Weekly Recap

Achieving classification accuracy comparable to digital models, this novel XL MIMO system leverages wave-domain learning to drastically reduce hardware complexity.

Data argue against a reliable standalone prognostic value of PD-L1 in the early postoperative period while the limited observation time does not permit conclusions on late recurrence or long-term survival.

LLMs falter in maintaining coherent multi-turn interactions, with only 37.3% successfully completing five-turn episodes, revealing a major gap in generative UI capabilities.

CE3D++ enables flexible, dialogue-driven 3D and 4D scene editing by leveraging a large language model to interpret user intentions and coordinate visual tools seamlessly.

Data argue against a reliable standalone prognostic value of PD-L1 in the early postoperative period while the limited observation time does not permit conclusions on late recurrence or long-term survival.

LLMs falter in maintaining coherent multi-turn interactions, with only 37.3% successfully completing five-turn episodes, revealing a major gap in generative UI capabilities.

CE3D++ enables flexible, dialogue-driven 3D and 4D scene editing by leveraging a large language model to interpret user intentions and coordinate visual tools seamlessly.

Static sample selection in reinforcement fine-tuning is a recipe for suboptimal updates—DIEM adapts dynamically, leading to superior performance on reasoning tasks.

Generation-first training eliminates latent collapse, achieving unprecedented stability and performance in latent generative modeling.

Models trained on CoVA-SFT outperform existing multimodal reasoning baselines by over 2x, revealing the potential of structured visual-textual integration in AI.

Models show a striking 40% drop in verbalized commitment when cues come from tool returns instead of user messages, challenging assumptions about reasoning trace reliability.

Achieving a 40% reduction in code redundancy while maintaining functionality could revolutionize how organizations manage their software portfolios.

SafeAtlas-VL reveals that nuanced, multi-level safety assessments can significantly enhance the performance of multimodal safety models, outperforming existing benchmarks by 4%.

The venous thromboembolism (VTE) rate among patients who received chemotherapy for initial endometrial cancer treatment is described, and factors associated with VTE diagnosis in this population are identified.

A cooperative multi-agent orchestration framework, in which each edge node is equipped with a scheduling agent to route tasks to local ASPs or neighboring edge nodes, and an attention-guided centralized critic to estimate per-agent values from cross-agent states under GPU memory heterogeneity is proposed.

Key predictors of mortality included delayed antifungal therapy, delayed catheter removal for source control, Echinocandin resistance, and septic shock requiring ICU admission, and resistance rates were notable.

Frontline healthcare workers experienced considerable emotional and psychological challenges during the COVID-19 pandemic, driven by occupational stressors, fear of infection, stigma, and inadequate institutional mental health support, highlighting the importance of integrating mental health protection into emergency preparedness and healthcare workforce management.

The study innovatively introduced a physically constrained scenario scaling factor based on historical correlations and future climate mean values, thereby addressing the bias where machine learning models’ predicted NDVI means converged across different SSP scenarios.

This review summarizes the major DDR pathways, their roles in tumor evolution and immune remodeling, and the rationale and limitations of combining DDR-targeted therapies with immunotherapy and discusses biomarker refinement, resistance mechanisms, and future strategies for translating genomic stress into durable antitumor responses.

This invited review introduces the Global South Radiology AI Implementation Framework (GSRAIF), a literature-derived synthesis designed to guide the safe, equitable, and sustainable deployment of AI in medical imaging.

This review paper presents an in-depth analysis of dimensionality reduction techniques used in image processing applications, including Correlation-Based Feature Selection, Forward Feature Selection, Linear Discriminant Analysis, Principal Component Analysis (PCA), and Empirical Mode Decomposition (EMD).

To systematically evaluate the clinical applications of AI in implant dentistry, with particular emphasis on diagnostic accuracy, treatment planning, prognosis, external validation, generalizability, and clinical applicability, a systematic review will be conducted according to PRISMA 2020.

Key insights are identified into the barriers and facilitators of ABG implementation in LMICs to strengthen ABG use in LMIC pediatric oncology programs and guide refinements in terminology, question sequencing, and scaling for contextual relevance.

Pediatric cancer patients with COVID-19, particularly those with mild or asymptomatic infection, can continue chemotherapy with favorable outcomes, and these findings support the individualized continuation of cancer therapy during active infection.

Prenatal maternal depression and anxiety emerged as the most frequently studied mechanisms, showing some evidence of independent effects irrespective of postnatal factors, but most of the findings point towards a 'chain-of-risk' model, where prenatal mental health interacts with the postnatal environment.

This review addresses the current gap by critically examining recent advances in biologic therapies and evaluating emerging invasive and non-invasive delivery strategies to enhance their therapeutic efficacy in brain cancer.

Overall, group-based interventions improved health, mental well-being, and biopsychosocial outcomes in CALD populations with chronic conditions, however, evidence remains limited.

There is no other digital intervention available for school nurses in their work with mental ill health that includes interactions with the student and other features to support mental health, and a digital intervention in the form of a mobile app with various features was considered a helpful way forward.

Disentangling overlapping, semi-transparent objects no longer requires rigid shape priors: recombining latent instance queries directly resolves ambiguous visual evidence, yielding a +2.7 AJI jump on ISBI2014.

RLVR chokes reasoning diversity at the front door rather than during execution: an 11x–16x likelihood collapse occurs before the very first operation, leaving downstream solution paths intact and fully recoverable via targeted late-layer weight interpolation.

Models already know when a question is fundamentally unanswerable—they hallucinate anyway because their internal "impossibility" representation is completely orthogonal to their safety-refusal circuitry.

This paper proposes World Model RL (WMRL), which replaces environment execution with a world model to remove this bottleneck and accelerates training by 3-4x on various tasks at different agent scales, while exceeding the performance of standard RL baselines.

Self-evolving LLM agents can introduce irreversible changes, but EvoUndo reveals that careful design of recovery mechanisms can recover from 99.3% of these failures.

Local decision failures, not just knowledge gaps, are the key to improving small-model dialogue agents, leading to a dramatic clemscore increase of over 200%.

Sliding Window Attention outperforms Linear Attention by 2 to 10 times on long-context reasoning tasks while being faster and requiring less memory.

AI systems can now create and revise deliverables in ways that demand new standards for accountability and iterative improvement.

LoopArena reveals that even the best models struggle with long-horizon task guidance, achieving only a 24.69% success rate in complex coding scenarios.

Achieving a staggering 79x reduction in decoding latency, Parallel Tube Decoding redefines efficiency in spatio-temporal video grounding.

LayerRecall boosts long-range video generation consistency by intelligently routing historical context to specific model layers, outperforming traditional methods without sacrificing local detail.

Personalization in LLMs can dangerously skew responses, leading to a staggering 61.7% increase in sycophantic bias.

In real-world practice, GPS provides regimen-independent prognostic stratification, whereas PLR shows treatment-dependent prognostic relevance, supporting its potential utility for biomarker-guided and individualized treatment selection in ESCC.

A physics-constrained conditional generative adversarial network that bypasses iterative constrained inversion by directly learning a forward generative mapping conditioned on Pauli expectation values is introduced.

SWE-bench leaderboards mask a massive real-world distribution shift: casual, under-specified user prompts drop agent resolution rates by 6.4 percentage points and reveal that reproduction steps provide zero measurable utility compared to explicit user intent.

Conversational avatars no longer require latency-heavy cascades: generating speech and full-body gesture end-to-end from shared LLM hidden states runs 5.4× faster than two-pass pipelines while matching cascaded teacher fidelity.

Across diverse tasks, models, and datasets, HyQuant maintains nearly lossless accuracy with an extremely simple design, demonstrating the efficiency and practical feasibility of hybrid quantization for LLM attention.

Live self-improvement can boost agent performance by over 14 percentage points while reducing output token usage by nearly 43%.

EditaLive achieves real-time character video editing with state-of-the-art performance, preserving facial expressions while eliminating latency issues in live streaming.

A clear threshold reveals that fitting ellipsoids can succeed or fail dramatically based on a single statistical moment, reshaping our understanding of high-dimensional data fitting.

Randomized tree ensembles can reduce forecasting errors in wind energy optimization by over 75% compared to traditional models, revealing a critical advantage in structured data scenarios.

Current block drafting models are operating at only 71% acceptance efficiency, leaving a staggering 43-64% of rejection unexplained by their design.

TADP achieves an impressive 80.91% mAP on the KITTI dataset, setting a new benchmark for single-stage 3D object detection.

MLLMs can recognize urban scenes but fail to maintain reliable navigation and goal-directed behavior over extended exploration in complex environments.

Self-OPD achieves state-of-the-art performance in flow matching models by eliminating the need for a separate teacher, transforming self-exploration into effective supervision.

Editing long videos with multiple instructions can be done without hallucinations or loss of temporal continuity, thanks to a novel agentic framework that combines LLMs and VLMs.

Task-specific image editing can boost MLLM performance by over 29%, but not all tasks benefit equally—some are left behind.

TTPO achieves label-supervised performance without any ground-truth labels, outperforming traditional methods on key benchmarks.

Evolved skills can enable smaller models to outperform their larger counterparts, highlighting the transformative potential of systematic knowledge accumulation in AI agents.

Current video generation models fall short of capturing the full distribution of possible behaviors, revealing a critical gap in probabilistic alignment that needs to be addressed.

Generating agentic data isn't just about quantity; it's about crafting valid, informative experiences that evolve alongside agents and their environments.

ES not only outperforms GRPO in reasoning tasks but also reveals that substantial parameter updates don't equate to widespread functional changes in LLMs.

Magpie revolutionizes game development by enabling real-time visual rendering without compromising gameplay design or requiring complete asset production.

LLMs reveal a complex geometric structure of moral knowledge that captures the tension in moral dilemmas, rather than offering simplistic moral resolutions.

Transition realization is a game-changer for World Action Models, with LEON boosting performance and robustness beyond traditional methods.

Unsupervised speech models can be dramatically improved for accented speech, achieving a 23.6% boost in performance with minimal adaptation.

Evolving interaction graphs can obscure peer effects, but DynaNet-DR reveals how to disentangle these complexities for more accurate causal estimates.

The Gaussian kernel's smoothness can lead to catastrophic overconfidence in predictive uncertainty, making it a risky default choice for regression tasks.

Non-survivors scored significantly higher on a novel sepsis severity index, revealing critical insights into patient trajectories that traditional models miss.

Removing the least important attention heads from transformers can preserve performance in tabular data tasks, challenging conventional wisdom about model architecture optimization.

Common geodesics fail to ensure Fisher consistency, revealing that even optimal score vectors can yield strictly non-Bayes maximizers.

Multi-hash embeddings can shrink GNN ID tables by over 98%, revolutionizing friend recommendation systems at scale.

LLMs falter in corporate Q&A tasks, with performance dropping sharply as document complexity increases.

LAMA not only boosts platform revenue but also maintains user response quality, challenging the status quo of traditional advertising models.

Regularizing model selection with a coding-theoretic approach reveals a new pathway to consistent performance even under complex dependencies and uncertainties.

The leading last-iterate fluctuation in quantile temporal-difference learning can be controlled without polynomial dependence on the number of quantiles, challenging conventional assumptions about sample complexity.

Composing distinct polynomials can lead to linear independence, fundamentally reshaping our understanding of neural network identifiability.

SecureDrive-FL withstands Man-in-the-Middle attacks without sacrificing accuracy, achieving a remarkable balance of privacy and performance in federated learning.

Achieving over 16 seconds of I/O-compute overlap, FoldPipe redefines efficiency in training molecular ML models on constrained resources.

QuantumBoostNet outperforms traditional models in cardiac ultrasound view identification, showcasing the potential of hybrid classical-quantum approaches in medical imaging.

HALO achieves a 13.7 percentage point boost in zero-shot open-set accuracy with 10x fewer parameters than leading models, redefining efficiency in human activity recognition.

Machine learning models for nitrogen recommendations in winter wheat fail to outperform standard advice, but a simple correction can cut profit losses by 43%.

LLM agents excel in time-series analysis when given domain context, but their performance drops significantly when required to generate code for predictions.

LLMs can now generate algorithms for operations research problems that rival traditional methods, even with minimal human input.

A single destructive image can now yield comprehensive population-level statistics on crack degradation in lithium-ion cathodes, transforming how we assess battery aging.

Achieving classification accuracy comparable to digital models, this novel XL MIMO system leverages wave-domain learning to drastically reduce hardware complexity.

A Hybrid ensemble of XGBoost and TabNet yields a striking annualized return of 51.26%, highlighting the critical role of regime-robust hyperparameter tuning in algorithmic trading.

Critical adjacency from expert critics can significantly enhance music recommendation systems, achieving a notable AUC of 0.767 in cold-start scenarios.