Weekly Recap

State-space models can achieve $1.8\times$ speculative decoding throughput and $1.47\times$ faster standard decoding by caching SSM inputs instead of their full state.

This narrative review explores current applications of AI in the intensive care unit (ICU), along with associated ethical, medico-legal, and regulatory challenges, and highlights the need for prospective validation studies.

This systematic review synthesizes current evidence (2018–2026) on the effects of anthropogenic warming on amphibians, reptiles, birds, mammals, and fishes, examining physiological stress, range shifts, phenological mismatches, and extinction risks, and discusses conservation strategies including thermal refugia protection, wildlife corridors, and nature-based solutions.

This review aims to highlight the philosophy of permissive cardiotoxicity, provide a framework for risk stratification of cardiovascular toxicities, describe systems of care considerations for successful implementation of permissive cardiotoxicity, and review the existing literature and guideline recommendations for specific cardiovascular toxicity mitigation.

This narrative review explores current applications of AI in the intensive care unit (ICU), along with associated ethical, medico-legal, and regulatory challenges, and highlights the need for prospective validation studies.

This systematic review synthesizes current evidence (2018–2026) on the effects of anthropogenic warming on amphibians, reptiles, birds, mammals, and fishes, examining physiological stress, range shifts, phenological mismatches, and extinction risks, and discusses conservation strategies including thermal refugia protection, wildlife corridors, and nature-based solutions.

This review aims to highlight the philosophy of permissive cardiotoxicity, provide a framework for risk stratification of cardiovascular toxicities, describe systems of care considerations for successful implementation of permissive cardiotoxicity, and review the existing literature and guideline recommendations for specific cardiovascular toxicity mitigation.

This review critically evaluates recent advances in AI-driven intelligent agriculture, with emphasis on machine learning, deep learning, explainable AI (XAI), generative and agentic AI, multimodal foundation models, and digital twins, together with the sensing and computing infrastructures required for their deployment.

The experimental results show the effectiveness of the proposed DLUMIS-TCF model in predicting traffic congestion compared to other methods, and can be useful for urban traffic management, congestion control, and intelligent transportation systems.

Participation in a wellness program was associated with improvements in self-reported dietary behaviors and physical activity frequency and larger controlled studies with validated measures and longer follow-up are needed.

This review discusses marine holobionts around three integrative themes, including bioactive compounds, industrial enzymes and probiotic interventions, and discusses the growing biotechnological potential of marine holobionts.

A strong evidence base is identified for multimodal imaging-guided prediction, particularly cardiovascular magnetic resonance (CMR) and transthoracic echocardiography (TTE)-derived global longitudinal strain measurements and how these diverse data can be integrated into multimodal risk models is considered.

The notion of an “operational-protective trajectory” is introduced to describe certain forms of regulated action that remain functionally organized under extreme stress exposure and the potential operational benefits and psychological costs associated with these adaptive configurations are discussed.

Millions of genome-scale metabolic models can now be solved orders of magnitude faster, revealing previously intractable metabolic pathway alternatives.

Diet-induced metabolic shifts in mouse muscle are revealed to be muscle-specific, with some showing reduced glucose utilization while others maintain or enhance TCA and redox activity.

A data-driven CANN framework discovers an eight-term hyperelastic strain-energy function that accurately models diverse venous biomechanics, outperforming existing constitutive models across multiple vein types.

Random RNNs exhibit a rich Lyapunov spectrum beyond the maximal exponent, with finite-time dynamics governed by a matrix Dyson equation and stationary behavior by an operator equation that captures the full history of local gains.

Unlocking cognitive state prediction from single-channel EEG by aligning brain signals with task semantics, achieving 30% zero-shot accuracy and paving the way for wearable neurotechnology.

A Bayesian adaptive design recovers full neural recruitment curves with the efficiency of threshold-hunting methods, enabling faster, more powerful intervention studies.

The Allee effect, often overlooked, is critical for IAS eradication feasibility, transforming the management landscape from inevitable persistence to a basin of attraction for extinction.

Selfish learners can converge to persistent cooperation in social dilemmas, but only if their learning rates are heterogeneous.

Molecular motor slack, not just attachment/detachment, fundamentally alters transport dynamics, requiring new analytical tools beyond standard quasi-stationary assumptions.

Training a single surrogate for amortised GSA slashes computational costs for comparing dosing protocols and patient groups in simulated clinical trials, recovering parameter rankings with 20x fewer evaluations than direct analysis.

Phototoxicity in optogenetic bioprocesses can be mechanistically modeled by a latent intracellular stress state that modulates growth, transcription, and translation.

Human reference labels for medical AI are more variable than previously assumed, with nearly 20% of reader-anchor pairs flipping AI-agreement state across annotation conditions.

Chromosomes aren't just miniatures of their genomes; they exhibit structured heterogeneity in functional element composition, revealing conserved organizational principles across vastly different life forms.

Key contribution not extracted.

Concept-based attribution for any black-box vision model is now possible without model weights or concept datasets, even for generative tasks.

Maximum degree in tournaments is provably $Θ(n^{3/2})$ for quantum algorithms, matching a long-standing conjecture.

This systematic review synthesizes the evolution of automatic paper ECG digitization following the PRISMA framework and analyzes 78 studies, and examines the digitization pipeline across three stages: image preprocessing, waveform localization and extraction, and signal calibration and reconstruction.

Ongoing efforts to optimize treatment sequencing, refine perioperative strategies, and integrate comprehensive molecular profiling are expected to further improve outcomes and continue generation of high-quality clinical evidence and international collaboration will be essential to advance the management of BTC.

The ITEK framework aided in understanding critical factors for AI/FN wellbeing and land-based healing efforts hold promise in preventing substance use and negative affect and promoting overall wellbeing among AI/FN adolescents.

With the utility of multidisciplinary care models, careful employment of evolving new treatment strategies, it is possible for patients to achieve their fertility desires while at the same time being protected from the risks imposed by their genetic mutations.

These findings identify a stromal-mediated metabolic immune evasion pathway in gastrointestinal tumors, uncovering a potential strategy to potentiate CAR-engineered therapies.

Satellite data can forecast regional desertification trends years in advance, but individual cell-level degradation remains largely unpredictable.

Domain knowledge injection unlocks superior performance in weakly annotated bioimaging, revealing nuanced ALS pathology and drug efficacy beyond automated feature extraction alone.

Wi-Fi HaLow enables stable, high-throughput wireless image transmission from within the body for capsule endoscopy, achieving over 12 Mbps with sub-500mW power consumption.

Multimodal decision models struggle with breast ultrasound classification, with even the best configurations falling below majority-class baselines on challenging datasets, revealing significant benign-to-malignant confusion.

Medical imaging JEPAs systematically bias learned representations away from clinically relevant pathologies by prioritizing predictable anatomical context.

Asynchronous Q-learning's last iterate converges to $\varepsilon$-accuracy in $\widetilde{O}(H^4/(\mu_{\min}\varepsilon^2)+\mathfrak h_\star)$ transitions, with initialization cost additive and independent of the effective horizon.

Neural networks can precisely encode arbitrary topological spaces, with their expressivity directly tied to the Betti numbers of their output, not just decision boundaries.

Nanobody design yield jumps 5x by dynamically allocating compute across scaffolds and epitopes at inference time, eliminating the need for upfront configuration commitment.

Chronologically ordered pre-training creates "specialist" models that learn new knowledge orders of magnitude faster than generalists.

Rapidly reconstruct high-fidelity MRIs by learning direct, stochastic transitions between k-space states, bypassing the need for iterative local updates.

Parallel sampling can achieve $\log \log n$ rounds for certain distributions, a stark contrast to the $(\log n)^2$ rounds required by previous analyses.

Finite-sample guarantees for discrete conditional diffusion models reveal that optimal error rates are achievable with a finite number of denoising steps, matching theoretical bounds for distribution estimation.

Hidden neural networks can be cryptographically secured within model weights, making them provably undetectable by current elicitation methods.

Deep-space relays can achieve superior information freshness and fairness by jointly optimizing bundle scheduling, dropping, and flow control using a masked Lagrangian PPO that prioritizes mission-critical data.

Quantile loss consistently beats MSE on tabular regression, but global early stopping still compromises performance in noisy regions, necessitating input-dependent label smoothing.

A new optimization framework, PISO, achieves state-of-the-art function-evaluation complexities for partially informed gradients by cleverly tracking and centering on the unknown residual.

Even

Experts miss significant predictive signal in raw data that can be recovered by removing only the *shared* canonical components between raw and expert embeddings.

A single denoising network can now serve as a plug-and-play prior for both diffusion and flow-based models, achieving state-of-the-art inverse problem solutions through exact Gaussian conditioning.

Industrial recommender systems can now deprecate nearly half their features without performance loss, unlocking significant efficiency gains.

Optimal $\widetilde O(\delta^{-1}\varepsilon^{-3})$ convergence for nonsmooth, nonconvex stochastic optimization with constraints is now achievable, matching theoretical lower bounds.

Coding agents can achieve higher solve rates by restoring and branching from full sandbox snapshots, especially when progress relies on external service states.

Imagine learning new robot skills without forgetting old ones, purely through internal simulation.

Inequality constraints, previously a post-processing step, now directly inform probabilistic PDE solvers, slashing reconstruction errors by up to 63% and arbitrage rates by 90%.

Achieve state-of-the-art 4-bit LLM inference without calibration data by deriving activation scales directly from weight reconstruction error.

A novel $\mathrm{SemiCat}$ architecture achieves discrete diffusion-level performance on TinyGSM and outperforms one-hot flows at scale by addressing self-conditioning and embedding geometry in continuous diffusion language models.

Pretrained tabular foundation models, when adapted with a simple readout and contrastive learning, unlock state-of-the-art dataset representations for retrieval and shift detection.

Discrete graph diffusion models, believed to generate novel graphs, are actually memorizing training data, with this bias appearing only when conditioned on corrupted training examples.

A single algorithm now achieves optimal convergence rates for stochastic non-convex optimization across average smoothness, standard smoothness, and noiseless settings, adapting dynamically without knowing any smoothness constants.

A novel multi-objective optimizer achieves $O(T^{-1/4})$ convergence with stochastic gradients, independent of task count, by integrating norm-constrained oracles and entropic weight updates.

Rare events under the prior pretraining distribution expose persistent amortization error in prior-data fitted networks, even at scale.

Joint safety violations are now traceable to individual agents, unlocking robust credit assignment in cooperative MARL even with team-level cost functions.

Halving oracle calls in momentum methods can slash training loss by up to 43% without sacrificing convergence.

Geospatial foundation models for disaster assessment exhibit severe shortcut learning, with performance plummeting from 0.80+ to 0.63 accuracy when evaluated on unseen disaster types, and simple building counts alone achieve comparable performance.

LLM agents trained with MCGPO achieve up to 5.16% higher success rates by intelligently reusing rollout data and constructing context-aware action preferences.

GFlowNet sampling accuracy is fundamentally limited by the square root of balance loss and a parameterization incompleteness residual, with convergence rates established for diverse backward policy update strategies.

DP-SGD's $(\varepsilon,\delta)$ guarantee is insufficient; a new information-theoretic framework provides a computable lower bound on reconstruction distortion, even for an adversary with full training knowledge.

Counterfactual predictions can be made with significantly higher accuracy and generalization by bypassing explicit causal graph estimation and directly learning from observational data via amortized GNNs.

Speculative decoding now achieves 4x concurrency and 3.7x output throughput on 100K context LLMs by intelligently managing KV cache residency across GPU and host memory.

Outlier dimensions, not global coefficients, are the key to accurate Transformer attention linearization, enabling a 9% MSE reduction and improved downstream performance.

State-space models can achieve $1.8\times$ speculative decoding throughput and $1.47\times$ faster standard decoding by caching SSM inputs instead of their full state.

Task diversity, not just task complexity, significantly enhances the richness of neural network internal representations in controlled mathematical environments.

A single benchmark now enables city-scale, multimodal OD flow forecasting across six modes and sub-district granularity, revealing that scaling spatio-temporal models and naive joint training often fail.

Diffusion models can now adapt to entirely new data distributions at inference time without retraining, by learning an in-context score operator from a few examples.

LLM test-time sharpening via power sampling doesn't actually target the intended power distribution; its gains come from simple sequence selection.

Unconditional graph generation can now capture complex node heterogeneity and higher-order structures by jointly diffusing graph edges and latent node sociability, bypassing the Erd\H{o}s--R\'enyi graph limitation.

A single attention module dynamically switches between linear prefilling and full decoding, achieving $21.9\times$ faster prefilling and Transformer-level long-context recall without the quadratic bottleneck.

First-order policy gradient methods can achieve minimax-optimal sample complexity for general parametric policies without restrictive assumptions or second-order operations.

Automated LLM agents achieve state-of-the-art performance by tuning specialized neural network components, outperforming baselines by up to 2491x on differential equation solving and 13pp on classification.

A simple scoring signal is insufficient for self-distillation; explicitly generating a critique before scoring unlocks corrective credit and enables models to solve previously intractable reasoning problems.

A 2B-parameter WAM can be distilled to run on edge devices at 11x the speed and 18.5x fewer FLOPs, enabling real-time robotic manipulation with minimal performance degradation.

Unlearning data deletions in federated fair k-means can be orders of magnitude faster than retraining, especially when entire clients are removed.

Synthesized audio-visual datasets now retain strong cross-modal semantic correspondence, boosting downstream model accuracy by over 15 percentage points and slashing distillation costs.

End-to-end calibration is achievable by treating the calibrator as a bilevel optimization problem, yielding accurate and well-calibrated models without extra parameters or datasets.

Calendar encoding choice can degrade holiday forecasts by over 80% in state-of-the-art TSFMs, revealing a critical blind spot in current covariate-informed model evaluation.

Robot controllers can drastically alter motion smoothness, with simple interpolation reducing peak jerk by 83% in VLA-driven manipulation without significant tracking degradation.

A unified cognitive architecture, built from just a few core principles, predicts human choices across thousands of experiments and reproduces 75% of known behavioral effects.

Discrete-time self-attention dynamics converge to a consensus equilibrium dictated by the principal eigenvector of the value matrix, a finding confirmed in real-world Transformers.

Quantum interference in a trainable Szegedy walk unlocks interpretable, high-accuracy perturbation response ranking on graphs, outperforming classical and neural methods even without explicit physical parameters.

Normalizing-flow conditionals and amortized variational inference unlock scalable, non-Gaussian spatial field modeling previously hindered by intractability.

Most feature selection benchmarks are misleading; tree-based methods and F-Test emerge as surprisingly robust and efficient choices on real-world tabular data.

Randomized LMO methods still face a $1/T^2$ convergence rate bottleneck for strongly convex optimization, even with adaptive oracles.