded proof obligations without project-specific axioms and producing executable Linux binaries.

arXiv

Together Compiler turns inference tuning into a continual loop: up to 2.6x throughput on frontier models, compiled in under 24 hours

Together AI released Together Compiler (TKCC), which turns optimization of an AI inference engine from a one-off compile into a continual loop driven by agents. Existing agent systems either tune individual kernels on offline benchmarks with low measurement fidelity, or restart the whole engine for every measurement and iterate slowly. TKCC decouples engine startup from forward-pass computation with a tiered persistent engine and adds a cost model for parallelism configurations; it beats hand-written Megakernels on Llama 1B and the best open-source engines on Kimi K3, GLM 5.3, MiniMax M3.1 Flash and DeepSeek v4.1 Flash on the InferenceX benchmark, within 24 hours and at most 2 nodes per model.

arXiv

An agent-written scheduler beats commercial solvers: 525,515 on a 100,000-operation job shop, below OptalCP, Hexaly and CP Optimizer

A frontier large language model, used as a coding agent with open-source tools only, reformulated shop-scheduling problems and wrote its own heuristic solvers. On the 100,000-operation instance tai_100_1000_1 its solver reached 525,515, below OptalCP, Hexaly and CP Optimizer; with equal budgets on one workstation it ended 0.0–1.8% above the lower bound across ten instances of the family, against 2.6–5.8% for CP Optimizer and 2.4–4.4% for Hexaly. On transport-robot MKT instances it improved seven of ten, but on Chambers–Barnes instances, where processing times dominate travel times, it was weaker than a CP Optimizer approach.

arXiv

Chemists write only the target, and VibeSyn's agents run a ten-step total synthesis: active human input is 7.3%

A team led by Zhejiang University built VibeSyn, a hierarchical agent system that turns a chemist's molecular objective into synthesis decisions revised as experimental evidence accumulates, then drives computational tools and automated devices to execute them. In a ten-step catharanthine total synthesis the system recorded 722 actions, of which agent reasoning was 69.7% and tool or device outcomes 18.8%, while active human input and human confirmation were 7.3% and 4.2%; starting from 29.8 mmol of 3-ethylpyridine it produced 160 mg of product. Across three campaigns synthesizing cyclic peptides containing unnatural amino acids, active human input and confirmation were 7.5% and 5.2%.

arXiv

CRAM turns Sudoku into energy descent: 99.996% of Sudoku-Extreme solved after 42 ms of training, 1,000 times faster than looped models

A trainable energy model called CRAM turns reasoning into repeated descent on an energy function whose weights are learned directly from solved puzzles. It solves 99.996% of Sudoku-Extreme test puzzles and every Maze-Unique test maze, with training on one GPU taking 42 ms and 132 ms respectively, at least 1,000 times faster than looped models. It also cautions that high scores on puzzles like Sudoku and mazes do not indicate how looped reasoning methods will scale to harder tasks.

arXiv

X-Strata splits a 2D angiogram into three layers, recovering the aneurysm hidden behind crossing vessels

A 2D X-ray angiogram sums every structure along each ray into one pixel, so vessels crossing an aneurysm hide part of the sac and mix their contrast into its filling curve. X-Strata separates a cine angiogram into the complete sac, the vessels in front of it and the vessels behind it: contrast timing indicates how much each layer contributes, and parallax between two views indicates their order, after which the layers are lifted into 3D. On 220 held-out patients from three centers, two-view X-Strata reached overlap-region sac Dice of 0.693 internal and 0.641 external and balanced depth-order accuracy of 0.703 and 0.659, with a median 3D sac surface Chamfer distance of 0.34 mm on the internal cohort.

arXiv

Bayesian neural operators miss systematic error: residual-informed correction cuts thermal-break online optimization from 178 hours to 18 minutes

Bayesian neural operators treat weight uncertainty as the only source of doubt, yet a learned operator keeps a systematic gap to the PDE solution that does not vanish as training data grow. The residual-informed Bayesian neural operator (RI-BNO) turns a one-step Newton correction of the governing residual into error observations, fits a probabilistic inadequacy model, and folds its mean and covariance into the likelihood. On a reaction-diffusion benchmark, glioma MRI inversion and 3D thermal-break risk optimization it agrees better with finite-element references than the vanilla Bayesian neural operator while using half as many full nonlinear or transient training simulations; thermal-break online optimization falls from about 178 hours to 18 minutes.

arXiv

Turing Agent turns formal verification into an agent pipeline: 57–62% faithful formalization on TCS benchmarks, plus three previously unresolved proofs

AI systems now generate mathematical proofs faster than anyone can check them, and a Lean statement can compile while quietly changing the claim it was meant to capture. Turing Agent splits the job into a theorem-formalization agent and a proof-formalization agent, coordinated by structured decomposition, parallel verification, an experience pool and selective semantic review. On the 143 problems of Formal TCS Bench it reaches 57.34% and 61.54% semantic success with GLM-5.3 and DeepSeek-V4-Pro, at compilation rates of 97.20% and 99.30%, and above 82% overall pass@1 on the 248-theorem RLMEval subset.

arXiv

Physics-informed operator learning for nonlinear PDEs gains an error guarantee: the first-order least-squares residual matches the error near a solution branch

Physics-informed operator learning trains on a PDE residual instead of solution snapshots, but whether a small residual means a small error was unproven for nonlinear equations. Writing the state as a first-order system of flux and primal variable, this work proves that the least-squares residual is equivalent to the error in the equation's own norm near a solution branch, provided the linearized residual is stable and its derivative Lipschitz continuous, and verifies the conditions for monotone reaction, solution- or gradient-dependent diffusion, and compressible Neo-Hookean hyperelasticity. A reduced basis neural operator trained on this loss kept the ratio of error to the square root of loss excess at medians comparable to the linear case.

Axiom Math

The CFSG Lean corpus does more than prove classification: for a typical one of 30 published theorems, over 99% of the code is reused

Axiom Prover treated its Lean formalization of the classification of finite simple groups (CFSG) as a reusable library and used it to formalize 30 published theorems that lie outside the classification proof and do not simply follow from the classification itself. A typical result cites corpus theorems whose proofs total about 880,000 lines of Lean and adds about 1,200 lines of its own, so more than 99% of the code behind it was already in the corpus; none of these results had reportedly been formalized before. The experiment is a sample of 30, each completed in under a day of agent time.

bioRxiv

No spike sorting required: C3PO learns interpretable population states straight from raw spike events, from milliseconds to hours

C3PO models neural spikes directly as a marked point process, skipping spike sorting and binning, and uses contrastive prediction over event history to infer a low-dimensional context state. On synthetic data the learned context forms a ring organized by latent phase and recovers tuning curves for 48 phase-tuned neurons; on several published recordings it recovers known variables such as position, head direction and theta phase, and maps population features directly onto single neurons.

arXiv

Tokenization ambiguity no longer sinks single-history steganography: coin projection keeps over 90% of string entropy at full vocabulary

Provably secure linguistic steganography assumes the receiver recovers the sender's tokens, but a text channel delivers only text, and one text admits many tokenizations, so retokenizing can lose the message. SyncPool kept the distribution exact by encoding pools of prefix-related tokens, yet its capacity collapsed at full vocabulary as whitespace and UTF-8 lead-byte hubs merged subtrees. Coin projection fixes each prefix's stop decision with shared randomness and leaves only root and child choices to the coder; on 3–4B models in English, Chinese and Japanese it leaves 91–93% of string entropy to the coder and turns 2.16–6.85 times as much entropy into message bits as SyncPool at full vocabulary, with zero error.

arXiv

Iris-Z folds depth into image generation: one set of weights reaches mean AbsRel 2.92 on Marigold's benchmarks, below MoGe-3's 3.12

Iris-Z is a general image generator that encodes image and depth into one token sequence and denoises them together, so a single set of weights generates images, estimates depth from photographs and paints images onto a given depth. Monocular geometry accuracy had belonged to task-specific ViTs such as MoGe-3, and image generators repurposed as depth estimators, the Marigold family, stayed behind it. Iris-Z-big reaches a mean AbsRel of 2.92 on Marigold's benchmark sets against 3.12 for MoGe-3 while keeping its generative abilities, at about 15 seconds per image versus about 0.41 seconds for MoGe-3.

arXiv

Irregular time-series generation stops freezing its imputations: PSLR co-evolves filler and generator, cutting discriminative score 59% below the strongest baseline

Irregular time-series generation usually fills missing values once, freezes them, and then trains a generator; PSLR instead runs a Monte Carlo EM loop in which each round samples complete series from the current generator's posterior and retrains the generator on them. Because the diffusion prior lives in a lifted 2D representation while observations live in time-series space, PSLR enforces operator-consistent conditioning, uncertainty preservation and manifold consistency to connect posterior sampling across the two spaces. Across ten datasets and four corruption regimes, it improves discriminative score, predictive score, Context-FID and feature-correlation error by 59%, 15%, 73% and 71% over the strongest baselines.

arXiv

Images Stay in Context but Slip Out of Reach: VPR Shifts Only Cached Visual-Key Rotary Phases, Lifting 97 of 100 Model–Task Scores

Keeping an image inside a multimodal model's context window does not mean its details stay usable: with 32K text tokens inserted between image and question, Qwen2.5-VL-7B's DocVQA score falls from 93.5 to 52.6 ANLS. Visual Position Re-Anchoring (VPR) shifts the rotary phases of cached visual keys toward a short-context reference before question processing, leaving visual values, text cache entries and causal masks untouched and adding no forward pass. Across ten checkpoints and ten benchmarks, static VPR improves 97 of 100 model–task scores; in the TextVQA length study, all six models remain below their short-context references at 32K, so recovery is incomplete.

arXiv

Quantum circuits can be screened for trainability before costly optimization: a representation-theory criterion plus a sampling algorithm decide with polynomially many runs

Gradients of variational quantum circuits can vanish exponentially with qubit count, creating barren plateaus that stall training. A new analysis splits the gradient variance over the invariant subspaces of the unitary representation of the circuit's compact Lie group and gives a sufficient condition for trainability: if the input state and the observable concentrate on the same polynomially sized invariant subspace, and the differentiated generator moves that subspace by an inverse-polynomial amount, the gradient variance has an inverse-polynomial lower bound. A companion sampling algorithm runs the circuit itself, without constructing representation-theoretic projectors, and decides whether the variance exceeds a tolerance with polynomially many runs.

arXiv

ARoute decides which model answers before any call: highest mean accuracy in all 15 task-cost settings, 4.1 points over the strongest baseline at matched cost

ARoute predicts both whether upgrading to the stronger model will help and how much more that upgrade will cost, using the query text alone, and picks the model before either model is invoked. It fits two lightweight heads on a small set of paired records over one frozen BERT encoder. On MMLU-Pro, LiveCodeBench, SuperGPQA and two SWAN value-filling tasks, ARoute-Batch attains the highest mean accuracy in all 15 task-cost settings, 4.1 points above the strongest baseline on average and up to 11.9 points, and completes a batch up to 2.8 times faster than cascades.

arXiv

LENS stops at an interaction score: a second model on held-out data must reproduce the pattern before a feature pair enters the model

A high interaction score can reflect a model fitting noise, and ranking methods leave the cutoff to the user. LENS trains two models with the same learning algorithm on disjoint data and accesses them only through predictions: it ranks pairs by interaction strength after removing main effects, then tests whether the second model reproduces the pattern at a meaningful strength, with one multiple-testing correction setting the retained set. On F1–F10 with XGBoost, LENS recovers 91% of material interactions against 45% for DIAMOND, without selecting false pairs.

arXiv

When verifiers can reject but not certify, NOPO learns only from failures and lifts average pass@16 above GRPO on code, logic and math

When a verifier can reliably reject wrong answers but cannot reliably certify correct ones, standard reinforcement learning with verifiable rewards reinforces shortcuts that merely pass partial checks. NOPO treats verifier-positive samples as unlabeled candidates and updates the policy only by suppressing trusted negatives, scaling each update by the share of non-falsified responses in the same rollout group. On Qwen3-4B and Qwen3-8B, average pass@16 on code, logic and math benchmarks exceeds GRPO and other baselines, and the advantage persists under inference-time scaling to 512 samples and cross-benchmark transfer.

arXiv

Theorems no longer stand alone: a graph of how results build on each other yields conjectures judges prefer

Mathematical agents have mostly worked on proving existing conjectures, while deciding what is worth proving lacks an operational criterion. The Theorem Evolution Graph records how theorems build on one another: nodes are theorems, and edges capture changes in setting, hypotheses and conclusions. Agents search citations backward and forward from a seed paper, compare theorems, and propose conjectures from those differences. Across 30 seed papers, three model judges preferred graph-generated conjectures in 63% of non-tied judgments, and an expert preferred them on 8 of 12 papers; baseline conjectures were favored for plausibility.

arXiv

Obaleth puts GPU-initiated networking inside the kernel IR, beating NVSHMEM by 1.25x on a 256-GPU all-to-all

GPU-initiated network communication has meant calling hand-written device libraries such as NVSHMEM and NCCL, whose transport decisions stay invisible to the compiler and whose every call pays for reservation, fencing and completion polling. Obaleth writes put, signal and wait into the kernel's intermediate representation, optimizes them together with the computation, and lowers them to transport-specific code. On H100 GPUs it sends 8-byte messages at up to 3.3 times NVSHMEM's rate and 4.2 times NCCL GIN's; on 256 GPUs its small-message all-to-all is 1.25 and 2.3 times faster, using 22 MiB of device memory per rank.

arXiv

Discrete codes can carry vision too: dNaViT uses one codebook for understanding, reconstruction and generation

dNaViT encodes images into discrete codes of the same kind language models predict, and one shared vocabulary serves visual understanding, image reconstruction, generation and editing. It builds on a frozen pretrained vision encoder, adds reconstruction-relevant detail through residual modulation, and uses grouped product quantization to widen discrete capacity without lengthening the spatial sequence. On ImageNet-1K, the 64-group dNaViT-G64 reaches 32.90 dB PSNR and 0.91 SSIM at 1024 spatial positions, and dNaViT-8B paired with Qwen3-8B scores 885 on OCRBench.

arXiv

Frontier models drive real robot arms: Reins gets GPT-6.1 Sol to 88% success on five long-horizon tasks

Frontier multimodal models can plan and call tools, yet they struggle to drive real robot arms through long-horizon, contact-rich manipulation, blind while the arm moves and while the model thinks. Reins lets an off-the-shelf model operate a dual-arm robot through 3D perception tools and atomic actions that report contact and grip, chains actions with stopping conditions, and has a second model instance watch the cameras and report what changed. On five real tasks, GPT-6.1 Sol succeeds in 88% of episodes from natural-language prompts with no task-specific code; without the harness it never lifts the correct mug in the shell game and takes 8 times as long.

arXiv

A frozen LLM learns full-duplex talk in 30 GPU-hours: 97M trained parameters, and a 0.170 floor-take rate when users merely pause

Full-duplex spoken dialogue usually needs hundreds of thousands to millions of hours of speech training and rewrites the language model that supplies knowledge. This system freezes the thinker of Qwen3-Omni-30B-A3B and trains only low-rank adapters, an audio adapter, a channel-fusion layer and 17 new token embeddings, 97M parameters, converting it in about 30 GPU-hours through a four-stage curriculum. On Full-Duplex-Bench it is the least likely open-source system to take the floor when the user merely pauses, with a turn-over rate of 0.170 on the synthetic subset against 0.584 for the next-best system; on InteractionBench it leads the open-source systems in spoken-task accuracy, groundedness and long-conversation coherence.

arXiv

Diffusion OOD detection stops trusting reconstruction alone: DUO adds a feature-sensitivity readout and reaches 0.897 average AUROC on 12 pairs

Diffusion-model out-of-distribution detection often relies on reconstruction error, yet OOD inputs can be denoised just as well or better, so those scores overlap or even invert. DUO adds a second readout on a frozen diffusion model: the sensitivity spectrum of intermediate denoising features to small input perturbations, with the feature layer, noise levels and two-channel fusion calibrated on in-distribution data alone. Across 12 image pairs DUO reaches 0.897 average AUROC versus 0.838 for the strongest baseline, EigenScore, at 0.28 seconds per sample.

arXiv

A legitimate deletion request can erase an unnamed entity: proxy unlearning turns shared representations into a targeted threat

Large language models encode distinct entities in shared parameter subspaces, and this knowledge superposition lets a legitimate unlearning request degrade sensitive entities that were never requested while removing the entities that were. The researchers name this exploitation the Proxy Unlearning Attack (PUA) and propose Superposition-Aware Unlearning Protection (SAUP), which detects at-risk entities at the representation level before unlearning and, afterward, decouples shared representations and restores them while keeping the requested entities suppressed. Across three knowledge domains and two model families, PUA produces selective target degradation and SAUP recovers most of the damaged knowledge.

arXiv

Discovery agents stop asking only true or false: swapping a hypothesis for rival explanations resolves 67% more questions at the same budget

Scientific discovery agents usually test one hypothesis at a time, so when evidence rejects it the agent learns that one explanation is unlikely but not which alternative the evidence supports. Contrastive Discovery replaces each discovery step with a set of mutually exclusive, collectively exhaustive competing explanations for the same question, and scores an experiment by how far its evidence redistributes belief across that set. Across six datasets spanning materials, astronomy, mortgage discrimination, online discourse, software engineering practice and social behavior, guiding search with contrast sets resolved 67% more questions than pointwise surprisal under the same experiment budget, and 38 blinded experts chose the same explanation on 78% of resolved questions.

arXiv

Foundation tensor networks generate ground states straight from the Hamiltonian, pinning glass-transition parameters across 500,000 disorder realizations

In disordered quantum systems, every disorder realization demands a fresh ground-state solve, so critical-phenomenon studies carry two layers of averaging. Foundation tensor networks (FTNs) use a transformer to generate the projected entangled pair state of the ground state directly from a graph embedding of the Hamiltonian, amortizing optimization over the whole Hamiltonian distribution. With GridBP and replica moment-generating functions, generating, fine-tuning and measuring the overlap parameter for one sample takes under three seconds, and finite-size scaling over 500,000 disorder realizations yields the critical parameters of the quantum Edwards-Anderson model.

arXiv

Matrix formulas compile straight into quantum block encodings: MARBLE matches or beats 37 of 39 published baselines

Quantum algorithms need exponentially large matrices encoded efficiently into circuits, yet such circuits have mostly been built by hand or assumed sparse. MARBLE takes a formula for each matrix entry, uses static multiplicity to guarantee a fixed number of distinct entry expressions, and applies algebraic rewriting plus cost analysis to emit OpenQASM. Across 74 matrices from fluid dynamics, PDEs and physics Hamiltonians, it matches or surpasses 37 of 39 published baselines in subnormalization and compiles very large dimensions in under five seconds on a laptop.

bioRxiv

Gene regulation beyond the single cell: Multicellular Transformer takes a whole tissue sample as input and predicts receptor-ligand pairs zero-shot

Single-cell foundation models usually take one cell at a time, so their gene representations lean toward cell-intrinsic regulation. Multicellular Transformer instead takes a set of cells from the same tissue sample as input, so each gene's representation carries its expression across that sample together with the sample's cellular composition. On known receptor-ligand pairs from CellChatDB it reaches a zero-shot AUROC of 0.71, against 0.65 for the second-best method, scGPT, and the gain holds across cell-cell contact, secreted signaling and ECM-receptor interaction types.

arXiv

SOAP's gains without SOAP's bill: FOAM halves the optimizer state and reaches Muon-class wall-clock time

The second-order optimizer SOAP leads on Transformer pre-training quality but pays higher optimizer state and a slower step, so practitioners often default to AdamW or Muon. FOAM keeps only the input-side preconditioning factor and replaces the dense rotated second moment with a row-and-column factored Adam-moment estimator; on GPT-2-scale pre-training from 124M to 760M parameters it matches SOAP and Muon on validation loss, uses half of SOAP's optimizer state, and runs at Muon-class wall-clock time.

arXiv

A reasoning model gains a no-token switch: HDM reads one forward pass's hidden state to decide whether to write a chain of thought

The hybrid decision model HDM puts token-free answering and full autoregressive reasoning inside one frozen backbone, with a lightweight self-routing block choosing between them before any token is generated. The block estimates the value of thinking, the accuracy gain the slow mode is expected to add over the fast one, and escalates only when that gain exceeds a price set per request. On MiniWoB, HDM reaches a reward of 0.469 against always-thinking's 0.519, cutting thinking tokens by 80% and wall-clock by 67%.

arXiv

Imitation policies learn a proprioceptive anchor: one measurement predicts sensitivity at new noise levels, but the protection fails in deployment

Imitation policies trained on noisy object-position readings do not discount the reading by the classical reliability ratio. They combine it with a proprioceptive anchor, weighting the reading by the ratio of reading noise to the anchor's error variance, and that variance belongs to the learner and is far smaller than regression on the demonstrations implies. Measured once, the law predicted ten held-out sensitivities across two robomimic tasks and two learner classes with mean absolute error 0.025, against 0.37 for classical errors-in-variables regression. In closed loop the policy produces its own state, and none of the 15 policies that passed the registered gates kept its open-loop protection against a persistent reading offset, though 8 followed the offset fully.

arXiv

Few-step diffusion students lose directions, not variance: ReDiverse restores diversity at one intermediate block

Distilling a diffusion model into a four-step student makes different random seeds produce nearly identical images and videos for the same prompt. By comparing teacher and student activations block by block, ReDiverse finds that the student keeps its total activation variance but concentrates it into far fewer directions, at one intermediate block during the first denoising step. ReDiverse-FT then uses LoRA at that block to regress the student's first-step activations onto the teacher's, while ReDiverse-TF perturbs activations along teacher principal directions; across four teacher-student pairs on SDXL, SD3, Wan 2.1 and MiniMax-H3, both restore diversity while largely preserving generation quality.

arXiv

Decoding-aware sequence-level losses are proved unbounded below: greedy and sampling objectives both diverge, so teacher forcing stays the practical choice for supervised autoregressive training

Autoregressive language models are usually trained by teacher forcing, predicting each token from the ground-truth prefix, while at inference they condition on their own generated prefixes, a mismatch called exposure bias. Two sequence-level losses built to remove that mismatch, one based on greedy decoding and one on autoregressive sampling, are both shown to be unbounded below and therefore unusable for learning. An autoregressive Gumbel trick shows the sampling loss is the expected greedy loss under Gumbel perturbations; a stop-gradient moment-matching direction is not the gradient of any loss yet is a descent direction for the autoregressive negative log-likelihood in function space, at the cost of high variance.

arXiv

Anomaly detection turns to diversity: VAS tops 148 embedding benchmarks without hyperparameter tuning

The Vendi Anomaly Score (VAS) redefines an anomaly as how much dataset diversity drops when a sample is removed, scoring samples through changes in the eigenvalue spectrum of the Vendi score rather than through density estimation or space partitioning. Across 148 ADBENCH image and text embedding benchmarks, VAS holds a better average rank than compared methods including ADERH, DN2 and SLAD, with no validation-set tuning; in controlled ImageNet contamination experiments it remains strongest as contamination rate and dataset scale grow.

arXiv

One long document, hundreds of questions: Parallel Prompting decodes them as parallel streams for a 3.7x speedup on SGLang

When tens to hundreds of questions are asked about the same long document, prefix caching removes only the repeated prefill; every decode step still re-reads the document's key-value cache once per question, so decoding is bound by memory bandwidth. Parallel Prompting packs those questions into a single request and decodes them as independent streams, reading the shared document once per step, with a stream-id mask evaluated inside the attention kernel so each answer sees only the document and its own question. On CodeLlama-7B with 8k–16k-token documents and 64–256 questions per document, it is 3.7x faster than stock SGLang with the FlashInfer backend on an A100 and 5.9–6.0x faster on an RTX A6000, winning all 144 paired runs.

arXiv

Diffusion models no longer need a reference to find memorized regions: JaRS reads Jacobian row sums

Diffusion models sometimes reproduce regions of their training images, and earlier localization methods read the diagonal of the denoiser Jacobian, which is also small in smooth regions and therefore needs a reference map to subtract; unconditional models have no null prompt and intermediate checkpoints are rarely released. JaRS instead takes the row sums of that Jacobian: local averaging keeps a row sum near 1 while memorized regions drive it toward 0, so no reference evaluation is needed. Paired with an early-step image-level detector as JaRS+, it reaches the best localization mIoU in seven of ten settings across Stable Diffusion v1.4, v2.1, Realistic Vision v5.1 and an unconditional DiT-B/2.

arXiv

Cross-view attention stops growing with views: TTB bakes it into a fixed-size MLP and holds 23 FPS under dense views

Decoder-only view synthesis models cache context views in a KV cache that grows linearly with view count, making rendering impractical under dense views; TTB instead bakes cross-view attention outputs into a fixed-size fast-weight MLP during reconstruction and queries only that state at render time. On DL3DV at 16, 32 and 64 views and 960p, TTB renders at a constant 23 FPS with quality close to DVSM and clearly above LaCT at the same constant complexity.

bioRxiv

ProtCOVER swaps per-variant conformal calibration for homology families: with no shift, pathogenic coverage on new families returns to the 0.90 target from 0.823–0.873

Zero-shot variant effect predictors calibrated conformally per variant let well-studied protein families set the error rate, leaving pathogenic coverage on new families at 0.823–0.873. ProtCOVER treats homology families rather than individual variants as the exchangeable unit and calibrates conditionally on label-free family descriptors; under the shift those descriptors define, its worse-class shortfall is 0.000, against 0.010 for group weighting and 0.029 for hierarchical conformal, at the price of larger sets. On 12,225 proteins outside the benchmark every method still falls short, with ProtCOVER at 0.027 and group weighting at 0.034, within the 0.01 margin.

TokenEconomizer lets coding agents learn from past sessions: 18.4%–30.9% lower cost across 277 real bug-fixing tasks

Developers use AI coding agents in the same codebases day after day, yet agents often start each session from scratch, rediscovering familiar facts, loading unused tools and repeating unsuccessful approaches. TokenEconomizer, an open-source plugin from UC Berkeley, Microsoft Research and Microsoft Security, turns completed session logs into guidance for later sessions without modifying the agent or its underlying model. Across 277 real bug-fixing tasks in six repositories, it cut average cost by 18.4%–30.9% for five agent and model combinations, saving in all 30 repository-combination pairs while task success rates stayed broadly stable.

bioRxiv

A digital twin of freely moving mouse V1 reproduces classic orientation and cell-type tuning when frozen and replayed with gratings

Activity in mouse primary visual cortex (V1) during free movement can now be predicted by a multimodal digital twin from retinocentric video plus eye and head movements, locomotion speed and pupil diameter. Across 223 neurons from three mice the model reached a mean prediction correlation of 0.31; frozen and replayed with drifting gratings, it recovered canonical orientation, spatial-frequency and contrast tuning and the systematic differences between broad- and narrow-waveform cells, and without retraining it matched the population representational geometry of an independent head-fixed natural-movie dataset as closely as two recorded mice match each other.

arXiv

Remote-sensing segmentation stops adopting every revision: GeoReVis keeps evidence from rejected masks and reaches 72.00% gIoU on LaSeRS

Language-guided remote-sensing segmentation often revises masks through feedback, yet a revision can damage regions that were already correct, and even the best whole mask can omit correct fragments held by weaker candidates. GeoReVis unifies execution-verified revision, candidate selection and cross-mask evidence consolidation: it adopts only corrections that improve the executed mask, partitions the spatial disagreements among candidates into composable regions, and jointly decides which regions to keep under the full instruction and image context. On LaSeRS it reaches 72.00% gIoU and 77.01% cIoU, and on ReVis-Eval it lowers the degradation rate from 17.90% for SocioReasoner∗ to 6.59%.

arXiv

Hyperlyve splits ReLU activation regions into parallel subtrees, enumerating 6.3 million regions in under 16 minutes

Hyperlyve turns activation-region enumeration for deep ReLU networks into parallel work over independent hidden-layer subtrees, then checks the Lyapunov conditions on the true nonlinear dynamics inside each affine region with sound and complete regional tests. Earlier enumerators were bottlenecked by per-neuron checks or network flattening: on Rejoin RL's 6.3 million 8-dimensional regions, RPM and Relucent both hit a one-hour timeout while Hyperlyve finished in under 16 minutes. It also characterizes verification of continuous piecewise-affine neural Lyapunov functions as co-NP-complete and embeds the verifier in a counterexample-guided synthesis loop.

arXiv

Self-driving labs drop the Q-learning rate: heteroscedastic Gaussian processes set a noise level per target and near the best return in fewer episodes across five benchmarks

Each campaign in a self-driving laboratory is a real reactor or culture run, and tens to hundreds of episodes leave no room to tune a learning rate as well. Heteroscedastic Gaussian Process Q-learning (HGPQ) places a Gaussian process directly on the Q-function, uses the Kalman gain as the learning rate, and estimates each Bellman target's noise variance with a second heteroscedastic Gaussian process so noisy targets move the estimate less. Pooled over five simulated benchmarks that include a chemical reactor and a bioprocess, HGPQ approaches the best return found by any method in fewer episodes than GP, kernel and deep RL baselines.

arXiv

SeisDFE, a 7B seismic foundation model, works across four tasks with a frozen encoder: two annotated F3 sections reach 0.82 whole-volume mIoU

Seismic interpretation is constrained by scarce high-quality annotations and by models trained separately for each task. SeisDFE, a 7B-parameter model pretrained on 5.34 million 2D images from more than 500 seismic volumes, uses three-stage pretraining with discriminative self-supervision, Gram feature anchoring and multi-resolution adaptation, and with its encoder frozen supports facies classification, salt detection, fault detection and relative geologic time estimation. Using only two annotated F3 sections, it reaches a whole-volume mIoU of 0.82, 12 to 19 percentage points above the baselines, with the largest gains under the scarcest supervision.

NVIDIA releases cuPotential molecular-dynamics primitives: its OpenMM plugin tops 3x end-to-end speedup on selected benchmarks

NVIDIA released cuPotential, a molecular-dynamics library that exposes modular primitives through a C API and ships an OpenMM plugin replacing the nonbonded force, Langevin integrator and Monte Carlo barostat. Across six systems from 24,000 to 400,000 atoms under NPT, newer RTX PRO 6000 Blackwell GPUs reached over 3x end-to-end speedup in selected benchmarks, while other tested GPUs often gained around 2x or more; these numbers come from standard, non-alchemical systems.

arXiv

Frozen DINOv3 heads already separate lesions; SMIA trains just 0.51M parameters to reach 80.99% mean mIoU across 20 medical and industrial tasks

Cross-domain few-shot segmentation transfers natural-image pretrained encoders to medical or industrial images using only a handful of pixel-annotated support masks. SMIA inserts a conditional branch after each attention head's aggregation inside a frozen DINOv3, letting the receiving patch's pre-attention local state decide how that head's message updates, so similar contextual messages can produce different or opposing corrections at different positions. Across 20 tasks covering five medical datasets and all 15 MVTec AD categories, it trains only 0.51M parameters and reaches six-dataset mean mIoUs of 77.40% at 5 shots and 80.99% at 10 shots, 3.61 and 3.93 percentage points above the strongest prior method at each budget.

arXiv

Pruned visual tokens keep the answer but shift the odds: DiFiT trains selection on the full model's response distribution

Visual token pruning for multimodal large language models is usually judged by whether answers stay correct, yet pruning can change the probabilities of alternative responses even when greedy answers do not move. DiFiT learns which visual tokens to keep from both the full-token model's response distributions and reference answers, with one ranking serving several retention ratios. On Qwen2.5-VL-7B it cuts sampling-induced negative flips by 30.3% relative to FastV at 10% retention and preserves 98.74% of baseline performance on MME at 30% retention.

arXiv

ElasticDino upsamples DINO features with elastic deformation: sharper segmentation and depth at 0.71 GB peak memory

Self-attention in vision transformers has quadratic cost, so feature maps stay at patch resolution and carry positional-encoding artifacts and boundary bleeding. ElasticDino is a lightweight adapter that exploits the observation that these artifacts are consistent across channels, using image-guided elastic deformation fields to correct all channels together during progressive upsampling and output pixel-aligned high-resolution features. On DINOv2 features it leads on several semantic segmentation, open-vocabulary segmentation and depth estimation benchmarks, at 108.41 ms average time and 0.71 GB peak memory, below the compared methods.

arXiv

A shared decision rule can now be corrected one owner at a time: 96.6% at 30 records, non-inferior to all seven baselines

Several organisations apply the same published decision rule while each holds private records showing it is wrong for its own items, and none will pool them. One approach compiles the published rule exactly into a small ReLU network, lets each party submit only typed declarations, and gates them on evidence before compiling each into a weight patch confined to that party's reserved units. In two pre-registered synthetic tests, the learner scored 96.6% at 30 records per party against 91.7% for the unchanged rule and was non-inferior within 2 points to seven tuned baselines; the same search without the compiled prior scored 84.0%.

arXiv

Deleting a message no longer deletes the memory: LinEra voids an entire derived chain with one forget request, and a thief holding every file reads nothing

An LLM agent's long-term memory is mostly written by the model itself, so deleting the original message leaves summaries, facts and embeddings in storage and backups. LinEra ties each memory's key to the keys of its parents, so forgetting a source is a single puncture of a puncturable pseudorandom function and every descendant becomes undecryptable at once. On LoCoMo conversations written by Qwen2.5-7B, an attacker holding every file and the current keys recovers 70 of 71 forgotten secrets after logical deletion and 67 with per-object or epoch keys, and none with LinEra.

arXiv

Refractory high-entropy alloys move past equimolar: a dislocation-genome plus dynamics framework predicts high-temperature strength and room-temperature ductility across the NbMoTaW space

High-temperature strength and room-temperature ductility in refractory high-entropy alloys have long traded off against each other, and most studies stay near equimolar compositions. This multiscale framework builds energy landscapes for edge-dislocation glide from atomic configurations with a dislocation-genome method, feeds the resulting local configurational forces into stochastic discrete dislocation dynamics, and obtains a linear scaling law between critical shear stress and force fluctuation. That law predicts high-temperature strength across the NbMoTaW composition space, which is then screened jointly with a ductility criterion based on the relative mobility of screw and edge dislocations.

arXiv

Erasing a group gap from a learning model does not close it: an audit prices the removal at about 0.021 Brier score on new Eedi classes

Adaptive tutors often pursue fairness by erasing the group signal from a learned model, but what is erased is the mean group contrast in predictions, not the real gap. The work gives a closed-form minimum extra error for this removal and a three-stage audit that fits on training units, announces the cost on validation units and tests on independent units. On three real data sets the erased model predicts almost none of an observed gap; on Eedi, with a knowledge-tracing network selected on separate units, the audit accepted the announced cost in all five splits, the measured cost fell inside the announced interval in all five, and the gap reduction transferred to new classes in four.

OpenAI News

OpenAI opens its internals to four outside assessors: Apollo, METR, Transluce and Goodfire move on-site to probe alignment

OpenAI has signed agreements with four independent assessors—Apollo Research, METR, Transluce and Goodfire—for a first phase of embedded assessments in which they work on-site at OpenAI offices on OpenAI devices and reach model checkpoints, source code, training and evaluation data, monitoring systems and the researchers behind the work, focusing on alignment and safeguards for models in training and internal use. Earlier external testing was episodic and limited in access; this round lengthens both access and duration, but no assessment findings have been released.

arXiv

Neural population dynamics without an explicit generative model: RP-GPdyn recovers interpretable latent dynamics from synthetic spikes and real recordings

Models of neural population dynamics usually specify three parts explicitly: the nonlinear flow field, the mapping from state to firing, and observation noise, and misspecifying any of them biases estimates of dynamics and latent trajectories. RP-GPdyn replaces the flow field with a Gaussian-process transition function and links latent states to neural activity implicitly through a recognition-parametrised model, avoiding a generative observation model. On synthetic spiking data, a random-dot motion decision simulation, mouse anterior lateral motor cortex delay activity and Neural Latents Benchmark motor-cortex data, it recovered behaviourally relevant latent variables and dynamics.

bioRxiv

NeuroMASS v4 learns curvature straight from spikes: near-positive in motor cortex, near-zero in a frontal timing circuit

NeuroMASS v4 makes the curvature of a hyperbolic latent space a learnable parameter trained end-to-end on extracellular spike recordings for neural population spike co-smoothing. Earlier latent-variable models such as GPFA, LFADS and NDT embedded population activity in Euclidean space, which cannot represent ring manifolds without distortion. Across four macaque-cortex datasets from NLB'21, learned curvature was 0.84±0.11 on motor-cortex MC_Maze and 0.71±0.14 on MC_RTT, but 0.07±0.04 on frontal DMFC_RSG, not significantly above zero.

arXiv

Weak residuals drive the network update directly: NeRI beats PINN and WAN on nonsymmetric and low-regularity elliptic problems

NeRI turns the solution of elliptic operator equations into a single network update driven by the current weak residual: at each step it builds a state-dependent linear functional whose derivative with respect to the trial function recovers that residual, so no global energy functional and no extra test network are needed. It can be read as a neural-tangent-kernel-preconditioned version of classical residual iteration, and penalty and hard-constrained parameterizations extend it to elliptic variational inequalities. In numerical experiments on a nonsymmetric problem with a corner singularity, a discontinuous-diffusion problem and an obstacle-type variational inequality, NeRI's relative errors were smaller than those of the compared PINN and WAN.

arXiv

Reaction models judged only by endpoints miss the point: UniRxnAxis splits the latent update, and its orthogonal residual carries most retrosynthesis utility

Reaction-prediction models are usually judged by endpoint accuracy, which does not show how a model represents a chemical transformation. UniRxnAxis unifies forward prediction and retrosynthesis as direction-conditioned endpoint redistribution in a shared latent space, making source states, transformed states and encoded target endpoints directly comparable, and decomposes the learned displacement into an endpoint-parallel component and an endpoint-orthogonal residual. On USPTO-50K retrosynthesis, decoding only the parallel state gives top-1 0.0997, while decoding only the orthogonal residual reaches 0.6303, close to the full state's 0.6441.

arXiv

COMPASS folds yield-table comparison and parameter inference into one diffusion model, and Milky Way data favor a NuGrid plus TNG hybrid

COMPASS puts two jobs of Galactic chemical evolution into one conditional score-based diffusion framework: comparing competing stellar nucleosynthesis yield tables and inferring galactic parameters, learning both posterior and likelihood directly from forward simulations. In controlled tests it identifies the generating yield table and recovers ground-truth parameters with accuracy comparable to existing simulation-based inference, at orders-of-magnitude lower cost than traditional sampling. Applied to high-precision abundances of 69 solar-type stars, it favors NuGrid asymptotic giant branch yields combined with the core-collapse supernova yields used in IllustrisTNG, and infers a high-mass initial mass function slope steeper than the canonical Salpeter value.

arXiv

Black-box processor probing moves off expert hands: ProcReverser's multi-agent workflow reaches 81.3% Pass@1 on the XuanTie C910

When processor implementation details are unavailable, security researchers can only infer hardware properties such as branch prediction and speculative execution from software experiments, redesigning tests, writing programs and interpreting results for every chip. ProcReverser hands that loop to five AI agents with separate roles, which use execution feedback to revise experimental plans and test programs and must check measurement evidence before a conclusion is issued. On 16 reference tasks for the XuanTie C910, five independent attempts gave 81.3% Pass@1 and 93.8% Pass@5; on 20 tasks for the gem5 RISC-V O3 model, 83.0% and 100.0%.

arXiv

Video diffusion drops the pretrained autoencoder: JVT reaches FVD 136.5 on UCF-101 with a plain pixel-space Transformer

JVT moves video diffusion back into pixel space, denoising with a large-patch DiT-style Transformer and no pretrained autoencoder, cascaded stages or U-Net-style skip connections. Pixel-space video generation had mostly relied on U-shaped backbones or cascades, while latent DiTs dominate. On UCF-101 class-conditioned generation JVT reaches FVD 136.5 against 477.97 for the latent DiT baseline Latte, and joint RGB-disparity generation raises RTA@5 from 72.4 to 79.5 in the camera-conditioned task.

arXiv

Robot multimodal alignment gains a structural check: UniOMA pulls vision, force and touch into one shared geometry

Multimodal contrastive learning for robot perception usually pulls paired samples of the same instance together without constraining how samples relate to one another inside each modality, so embeddings can be statistically aligned yet geometrically inconsistent. UniOMA adds a Gromov–Wasserstein regularizer to InfoNCE-style contrastive losses, builds an intra-modal similarity kernel per modality, and learns a cross-modal GW barycenter as a structural consensus that each modality aligns to. Across five robotic benchmarks, adding the term improves regression, classification and cross-modal retrieval in most paired comparisons and raises rank correlation between embedding structure and an input-kernel reference.

arXiv

DTW search stops using one ruler for every pair: WETW predicts per-pair weights and runs 40 times faster than the UCR Suite on long series

WETW changes the comparison rule in DTW search from a fixed distance to weights generated per series pair: two fixed-length embeddings go into a small symmetric network that predicts one weight per embedding dimension, without computing an alignment path. On eight datasets, the WETW-FV pipeline verifies a shortlist with exact DTW and reaches a mean recall@10 of 0.93 at 11.4 times the speed of the fastest exact method; on 48K-point series it stays 40 times faster than the UCR Suite at a recall@10 of 0.91.

arXiv

Euclid DR1 lens modelling drops per-image hand tracing: AstroVink-TREX separates lens, source and contaminants, with only 39 of about 3,000 systems failing on segmentation

Euclid DR1 holds more than 15,000 strong-lens candidates, and until now an expert had to mark the lens light, the lensed source light and the foreground contaminants to be masked in every image, which cannot keep pace with the catalogue. AstroVink-TREX replaces that step with a neural network that predicts three flux maps and derives the modelling centre, fitting region and conjugate points from them. Across about 3,000 expert-inspected systems, at a 75% inspector-agreement threshold, only 39 models fail because of a segmentation error.

NVIDIA Blog

NVIDIA unveils Rubin HPC GPU: over 50,000 CUDA cores per card to speed simulations for science agents

NVIDIA announced Rubin HPC at SC26, a new GPU in the Vera Rubin platform aimed at simulation and data processing for science, engineering and analytics workloads. The card carries more than 50,000 CUDA cores on a single GPU plus a sparse data acceleration capability that keeps those cores fed, and NVIDIA says the design scales data movement and compute together rather than raising compute alone. The product is expected to be available in 2028.

arXiv

EEG-to-image reconstruction stops leaning on alignment alone: an image teacher supervises denoising, and the method beats its same-backbone baseline on all seven THINGS-EEG metrics

EEG-to-image reconstruction has long aligned brain signals to pretrained visual embeddings and used them to condition a generator. Retrieval experiments here show that better alignment does not mean better reconstruction: retrieval improved by up to 77, while high-level reconstruction gains stayed modest and low-level fidelity showed no consistent improvement. An image teacher conditioned on DINOv2 features is trained first to express fixed encoder features as intermediate denoising representations, and an EEG-conditioned student generator aligns to these dynamic targets during denoising, with the teacher discarded at inference. On ten THINGS-EEG subjects, the method beats a SiT-XL/2 baseline trained on the same Brain-HiVE embeddings without representation alignment on all seven metrics.

arXiv

An EDA scripting agent stops guessing: reading back full execution state lifts Pass@1 from 19.7% to 76.5% across 449 tasks

For three EDA scripting interfaces, PyAether, SKILL and Tcl, an execution-grounded coding agent closes a generate-execute-read-back loop and reaches 76.5% pooled Pass@1 on the 449-task EDA-Scripting-Bench, against 19.7% for a pure LLM. With the model fixed, the observation axis spans 56.8 percentage points; with the harness fixed, the backbone axis spans 23.3 points. The authors frame this as a grounding gap: scripting correctness is bounded by how faithfully the agent observes the live design state, which model scale alone cannot supply.

arXiv

One pretrained transformer stops re-solving each dynamical inverse problem: shared computation yields both parameter estimates and trajectory forecasts

For a family of ordinary differential equations indexed by unknown parameters, a pretrained transformer can estimate those parameters and predict trajectories from a few noisy observations of a new system without updating its weights. The construction casts estimation as two-stage approximate maximum likelihood: a coarse search places the initial value inside a locally strongly convex region, then gradient descent contracts linearly; prediction evaluates a neural approximation of the solution map at the recovered parameters. Error bounds separate finite observations, finite pretraining tasks, context length and transformer size, and numerical experiments show error falling then plateauing as inference context grows.

arXiv

Clinical sleep staging without full-night labels: a 31k-parameter student trained on a quarter of patients closes in on full-label N1 scoring

Clinical sleep is more fragmented, with more awakenings, more N1 and less REM sleep, and automatic staging models trained on healthy sleepers degrade on patients. A 136M-parameter sleep foundation model, SleepGPT, is fine-tuned on labels from only 25% of patients per cohort and then supervises a 31k-parameter, 122 KB causal CNN-GRU that stages EEG and EOG in real time. A new transition-aware distillation loss, Gcv-TAD, raises per-patient macro F1 by 3.2 to 7.6 points on average over supervised training on the same labels in patient-wise 10-fold cross-validation, with the largest gains on N1, approaching full-label training.

arXiv

Read the biology before fitting: P2C cuts worst-form error from 13.096 to 1.095 as noise grows

Recovering ordinary differential equations from sparse, noisy biological observations tends to select structures that fit the data equally well while encoding different mechanisms. P2C first turns mechanistic prose from the literature into typed constraints tied to source spans, then rejects candidate equations that violate them before any numerical fitting. Across four published inflammation and immune-regulation models, adding the text constraints reduced the worst form error from 1.397 to 0.499 on clean data and from 13.096 to 1.095 under Gaussian noise perturbation.

arXiv

Vision-language models rewrite robot cost programs: 97.5% success on four tasks after five feedback rounds in simulation

Contact-rich robot manipulation has relied either on end-to-end learned policies or on hand-engineered rewards and constraints for known scenes. Objectives as Policies has a pretrained vision-language model write an editable multi-stage cost program, which one shared sampling-based model-predictive controller executes in a digital twin and revises from execution statistics. Across four tabletop tasks, simulation cumulative success reached 90% with at most two feedback rounds and 97.5% with up to five; programs that succeeded in simulation then reached 80% success in the real world with perturbed object locations.

arXiv

Verified SQL history need not be rebuilt each time: MEM-SQL lifts execution accuracy from 37.1% to 50.0% on HARE while cutting input tokens by 76.6%

When users query the same database repeatedly, previously verified SQL often already contains the joins or computations a new question needs, yet standard benchmarks score each question in isolation and never see that reusable history. MEM-SQL stores history in three layers, whole-query templates, reusable SQL blocks and schema concepts, and admits a layer only when every requirement assigned to it clears that layer's threshold, consulting the database schema only when no layer qualifies or the generator judges the evidence insufficient. Across six language models on the 1,600 HARE requests, mean execution accuracy rises from 37.1% for direct generation without history to 50.0%, with 76.6% fewer input tokens and 68.8% lower latency on average.

arXiv

Agent training data no longer needs pure expert demos: JudgeMix judges each step and lifts pass@1 by up to 8 points on AppWorld

Supervised fine-tuning of AI agents on expert trajectories leaves training data mismatched with the target model's inference-time state distribution, so the model is brittle once it deviates from the ideal path. JudgeMix uses a judge to decide at each step whether the target model's proposed action is semantically correct, hands control to the expert only when the model struggles, and reweights expert supervision by that judgment. On the AppWorld and AutomationBench tool-use benchmarks, JudgeMix reaches the highest pass@1 for both Qwen3.5-2B and Qwen3.5-9B, up to 8 and 10 points over pure expert demonstrations.

arXiv

A keyboard the attacker never heard still gives up its keystrokes: Decker cuts the cross-keyboard accuracy gap from about 40 points to under 18

Acoustic side-channel attacks had been validated only with a single keyboard, a fixed microphone, a quiet room, and training and testing on the same device; on unseen keyboards, strong unimodal and multimodal baselines lose roughly forty accuracy points. Decker combines keyboard signature normalization, domain-adversarial disentanglement, supervised cross-keyboard contrastive alignment and acoustic style randomization in one training framework to learn embeddings that discriminate keys but stay invariant to keyboard and user, narrowing the cross-keyboard generalization gap to under eighteen points, and a language model then decodes over the top acoustic candidates to turn noisy key predictions into coherent text.

arXiv

ChlCast forecasts chlorophyll-a at about 4 km: against the Copernicus operational system, errors fall by up to 70% on satellite reference fields

ChlCast is an observation-driven deep learning model that issues daily surface chlorophyll-a forecasts out to 10 days at roughly 4-km resolution between 60°S and 60°N. The operational Copernicus global ocean biogeochemical forecasting system previously ran on a 0.25° grid, where fine-scale structure is weakened by coarse-grid averaging. Against satellite Level-4 chlorophyll-a fields, ChlCast lowers mean absolute error and root mean square error by 70.74% and 47.63%; against independent BGC-Argo in situ observations, the two errors fall by 35.03% and 13.07%.

AG-UI 1.0 freezes its spec: one bidirectional event stream now carries the agent-to-interface link

AG-UI has reached version 1.0, fixing the connection between AI agents and user-facing applications as a single bidirectional event stream: the agent emits typed events and any client can render them, so each framework no longer needs its own wiring. Agents did not fit the request/response model, because they stream tokens, call tools mid-response, update shared state, pause for a human and spawn subagents, and every framework did this differently. The 1.0 specification is frozen and backed by a JSON Schema, the TypeScript, Python and .NET SDKs are generated from it, and 1.0 is backwards compatible in both directions with 0.x.

NVIDIA's Vera Rubin debuts on MLPerf Training with the fastest time on six of seven benchmarks, about 2x per GPU on DeepSeek-V3

NVIDIA's Vera Rubin architecture entered MLPerf Training v6.1 for the first time, and its NVL72 platform posted the fastest training time on six of the seven benchmarks, up to 2x the highly optimized GB300 NVL72 baseline. On the 671-billion-parameter DeepSeek-V3 workload, Vera Rubin NVL72 cut training from 8.03 to 4.14 minutes at 1,024 GPUs, while the round's fastest absolute result, 1.38 minutes, came from a 10,240-GPU GB300 NVL72. NVIDIA reports above 97% overall scaling efficiency for Vera Rubin from 256 to 1,024 GPUs.

OpenAI News

ChatGPT for Healthcare puts agents inside hospital software: reading screens, clicking, and moving records while clinicians keep review

OpenAI is bringing computer-use agents to ChatGPT for Healthcare that carry out multi-step work across existing healthcare software, alongside improved trusted clinical search and more than 20 new healthcare plugins. Healthcare teams previously had to piece together information by hand from the EHR, hospital policies, scheduling systems, and payer portals; the agents read screens, click, type, and move between applications, with access controls and audit logs. In a blind evaluation spanning multiple specialties and multi-turn conversations, physicians rated the updated clinical answers more highly across all eight dimensions over more than 1,300 reviews.

arXiv

Two policies for one rendezvous: a learned mixture guidance law meets 10 m and 1 m/s on all 600 multi-revolution low-thrust test missions

Multi-revolution low-thrust rendezvous must combine a long orbital transfer with meter-level terminal control, and a single policy struggles to cover both state scales. This guidance law has a far-range expert read modified equinoctial elements and a terminal expert read the relative state in the target's local frame, while a temporal-attention gate blends their thrust commands continuously. On 600 test missions it completed every rendezvous at 10 m and 1 m/s tolerances, cutting mean flight time by 50.6% and velocity increment by 44.5% relative to 4 km distance-based hard switching.

arXiv

Calpha lets a general chat model decide when to fold: mean TM-score 0.808 on single chains, still behind the direct baseline on long chains

Calpha wires protein folding into a general chat model: Qwen3.5-9B reads a sequence with ESM2 residue features, emits a <structure> token when asked to fold, and conditions a frozen SimpleFold-3B head on its own residue hidden states to produce all-atom coordinates. Under a fixed folding request this configuration reaches mean TM-scores of 0.808 on FoldBench-334, 0.635 on CASP16 domains and 0.900 on a CAMEO-style PDB set, but on 300–600-residue chains the mean paired differences are −0.044, −0.160 and −0.025, all below the direct ESM2→SimpleFold-3B baseline.

arXiv

Molecule-language models get a structure lesson: SAGA trains only, lifting retrieval R@1 by up to 7.82 points

Molecule-language models learn from paired structure-text contrastive objectives, yet they treat a structurally related nonmatch and an unrelated molecule as the same negative class. SAGA uses Centered Kernel Alignment during training to align pairwise relationships among neural molecular embeddings with fingerprint Tanimoto similarity, adding no parameters and requiring no fingerprints at inference. On MolCA and Atomas, every reported molecule-text retrieval metric improves, with Atomas-L text-to-molecule R@1 rising from 43.15 to 50.97, and non-exact retrieval shifts toward closer structural analogues.

arXiv

Same WZW current algebra, two neural realizations, different errors: level–rank duality carries charged correlators across paired theories

Neural network field theory writes field-theoretic correlators as expectations over network parameters, yet one continuum current algebra can be realized by different networks. This work builds positive-level WZW current algebras from bosonic Wakimoto fields and fermionic Fourier–Berezin fields, and uses level–rank duality to connect the charged sectors of paired theories. At fixed regulator and finite width, direct and coset constructions differ in error, spectral distribution and regulator choice shift accuracy, and source-sector measurements predict paired-sector responses without measuring the target preparation.

arXiv

Running weather backward: Rewind-AR concentrates the precursors of a pinned extreme, but up to 22% of its members go unstable within 66 hours

Given an observed extreme-weather end state, can the varied precursor paths that lead to it be sampled, as scenario planning needs? The video-diffusion emulator cBottle-video previously produced 66-hour final-conditioned ensembles, but they were underdispersed, geostrophically inconsistent, and reproduced the conditioning frame only approximately. Rewind-AR instead uses a FourCastNet 3 backbone as a backward-in-time probabilistic autoregressive model; for the 2021 Pacific Northwest heatwave and Hurricane Ian, its leading antecedent 500 hPa geopotential-height modes explain far more across-member variance than cBottle-video's, at the cost of up to 22% of members becoming numerically unstable within 66 hours.

arXiv

Rotation invariants can blur highly symmetric molecules: an extended HIP-HOP basis gives the lowest error among HIP-HOP variants on methane and still runs faster at inference

Rotation-invariant representations can map geometrically distinct atomic environments onto the same descriptor in highly symmetric settings, limiting the accuracy of machine-learned interatomic potentials. A lightweight benchmark built around configurations with multiple-fold symmetry exposes degeneracies shared by invariant and equivariant architectures, and guides an extension of the invariant HIP-HOP basis to higher angular order. On a randomly displaced methane dataset of more than seven million configurations, the extended basis gives the lowest error among HIP-HOP variants and, at about 36K parameters, nearly matches EquiformerV3 while running faster at inference; both still fail on the (3/4) ring configuration.

arXiv

Legal judgment models keep guilty verdicts after the decisive fact changes: PROSE cuts damage to supported rulings from 17.5% to 8.5%

Legal judgment prediction models often keep a guilty verdict after a legally decisive condition has been altered, a tendency the authors call guilty bias. PROSE frames the correction as a paired constraint: the unsupported member of a case pair should flip, while its counterpart, differing in one decisive condition, should stay. On Qwen3-8B and GLM-4-9B it performs this selective repair through a one-time offline parameter edit, and at matched repair rates it reduces damage to supported verdicts from 17.5% to 8.5% on Qwen and from 24.3% to 7.1% on GLM.

arXiv

Long-sequence distillation stops handing the teacher the whole answer: ATI-OPD rewrites only the earliest divergent segment, and students' average accuracy on their own beats standard on-policy distillation

On-policy distillation trains a student on its own responses using token-level teacher feedback, but as responses grow longer the student's prefix drifts from the teacher's own reasoning and later feedback weakens, bringing gradient spikes and extremely long responses. ATI-OPD has the teacher insert a short replacement at the earliest high-disagreement segment, hands generation back to the student, and excludes persistently divergent suffixes from the distillation loss. Across five mathematical reasoning benchmarks and two 1.5B teacher-student pairs, average student-only accuracy exceeds standard on-policy distillation.

arXiv

Candidate trajectory scoring drops per-candidate rollouts: RFWorld reads a shared scene memory and reaches 90.5 PDMS on NAVSIM

End-to-end driving must pick one of several candidate trajectories, and earlier world-model evaluators explicitly predicted future states for each candidate over the planning horizon, so inference cost grew with candidates and horizon while prediction errors accumulated. RFWorld instead builds one trajectory-agnostic shared scene memory, shapes it during training with a lightweight temporal readout under semantic supervision, and lets candidates query it through sparse spatial readout at inference. On NAVSIM-v1 with a ResNet-34 backbone, RFWorld reaches 90.5 PDMS, above DriveDPO at 90.0 and ResWorld at 89.0 in the reported table.

AI2's XYBench tests 8,000+ real questions to see whether AI assistants catch a user's mistaken premise

AI2, with Carnegie Mellon, NVIDIA and Johns Hopkins, released XYBench, a benchmark of 8,115 real questions drawn from Stack Overflow, Stack Exchange, hand-curated examples and WikiHow, testing whether AI assistants recognize the mistaken premise behind a user's request. On Stack Overflow questions, the five evaluated models offered a solution to the stated request in 75–86% of responses but a solution to the underlying goal in only 59–71%; even when the user's intent was supplied, misconception identification did not reliably rise.

bioRxiv

Liver senescence read beyond marker lists: spatial causal models link stress inputs to a matrix-deposition loop

Across 11 human liver Visium spatial transcriptomics samples, two causal-discovery algorithms with correctness guarantees were extended to spatial data, yielding BOSSsrsi and LGLsrsi models with 224 and 38 direct upstream genes and 659 and 53 direct downstream genes. Overlap with the SenMayo and SenUMN marker lists is limited, the largest being SenMayo covering 13.4% of the BOSSsrsi upstream list. Pathway analysis of the BOSSsrsi lists describes senescence-associated signaling as integrating cellular stress inputs and immune responses into a self-sustaining loop of integrin-mediated TGF-β activation and continuous matrix deposition.

Harmonic

Aristotle extends formal verification to Rust: it can prove unsafe blocks free of undefined behavior and caught a rounding bug in a Solana order book

Harmonic's Aristotle extends formal verification from C and Verilog to Rust, letting developers prove code matches its specification before shipping rather than relying on tests to catch flaws. It claims to prove code has no undefined behavior, including inside unsafe blocks, and that Rust rewritten from C behaves the same as the original. On the open-source order book Manifest, Aristotle proved 28 results across the whole order book and found a rounding bug the team has since fixed.

bioRxiv

RNA outside cell bodies stops being discarded: SCOUBI maps axon-dendrite interfaces across the mouse brain

Much of the RNA in spatial transcriptomics falls outside segmented cell bodies and is usually discarded, yet neuronal mRNAs can be carried far from the soma into axons and dendrites. SCOUBI bins this extrasomatic space, trains a shallow neural network on known axon and dendrite marker genes while requiring significant ligand-receptor co-expression between neighboring axonal and dendritic bins, and calls loci where the two converge with ligand-receptor colocalization interfaces. On mouse hemibrain Xenium data it produces axon, dendrite and interface maps that match known neuroanatomy, with interface transcriptomes enriched for synaptic genes and reproducible across independent samples and two imaging platforms.

arXiv

A team can be modeled as a ball in an energy bowl: ANCHOR forecasts four-person drift with 52 readable parameters

Collaborative mixed-reality headsets already record who speaks, who looks at the same image and who stands close, but prompting a team before it drifts needs a plant model that learns from few teams, stays bounded when iterated and can be read by a designer. ANCHOR treats a four-person team as a ball in an energy bowl whose bottom blends recent behavior, the team's own history and other teams' norms, written as a switched port-Hamiltonian system with 52 parameters. On 17 held-out four-person teams, about 24% of a deviation remains after 32 s against a predicted 30%; conversation-collapse warning reaches AUROC 0.85, and accuracy matches a black-box baseline with about 40,000 parameters.

bioRxiv

A cerebellar-medial temporal lobe circuit learns only to predict the next state, yet infers context and transfers to new ones

A circuit assembled from known anatomy pairs a cerebellar forward model, which predicts the next state conditioned on a latent from the medial temporal lobe, with a medial temporal lobe module that updates that latent from the cerebellum's prediction error. Trained only on next-state prediction, it cut held-out velocity error across twelve uncued motor contexts from 0.0348 to 0.0138, below the 0.0223 of a context-free model, and organized contexts by family and parameter so a downstream motor cortex could retrieve old contexts and compensate for unseen ones on first reach.

arXiv

Auction reserve prices now come from listing text: across roughly 65,800 used-vehicle auctions, the net automatic-sale revenue gap reaches $165.8 million

Online auction houses must set reserve prices for highly differentiated products, while established approaches require managers to specify product attributes in advance. This work fine-tunes a pretrained text embedding model against the structural likelihood of auction bids, so the description representation targets the valuation distribution that determines the reserve, then estimates a generalized Pareto tail on the other half of the sample and derives the reserve in closed form. Across roughly 65,800 US online used-vehicle auctions, with seller values calibrated to match the sale rate, the model implies a counterfactual net automatic-sale revenue increase of $165.8 million, or 14.6% above the incumbent reserve policy.

OpenAI News

OpenAI puts frontier models in NATO cyber defenders' hands: core institutions have used up to 300 billion raw tokens a month

OpenAI is working with NATO's Communications and Information Agency (NCIA) to give NATO's core institutions access to advanced cybersecurity models for reviewing software, investigating suspicious activity, validating vulnerabilities, and developing and testing fixes. NATO has used up to 300 billion raw input and output tokens per month in this work. The same models also serve European defenders: the EU's ENISA used them to identify four vulnerabilities in software used across EU institutions, with fixes published for all four, and CERT Polska used OpenAI models to help identify six vulnerabilities in MikroTik RouterOS, whose fixes CERT Polska confirmed prevented the attacks it had observed.

arXiv

Signature kernel ridge regression gets finite-sample error bounds, and the excess risk still decays on heavy-tailed path data

Path signatures can approximate functions on path space arbitrarily well, but how much data is needed to pin down a suitable signature functional was not settled. Working in the signature kernel ridge regression framework, this work derives finite-sample excess-risk bounds and consistency guarantees under moment conditions on the signature kernel and the response; when the driving path is Brownian motion and the standard moment condition provably fails, a weaker finite-moment assumption with the Fuk-Nagaev inequality yields the same order of decay.

Phones and cloud split generative inference: MAGS bids instead of stampeding, cutting p99 latency 17.1%

A generative query's cost is unknown until the answer is written, so devices reserve resources blind and many crowd one server at once. MAGS has each device forecast token count and generation entropy with a four-layer probe on its local model, then bid for cloud capacity against a broadcast congestion price, choosing its per-token privacy parameter in the same bid. Over ten seeds it cuts p99 latency 17.1%, energy per token 13.6% and raises throughput 20.7% against a reactive PPO baseline given identical observations; an eight-device, two-server testbed drops the cross-device correlation of offload decisions from 0.41 to 0.08.

arXiv

A quantum generative model's Fisher trace is pinned to its parameter count, so certification shifts to a two-metric escalation rule

For quantum circuits that prepare real-amplitude states from half-angle rotations, the classical Fisher information matrix equals the quantum one off a measure-zero set, so its trace is fixed at the parameter count 2n−1 and says nothing about fit quality. The analysis turns to the Fisher eigenvalue rank and to two distributional metrics, Wasserstein-1 and maximum mean discrepancy: the rank only counts the support of the converged distribution and falls to chance on a full-support benchmark. For two analytically solvable bitstring defect families, the tolerance-mismatch window in which the two metrics disagree opens only at certification-grade tolerances and closes as the qubit count grows.

arXiv

Coherent scattering beyond average relaxation: a predictive-sufficiency test shows which dynamics correlation functions leave out

X-ray photon correlation spectroscopy (XPCS) usually compresses a speckle movie into intensity correlation functions, and that compression can discard information needed for later dynamics. An "information microscope" compares prediction from measurable XPCS observables alone with prediction after adding a learned representation of the present speckle sequence; a gain means the conventional description has left information unused, and a library of candidate observables built from speckle intensities is then searched to recover it. In simulated Bragg scattering, two added observables raise the prediction score from 0.347 to 0.412, while the learned representation retains a 0.047 increment.

NVIDIA Blog

NVIDIA puts a four-node DGX Spark cluster on the desk: 512GB of memory runs 700B-parameter models locally, no shared-cluster queue

At SC26, NVIDIA introduced the DGX Spark HPC Developer Kit, a four-node ring of DGX Spark systems for a researcher's desk. Multi-node scientific workloads and AI models have typically waited on shared GPU clusters; the kit links nodes directly over ConnectX-7 with no switch, combines 512GB of memory and 20TB of NVMe storage, and can deploy models up to 700B parameters locally under NVFP4. The kit is on sale, while the HPC SDK playbook is still to come.

arXiv

Wildfire risk assessment goes geolocation-aware: RoPE anchors street-view semantics into satellite tiles, reaching 0.2476 harmonic F1 over five risk classes

Structure-level wildfire vulnerability assessment has struggled to combine parcel-scale ignition pathways with community-scale geographic coverage. Wildfire-SSR has a vision-language model turn pre-event street-view panoramas into structured text, then uses latitude and longitude to anchor each description at a relative pixel position inside its satellite tile before fusing it with satellite features through a decoupled cross-attention decoder. On roughly 37,000 structures affected by the 2025 Eaton and Palisades fires in Los Angeles, the configuration using Rotary Position Embedding reached a building localization F1 of 0.6735 and a five-class pixel-level risk harmonic-mean F1 of 0.2476, above both unimodal baselines and an unanchored multimodal configuration.

arXiv

Quantum repeater protocols stop splitting the network into routes: reinforcement learning lifts end-to-end entanglement rate above routed swap-ASAP on intersecting networks

On non-linear quantum networks, entanglement routing splits the network into disjoint paths and runs chain protocols on each in parallel, shrinking the search space. This work instead uses reinforcement learning to search the larger non-routing protocol space directly, and finds end-to-end logarithmic-negativity rates above routed swap-ASAP on an eight-station intersecting network and other small networks, even in the parameter regime where the latter is widely regarded as state of the art. The authors argue routing should be discarded in favor of non-routing methods, and give a network score tied to the size of the improvement.

arXiv

Low-dose CT denoising drops the whole-image trade-off: RouteDiff reaches 40.95 dB PSNR on Mayo 2020 within ten network evaluations

Diffusion-based low-dose CT denoising previously applied one structure–detail trade-off to the whole image, set each step's restoration emphasis by the timestep, and derived each next state recursively from the current one, carrying errors forward. RouteDiff routes detail pixel-wise to where a confidence map trusts it, judges the restoration stage from current features and route disagreement through a state hypergraph, and re-enters the trajectory from an anchored estimate at a fixed step. With the same ten network evaluations as its baseline, it achieves the highest PSNR and SSIM on four datasets, 40.95 dB and 95.13% on Mayo 2020 and 45.18 dB and 97.23% on Mayo 2016.

Hugging Face hands language-model routing to OpenRouter: one Hub token now reaches open-weight models, with automatic retry on another provider

Hugging Face has switched routing for language-model requests from its own router to OpenRouter's, so users still call open-weight models served by providers on the Hub with the same Hugging Face token and the same API. OpenRouter picks a provider for each request from live price, speed and reliability, and retries on the next provider inside the same call if the first fails; previously Inference Providers chose one provider and returned its error on failure. Image, video and audio generation still run through Hugging Face's own inference path, not through OpenRouter.

OpenRouter Blog

OpenRouter's routing now drives Hugging Face inference: 160 models pick price, speed and reliability per request

Chat-model inference on Hugging Face now runs through OpenRouter's provider routing, which picks a provider for each request from live price, speed and reliability and fails over automatically. Previously developers had to choose among providers themselves. OpenRouter reports that a live experiment when it rebuilt this routing in October measured 46% fewer requests hitting a capacity error, 16.9% faster time to first token and 16.3% higher throughput, at 6.0% higher cost per request. 160 models from 38 labs are available this way today.

bioRxiv

Cryo-FIB-SEM volumes stop drowning in charging artifacts: a two-stage restoration lets a standard FIB-SEM membrane predictor read them directly

Cryo focused-ion-beam scanning electron microscopy images cells in a near-native hydrated state in 3D, but charging artifacts, noise and weak membrane contrast have long hampered interpretation. A post-acquisition two-stage deep-learning pipeline translates cryo images toward the conventional room-temperature FIB-SEM domain: stage one suppresses large-scale artifacts slice by slice, and stage two uses five consecutive slices to enhance membrane contrast. On 164 held-out synthetic images, stage one raised mean paired PSNR by 16.74 dB over degraded inputs; across four experimental datasets, visible artifacts decreased and membranes became clearer, and the pretrained CebraNET membrane predictor could be applied directly to the processed volumes without retraining.

arXiv

SMEFT operator space without enumeration: a reinforcement-learning agent finds 26 low-overlap, high-quality solutions among 912 operators

With 912 baryon-, lepton- and lepton-flavour-conserving SMEFT operators, enumerating multi-operator combinations becomes a combinatorial explosion. ReinforceSMEFT casts operator-by-operator assembly as sequential decision-making: an actor–critic policy with about 40K trainable parameters learns which operator to add in an environment of 132 observables, 24 of which deviate from the Standard Model by more than 3σ, and is rewarded for solutions that overlap little with those already found. Against random-subset and random-sequential searches, the policy builds multi-operator theories more efficiently; combining ten independent training runs yields 26 high-quality solutions with Jaccard indices below 0.1.

arXiv

Two equivalent encodings of one rotated box mislead semi-supervised detectors: Oriented Semi-DETR lifts mAP at every DOTA-v2.0 label ratio

In aerial images a rotated box has two parameter encodings that describe the same rectangle, so a semi-supervised detector can penalize and discard a student box that is already correctly localized. Oriented Semi-DETR removes that ambiguity at the three points where a pseudo box enters training: the regression target follows whichever encoding is closer to the student prediction, the matching cost takes the minimum over both encodings, and pseudo-label denoising starts earlier and is weighted by teacher score and encoding margin. On DOTA-v2.0 at 10%, 20% and 30% labeled data, mAP exceeds the best DETR-based baseline by 1.78, 2.18 and 2.59 points, with larger gains at the strict IoU threshold.

langchain.com

LangSmith rebuilds around the agent: cross-framework agents get one owner, one ID and one governance view

LangSmith has been reorganized around an agent record that ties the build, run, improve and govern stages to a single agent identity. Previously, traces sent to LangSmith sat in a flat project hierarchy with no association to a particular agent, and traces from one agent across environments were not linked, so platform teams struggled to say which agents existed, who owned them, and what they cost or how they performed. Each agent record now carries an owner, a globally unique immutable ID and a type, and agents on other platforms can register through OpenTelemetry Protocol and be measured the same way.

OpenAI News

OpenAI opens its CSAM image classifier to vetted groups for free, with a dedicated signal for AI-generated abuse

OpenAI released an image classifier that detects child sexual abuse material (CSAM) and says it is the first broadly available CSAM classifier with a specific signal for whether an image may be AI-generated. In an evaluation on user-uploaded images, the new classifier was estimated to flag 48% fewer images than existing protections while finding roughly the same number of abusive images, and flagged images were about twice as likely to contain abuse. Access is free through the OpenAI API to vetted organizations, and the National Center for Missing and Exploited Children helped with training and evaluation.

arXiv

Structured neural Lyapunov functions guarantee positive definiteness by design, and PLNet reaches the largest certified region on the inverted pendulum

For known nonlinear discrete-time systems, a controller and a control Lyapunov function can be trained jointly while positive definiteness of the candidate no longer rests on penalties at sampled states. Input convex neural networks and Polyak-Łojasiewicz networks, which carry lower and upper quadratic bounds and are therefore positive definite by construction, satisfy positivity and radial unboundedness directly, leaving only the closed-loop decrease condition for formal verification. On an inverted pendulum and a path-tracking benchmark, the PLNet attains the largest base and expanded certified regions on the inverted pendulum, and all five seeds certify in a single round.

arXiv

A gated vision state-space model can erase closed tokens exactly, but its gate maps do not explain malignancy calls

In a lung-nodule malignancy concept bottleneck model, zeroing the mask on the selective state-space block's convolution input and giving closed tokens zero readout weight makes the concept and task logits exactly invariant to closed-token content; a path-by-path knockout over 15 cross-validated models reproduces this in every run, with the largest logit change exactly 0 when the conditions hold and between 4.1 and 6.7 when they do not. Gating is neither necessary nor sufficient, and the gate maps behave more like a mask-supervised support set than an explanation of the malignancy call.

arXiv

Invertible networks learn Morse Lyapunov functions: hyperbolic equilibria get stability certificates on compact sets

For dynamical systems with a hyperbolic equilibrium, invertible neural networks can approximate a Morse Lyapunov function on compact sets and thereby certify global asymptotic stability. Previously, neural learning of Lyapunov functions guaranteed approximation only in norms that do not control derivatives, which is not enough for a Lyapunov inequality. The approach writes a Morse Lyapunov function as the squared norm of a diffeomorphism and approximates that diffeomorphism with a neural-ODE invertible network in a Sobolev norm.

arXiv

Stochastic trajectory generation moves past one dimension: SINN constrains joint distributions with projected densities, and joint-density error falls monotonically as projection frames are added on a coupled-oscillator benchmark

Statistics-informed neural networks (SINNs) had been demonstrated only on one-dimensional stochastic processes, and independently trained one-dimensional models cannot reproduce the joint distribution across components. The multidimensional SINN constrains that joint distribution through densities of several one-dimensional linear projections and balances multiple statistical objectives with adaptive loss weights; on a benchmark of two coupled Duffing oscillators, joint-density error falls monotonically as projection frames are added, and that error enters neither training nor model selection.

mastra.ai

Mastra launches Custom Factory program: embedded engineers build a software factory around your codebase

Mastra has launched a Custom Factory program in which forward-deployed engineers embed with a customer's team and co-build a software factory for that codebase on Mastra, running in the customer's own stack. Most engineering teams still adopt AI one tool at a time, such as a hand-run coding agent, an issue investigation agent and a code review agent. Mastra says its own software factory now handles 70% of its issues and that its PRs have grown 6x since launch, with most work starting in Slack through Shipyard, its Slack agent, and running through to a pull request ready for review.

NVIDIA Blog

NVIDIA's Vera Rubin debuts on MLPerf: up to 2x performance per GPU in training and over 10x throughput per watt for agentic inference

NVIDIA's Vera Rubin platform entered MLPerf for the first time, with submitted results across training, post-training and agentic inference. In training, Vera Rubin NVL72 delivered up to 2x more performance per GPU than the highly optimized GB300 NVL72 in its debut; on the DeepSeek-V3 benchmark it scaled from 256 to 1,024 GPUs at about 97% efficiency, cutting training time from 16 minutes to about four. In its first MLPerf Endpoints submission, Vera Rubin NVL72 showed more than 10x the throughput per watt of GB300 NVL72 at the same response speed in a provisional Kimi K3 result.

Liquid AI

Liquid AI ships d1.2 decision models with audio input and five new task types, as d1.2-ultra leads the Decision Index by two points

Liquid AI released the decision models d1.2 and d1.2-ultra, adding audio input and five structured-output tasks: reading a value from text, extracting spans, locating objects in images, estimating a number with a range, and selecting every applicable option. On the Artificial Analysis Decision Index v2.1, d1.2-ultra scores 75, two points above the previous best decision model; across 13 benchmarks and 22,505 questions, its calibration error is 0.8 points, versus 1.2 for d1.2 and 7.2 for GPT-6 Luna on its Decisions API. No public benchmark yet measures the five new tasks, which were evaluated on internal test sets built from held-out splits of public datasets.

bioRxiv

No trillion-sequence library needed: MobGen designs monobodies from scratch that bind SARS-CoV-2 RBD and neutralize B.1

Monobodies, small binding proteins derived from a human fibronectin type III domain, previously required screening libraries of 3×1013 random sequences by phage display to find high-affinity binders against a viral antigen. MobGen chains RFdiffusion, AlphaFold2 backpropagation and ProteinMPNN into one computational pipeline; targeting the SARS-CoV-2 B.1 spike receptor-binding domain it generated 11,450 sequences, yielded 7 binders, and one of them neutralized B.1 pseudovirus at micromolar potency. Adding a C-terminal tail raised production about 7-fold on average and binding affinities reached 22.1–424.8 nM, but neutralization stayed around 2–25 µM and no monobody covered all tested variants.

arXiv

After a disruption, 3RO re-plans station routes by fixing then selectively releasing: it strictly beats the unrestricted MILP on most instances within 180 seconds

When a disruption hits a station area, routing, platforming and timing must be revised within minutes. The 3RO framework first fixes routes predicted to stay stable using route-specific classifiers, releases a few fixings when the reduced model is infeasible, and later releases more around interacting trains to recover alternatives. On 226 real station disruption instances, within a 180-second budget it found strictly better incumbents than the unrestricted MILP in 56.7% of all instances and 77.4% of hard instances.

arXiv

Dynamic survival predictions get time-resolved explanations: DynSurvX cuts computation by 41–80% and can refresh them during follow-up

Longitudinal follow-up data make survival predictions more accurate but harder to interpret, and existing explanation methods are either tied to specific model classes or limited to static covariates. DynSurvX splits temporal attribution into two steps: it first screens longitudinal features by overall importance, then adaptively subdivides time windows only for those features, yielding time-resolved explanations that update as new measurements arrive. On simulated and two clinical datasets, agreement with theoretical importance rises with prediction-model performance, and computation time falls 41–80% relative to the temporal explanation baseline WindowSHAP.

NVIDIA Blog

NVIDIA expands CUDA-X libraries and hands them to AI agents: Event Horizon Telescope correlation drops from over a month to eight minutes

NVIDIA has expanded its CUDA-X libraries with cuPhoton, cuPotential and cuSpa, and made them callable by AI agents to speed up scientific computing. cuPhoton.xBeam cuts Event Horizon Telescope correlation analysis from over a month to eight minutes; cuPotential accelerates existing molecular simulation software by up to 3x; the open-source cuSpa reaches a 900x speedup over CPUs for cell polygon calculations at Memorial Sloan Kettering. These figures belong to different libraries and different settings.

Vercel News

Seedance 2.5 Flash lands on AI Gateway: faster and cheaper, with output close to the standard model

ByteDance's Seedance 2.5 Flash is now available on AI Gateway as a faster, lower-cost version of Seedance 2.5. It generates video with synchronized audio from text prompts or image, video, and audio references, with clips of 4 to 30 seconds at up to 720p and 24 frames per second in landscape, square, or portrait, and it can also edit or extend existing videos. ByteDance says its output is close to the standard model in human realism, reference-based editing, and prompt adherence.

arXiv

A three-agent team now sits between buildings and the distribution grid, lifting task success from 38.0% to 69.0% on 200 synthesized tasks

Dispatch interactions between buildings and the distribution grid have long depended on specialists from both domains translating requests and hand-building simulation and optimization workflows. B2G-Agent splits that process across three collaborating agents: a Translator for bidirectional semantic alignment, an Actor that decomposes tasks into directed acyclic graphs and calls EnergyPlus and pandapower, and a Reviewer that combines semantic and evidence checks and routes failures back for correction. Across 200 synthesized interaction tasks, its interaction quality score rises from the baseline's 66.71 to 85.57 and task success from 38.0% to 69.0%, while operational constraints are satisfied.

arXiv

CT-to-PET synthesis learns from its own candidates: MedicalDPO cuts BBDM's LPIPS by 11.67% with preference fine-tuning

Brownian Bridge diffusion models can generate several PET candidates of differing quality under the same CT condition, yet standard training aligns each sample to ground truth and leaves those relative differences unused. MedicalDPO turns rankings along spatial, intensity and structural fidelity into automatic preference pairs, then fine-tunes with a relative denoising-error preference objective jointly with the original Brownian Bridge loss. On the BBDM and CPDM backbones, LPIPS, MAE, SSIM and PSNR all improve; with multi-pair rotation on BBDM, LPIPS and MAE fall by 11.67% and 13.22%.

arXiv

Laryngeal lesion classification stops relying on appearance alone: PADSNet adds vocal-fold vibration waveforms to a video model, lifting F1 about 0.20 over the strongest video baseline

In laryngoscopy, high-risk, low-risk and normal tissue often overlap in static appearance, while differences in vocal-fold symmetry, closure and periodicity may carry more discriminative information. PADSNet models video appearance features alongside glottal-distance waveforms extracted from vocal-fold segmentation masks, and uses nine biomechanical indices to gate and recalibrate the visual features. On a patient-level stratified three-class clinical video dataset of 397 patients, PADSNet reaches 0.8452 accuracy and 0.8089 F1, above the best compared video model, InternVideo2, at 0.6084.

ElevenLabs Blog

ElevenLabs to acquire Japan's Kotoba voice AI lab, putting its Koto model behind Japanese Agentforce Voice

ElevenLabs announced it is acquiring Kotoba, a Japanese voice AI lab, in a deal expected to close later in Q4 2026 subject to customary regulatory approvals and closing conditions. Founded by researchers in 2023, Kotoba makes Koto, a voice model covering Japanese text-to-speech, speech-to-text and speech-to-speech for simultaneous interpretation, already used by Salesforce to power Japanese-language conversations in Agentforce Voice. The companies say the combination will pair Kotoba's Japanese research with the full ElevenLabs stack, with plans to build out local sales leaders and continue investing in infrastructure in Japan.

NVIDIA Blog

Lambda links over 10,000 GPUs with co-packaged optics: rack-scale network latency falls about 4%, optics run at 36°C median

Lambda has deployed NVIDIA Photonics co-packaged optics (CPO) switches in an AI factory, packaging optical engines alongside the switch chip to connect NVIDIA GB300 NVL72 racks carrying more than 10,000 GPUs. Company tests show network latency down about 4% at rack scale and up to about 10% at pod scale, with throughput matching conventional pluggable optics. Against pluggable modules, the switches are claimed to deliver up to 5x better power efficiency and 15x higher reliability, use 4x fewer lasers, and run optics at a 36°C median versus 51°C.

arXiv

Medication-trajectory anomaly detection stops scoring whole records: L2G-Joint pinpoints the few suspicious visits

Suspicious deviations in longitudinal medication trajectories often sit in only a few visits, and scoring the whole account can dilute that local signal. L2G-Joint builds multi-granularity proxy anomalies constrained by a drug knowledge graph and reimbursement-catalogue categories, adds knowledge-consistent plausible variants as negatives, and gives every scoreable visit a shared anomaly score that is selectively aggregated by Top-20% into an account-level decision. On the City-A proxy benchmark it reaches a Macro AP of 0.7844, Macro AUROC of 0.7707 and Localization Macro AP of 0.7215, relative gains of about 2.4%, 3.6% and 9.2% over the strongest respective comparators.

PFN and SEMIFIVE team up on MN-Core L2000 inference chip: double the memory bandwidth, design handoff set for Q1 2027

PFN and SEMIFIVE will collaborate on the MN-Core L2000 series, an AI-inference semiconductor that is the successor to PFN's MN-Core L1000 series, planned for sale in 2027, with more than twice the memory bandwidth of L1000. To handle the structural complexity of stacking logic and DRAM vertically, SEMIFIVE provides design implementation and supply-chain management while manufacturing uses Samsung Foundry's platform, and the two plan to complete design and hand off final data in the first quarter of 2027. The approach bets on memory bandwidth as the bottleneck for generative-AI inference, but power, heat and yield issues in 3D stacking still have to be resolved before mass production.

arXiv

Landslide runout no longer needs a solver at every step: a physics-constrained emulator runs about 10 ms per step and roughly 50 times faster over a 110-second rollout

Landslide runout forecasts usually come from numerical simulations that solve depth-averaged flow equations, and repeated runs for parameter sweeps are costly. A physics-constrained neural network was trained as a single-step surrogate that advances the thickness and velocity fields by a fixed time interval. On the test set, single-step root mean square errors were about 0.06 m for thickness and 0.11 m/s for velocity; after 110 autoregressive steps they were about 0.21 m and 0.09 m/s, with the full rollout taking about 1.1 s against about 16.7 s for the numerical solver.

NVIDIA releases SkillEvaluator: static quality checks for agent skills, then a with-and-without skill comparison

NVIDIA has released SkillEvaluator, a tool that checks the quality of AI agent skills and measures their effect on task results. It runs quality checks first, then compares runs with and without a skill; a skill is a folder of instructions and supporting files defined by the Agent Skills specification. The bundled text-analyzer example scores 88.8/100 (B) on the static quality check against a default passing threshold of 70, while publication evidence remains marked INCOMPLETE.

Databricks Blog

Databricks opens an enterprise AI index: equal task success, but a roughly 32x gap in cost per task

Databricks has launched a public AI index that combines anonymized usage data from more than 10,000 enterprises on Unity Gateway with coding-agent evaluations built on real tasks from its private codebase, updated weekly. The index shows enterprise model choice fragmenting quickly, with open-weight models rising from 7.3% of enterprise token traffic in July to 30.9% in November. In the coding benchmark, GPT-6 Luna and Claude Opus 5.5 both reached an 89.1% task success rate, yet averaged $0.05 and $1.62 per task respectively.

NVIDIA releases Nemotron for Science with a post-training recipe: on drug-metabolism prediction, SFT lifts accuracy 12% and RL another 17%

NVIDIA released Nemotron for Science, a collection of models, datasets, tools and recipes that lets research teams turn experimental records into a verifiable-reward environment and adapt a general multimodal model into a domain-specific science agent. In a drug-metabolism example, the model takes a parent compound and its transforming enzyme and outputs the metabolite SMILES; at a 16,384-token output budget, supervised fine-tuning raised Nemotron for Science accuracy by 12% and RL post-training lifted it a further 17%. The recipe also offers an optional teacher-model distillation warm start for cases where RL converges slowly.

High accuracy inside the field, collapse across it: random forest drops from R² 0.90 to 0.329 on unseen blocks, while sequence models hold near 0.44

Pixel-level winter wheat yield prediction at one Swiss farm shows that high within-field interpolation accuracy does not carry over to unseen blocks. Random forest reached R² about 0.90 under a coordinate-grouped split but averaged 0.329 under contiguous within-year spatial block cross-validation, while CNN-A-GRU and Mamba held 0.444 and 0.434, rising to 0.485 and 0.464 with regime-stratified generative augmentation. Seven-pass phenological sampling also beat 90-day seasonal averaging for all three models.

Vercel News

Vercel's AI Gateway adds Private Inference: self-hosted and fine-tuned models now share one key, SDK and log

Vercel's AI Gateway has added Private Inference, which connects enterprise self-hosted, fine-tuned or prerelease models as private providers. Such models previously needed separate keys, client code and logs; they now share the same authentication, SDKs, APIs and observability as catalog models, and switching only changes the model ID to private/<slug> while the gateway translates each request to the endpoint's format. Private models are visible only to their team, stay out of the public catalog, store credentials encrypted, and can enable Zero Data Retention per provider.

PFN and SEMIFIVE team up on MN-Core L2000: a 3D-stacked inference chip aimed at more than twice L1000's memory bandwidth

Preferred Networks and SEMIFIVE announced a collaboration to develop the MN-Core L2000 Series, PFN's next-generation chip for AI inference, planned to deliver more than twice the memory bandwidth of the preceding MN-Core L1000 Series to ease the bandwidth bottleneck in generative AI inference. The series keeps the 3D structure that stacks logic chips vertically with DRAM; PFN handles processor architecture and logic design, SEMIFIVE provides 3D-IC design and supply chain management, and the logic chip is manufactured on Samsung Foundry's platform, with tape-out scheduled for the first quarter of 2027. The collaboration is already underway, but the chip has not been taped out and the bandwidth figure is a planned target.

arXiv

Quantum experts stop slicing brain regions by index: grouping by functional network lifts strong-connection retention from 43.9% to 62.4%

Quantum mixture-of-experts models for high-dimensional spatio-temporal data have split input features by index, cutting the dependencies that link them. Feature-Partitioned QMoE instead assigns 180 brain regions to four quantum experts by functionally coherent network, so each expert sees one neuroscience-defined circuit. On ADHD classification from ABCD resting-state fMRI, this grouping matches arbitrary splitting in test AUC while raising the share of the top 5% strongest functional connections kept inside experts from 43.9% to 62.4%, and it supports circuit-level attribution.

arXiv

A frozen vision encoder gets a narrower view: vividRGB enlarges local crops and stitches them back, beating DINOv3 and SAM3 on edge F1 across five ecological image domains

vividRGB 0.1.0 packages an inference-time procedure as a Python library: without retraining a vision encoder, it restricts each input to enlarged overlapping local crops and maps their descriptors back into source coordinates to build a dense feature field. On grassland, fungal networks, airborne forest, coral reefs and phytoplankton microscopy, a frozen DINOv3 ViT-B/16 with this procedure reaches domain-mean RGB-edge boundary F1 of 0.290 to 0.596 at 5% focus and 0.244 to 0.441 at 10% focus, versus 0.028 to 0.076 for original DINOv3 and 0.020 to 0.090 for automatic SAM3. The cost is substantially higher extraction time and GPU energy, about 175.8 s per image at 5% focus.

arXiv

PhaseCLIP brings the phase spectrum into frequency-domain adaptation of a frozen CLIP: 84.28% mean accuracy across five remote-sensing benchmarks and 4.36 points on the hardest one

PhaseCLIP extends frequency-domain adaptation of a frozen CLIP visual encoder from the amplitude spectrum to the phase spectrum, and lets one set of module weights learn phase first and refine with amplitude, so no parameters are added. Under a matched protocol on EuroSAT, PatternNet, UC Merced, RSSCN7 and LSRS, with 1 to 32 shots and eight baselines, it reaches 84.28% mean accuracy, 1.93 points above the amplitude-only FreqDIMFT and 4.36 points above it on the 113-class LSRS.

Asimov

Asimov and Emergent sign two cell-line campaigns: CHO Edge to build high-titer research banks for two Ebola antibodies

Asimov and Emergent BioSolutions announced a cell-line development partnership covering two full campaigns, running from vector design and synthesis through production and shipment of a high-titer research cell bank, for two monoclonal antibodies targeting Ebola. Emergent aims to obtain high-titer cell lines using Asimov's CHO Edge System, which combines a GS knock-out CHO host, a hyperactive transposase, a genetic parts library and Kernel, Asimov's AI-driven genetic design software. The work sits at the cell-line development stage, and the companies describe an intent to accelerate candidate antibody development without reporting titers or antibody results.

Hacker News

Axonnic's forearm sleeve reads muscle signals to decode finger motion and estimate fingertip force, leaving hands bare

Axonnic released a forearm sleeve that records muscle electrical activity with 64 channels of surface electromyography to decode finger joint angles and estimate fingertip force while the hands stay bare. The team argues video captures movement but not the effort behind it, and that barely visible muscle activity, such as holding a grip, is the key signal for sewing and delicate assembly. In one 215-second recording made four days after training without recalibration, the sleeve decoded 20 joint angles with 6.9° error, compared against Apple Vision Pro hand tracking.

arXiv

NVIDIA's 4GB Nemotron 3.5 Bolt E4B tops 21-benchmark average while emitting fewer output tokens

NVIDIA released Nemotron 3.5 Bolt E4B 4GB, an open small language model targeting roughly 4 GB of memory for local voice assistants, interactive characters and robotic or IoT interfaces. It is post-trained from Gemma-4-E4B-it with reinforcement learning alone, strengthening multi-turn tool use, instruction following and input grounding. Across 21 benchmarks it posts a higher average score than Gemma-4-E4B-it and Qwen3.5-4B in both reasoning and non-reasoning modes, while generating fewer output tokens on average.

AI

Ask Maps adds itineraries, agent-booked appointments and list search: itineraries and list search ship wherever it is offered, while route tips and booking start in the U.S.

Google added three features to Ask Maps: day-by-day itineraries with numbered map pins, agentic booking that finds beauty, pet care and repair appointments and can call businesses, and search across a user's saved lists with live availability. Itineraries and list search work wherever Ask Maps is offered, while route tips and appointment finding launch first in the U.S.; Ask Maps also begins expanding to 30 European countries, bringing it to more than 200 countries and territories.

Vercel News

Tencent's Hy4 Flash lands on AI Gateway: full Hy4 feature set at roughly a tenth of the per-token price

Tencent's Hy4 Flash is now callable through AI Gateway, with an introductory 50% discount on input, cached input and output tokens through November 29, 2026. It is the lower-cost model in the Hy4 family, keeping the full Hy4 feature set at roughly a tenth of the per-token price: reasoning with preserved thinking across turns, tool calling, structured outputs, prompt caching and image input, with the same 1M-token context window. Reasoning can be turned off for requests that don't need it, suiting high-volume coding, agent and document workloads.

arXiv

Cardiac cine MRI synthesis now traces motion: CardioTrace couples deformation fields and segmentation to curb drift over long sequences

CardioTrace ties cardiac cine MRI synthesis to explicit motion tracing: one branch predicts residual frames relative to the end-diastolic reference, another estimates time-resolved deformation fields from that reference, and segmentation labels of the left ventricle, right ventricle and myocardium at end-diastole and end-systole constrain anatomy. On the public M&M and ACDC cardiac cine MRI datasets it improves perceptual quality and temporal coherence over representative GAN- and diffusion-based baselines, while better preserving ventricular morphology through the cardiac cycle.

arXiv

Liquid food no longer needs a label: 2.4 GHz OFDM signals estimate carbohydrates, fat and protein directly

The carbohydrate, fat and protein content of liquid foods can now be estimated directly from 2.4 GHz OFDM radio signals passed through the liquid. Such quantification previously relied on package labels, laboratory analysis or manual user logging, while RF methods mostly identified liquid type or gave coarse nutrient fractions. ICOT first infers the ingredient classes that drive dielectric behavior, then regresses the three macronutrients; across 55 beverage classes and 28,420 captures, in-distribution MAE is 0.101 g/100 ml and leave-one-class-out MAE is 0.485 g/100 ml, though generalization to novel formulations remains the main open problem.

ClinicalKey AI now reads the whole chart inside the EHR: every answer carries patient data and citations from about 1,000 journals

Elsevier says ClinicalKey AI can now read a patient's entire medical record inside the electronic health record and evaluate treatment plans against the company's evidence-based clinical content, with every answer carrying chart data and citations drawn from full-text articles in roughly 1,000 journals. The tool is used at more than 1,200 hospital sites worldwide; previously it answered from content starting at a blank prompt, and now it starts from a specific patient. San Juan Regional Medical Center in New Mexico is using the patient-context feature to check each patient's care with clinical and quality teams smaller than those of large academic health systems.

AMD Newsroom

AMD opens Lux to scientists at SC26: the TOP500 No. 14 machine becomes the DOE Genesis Mission's first dedicated AI factory

At SC26, AMD announced that the Lux SI + HPC supercomputer at Oak Ridge National Laboratory is open to scientific users and serves as the first dedicated AI factory for the U.S. Department of Energy's Genesis Mission. Lux ranks No. 14 on today's TOP500 list with 220.79 petaflops of measured HPL performance, and it puts AMD EPYC processors, Instinct MI355X GPUs, Pensando networking and ROCm software on one platform for both AI and traditional modeling and simulation. AMD also reports its technology in 202 TOP500 systems, more than 40% of the list, and in more than half of the Green500 top 50.

Federated load event recognition stops sharing semantics unconditionally: FedReST edges an architecture-matched FedAvg by 0.00524 Event F1 on REFIT 2014 at 10% labels

Federated non-intrusive load monitoring lets households collaborate without centralizing raw power traces, yet sparse local labels and household-specific operating patterns make cross-client semantic evidence unevenly reliable. FedReST splits collaboration into two admissions: the server uses reliability-gated partial optimal transport to accept only semantic mass that both sides support, and each receiving client filters the resulting update through function-preserving gradient projection so locally mastered states are not disturbed. On frozen REFIT 2014 rolling folds with 10% visible labels and matched two-epoch local budgets, corrected-core FedReST shows paired differences of +0.00524 in Event F1 and +0.00104 in Window F1 against an architecture-matched Head+BN FedAvg reference.