Weekly Recap

Human-like robot trajectories scored an impressive 71.50 on a human-likeness scale, suggesting robots can now move with a level of fluidity that builds trust in human-robot collaboration.

VLMs struggle to provide accurate annotations in video games, revealing significant gaps in their understanding of dynamic environments.

Annotating video game datasets with VLMs can transform the way RL agents learn by simplifying reward extraction and conditioning.

A dynamic "structural firewall" can thwart model extraction attacks on GNNs without sacrificing performance for legitimate users.

A dynamic "structural firewall" can thwart model extraction attacks on GNNs without sacrificing performance for legitimate users.

NeuPAT recovers nearly all lost language capabilities in multimodal LLMs while ensuring robust performance across diverse tasks.

An integrative review of digital technology interventions in marine ecosystem monitoring is presented, synthesising peer-reviewed research on sensing platforms, molecular and acoustic methods, machine learning, and ocean data infrastructures published mainly between 2014 and 2026 to propose a four-layer sensing-to-decision architecture.

Even the most advanced LLMs struggle to maintain narrative consistency, with a staggering 68% of generated content conflicting under user interventions.

Gambit achieves up to 6.7% higher accuracy and over 2x throughput by intelligently reallocating compute resources during reasoning.

A comprehensive review of approaches for the detection of malware infecting sensor nodes in WSNs and a taxonomy of reviewed approaches for detection of malware is presented, which will be highly beneficial for researchers and practitioners designing approaches and systems for the detection of malware in WSNs.

Multi-turn interactions can be effectively optimized in LLMs using tailored RL strategies, overcoming significant challenges in credit assignment and reward design.

LSAR transforms audio retrieval by mapping audio directly into a sparse lexical space, enabling efficient and interpretable retrieval without the pitfalls of transcription.

OasisKV achieves up to 2.1x throughput gains in LLM inference while using significantly less memory, challenging the limits of current HBM constraints.

CED reveals that VLMs can be trained to prioritize evidence-based reasoning over language shortcuts, leading to more reliable visual understanding.

A unified toolbox for time-series dataset similarity could revolutionize how researchers select and evaluate datasets for AI model fine-tuning.

SymDiag reveals that existing verification methods fail to diagnose reasoning errors effectively, providing a robust framework that localizes failures and generates actionable insights.

Grounding tasks can be unified into a single correspondence prediction problem, leading to a 48% boost in performance on complex datasets.

Ego-OSCAR delivers a game-changing, budget-friendly solution for crowdsourced egocentric data collection, complete with a rich dataset and open-source tools.

Hybrid nested optimization outperforms traditional methods by effectively decoupling structural design from parameter tuning in LLM-driven evolutionary algorithms.

A self-developing agent that autonomously improves its own coding capabilities while navigating operational safety challenges has set new performance benchmarks.

CEDAR transforms demand forecasting by enabling accurate simulations of future sales outcomes based on planned actions, outperforming traditional methods that fail to account for decision-driven changes.

This comprehensive review delves into the fundamental concepts of radiomics and AI, summarises their current applications in BC imaging, and explores their evolving roles in clinical practice, highlighting recent advances and presenting case studies demonstrating the clinical impact.

The discussion will explore emerging directions such as reasoning and agentic systems, as well as hybrid approaches that integrate learning with structure, prior knowledge, or explicit models of the world that integrate learning with structure, prior knowledge, or explicit models of the world.

Self-improving coding agents can achieve unprecedented efficiency and generalizability by learning from multiple task trajectories simultaneously.

MirrorWorld achieves unprecedented fidelity in mirror reflection generation by effectively modeling both what to reflect and how to spatially arrange it.

CoinRAG redefines efficiency in RAG by achieving higher accuracy with lower operational costs through innovative cache reuse strategies.

Achieving substantial speedups in kNN regression calculations on large datasets by leveraging CUDA GPUs could transform how we approach real-time data analysis.

AutoPrune enables LLMs to autonomously design visual-token pruning strategies, achieving a remarkable 9.9x reduction in FLOPs without sacrificing performance.

AtlasVLA outperforms multi-view baselines by over 17% in long-horizon tasks, showcasing the power of proactive reasoning in embodied AI.

Cross-view correspondence, not geometry, is the critical bottleneck in achieving reliable 3D localization of rib fractures from CT projections.

Learned routers can outperform fixed-model baselines by 14.6%, revealing a new frontier in efficient LLM deployment.

Attention Collapse reveals a hidden vulnerability in RAG systems, enabling the detection of poisoning attacks that traditional methods fail to catch.

Previous contamination mitigation strategies may inflate model performance by over 40%, but a new evaluation method reveals the true extent of this overestimation.

YOLO-PEFT transforms the fine-tuning landscape for real-time detectors by replacing trial-and-error with structured, auditable planning, achieving notable performance gains.

The Skaling law slashes loss estimation errors by up to 3x and reduces compute needs by 90% for reliable model performance predictions.

SimWAM achieves state-of-the-art performance in autonomous driving while eliminating the need for future video generation at inference, drastically reducing latency.

Complex search behaviors typically requiring explicit controllers can emerge naturally from an agent's reasoning process, leading to substantial performance gains in optimization tasks.

WorldTrace redefines memory management in video world models, achieving up to 19.5% better episodic recall without any retraining.

No single model-harness combination consistently outperforms others, highlighting the critical need for tailored evaluations in agent deployment.

Small LLMs can achieve up to 27.2% accuracy gains by leveraging hierarchical memory from larger teacher agents, reshaping how we think about agent training.

Capek 0.5 reveals that organizing embodied capabilities by functional roles can significantly enhance performance in complex execution tasks.

Modular TTT reveals that the right combination of learning rate and weight decay can significantly boost performance, while deeper architectures may actually degrade results.

A multi-agent forensic reasoning framework outperforms leading closed-source models in deepfake detection by leveraging diverse analytical perspectives on forgery cues.

Artist-grounded image generation can achieve unprecedented fidelity by explicitly controlling for artistic intent, rather than relying on shortcuts that distort user vision.

TNT is increasingly considered the preferred treatment strategy, especially for patients with MRI-defined high-risk rectal cancer, and organ-preserving (“watch-and-wait”) strategies are increasingly favored and continue to be refined in ongoing clinical trials.

Only the ICI–sorafenib combination suppressed tumor growth, whereas both monotherapies showed limited antitumor activity, and findings support ICI–sorafenib combination therapy as a promising therapeutic strategy for TNBC.

It is concluded that targeting tumor immunosenescence represents a promising but incompletely validated therapeutic axis for enhancing anti-tumor immunity in breast cancer.

Expensive end-to-end omni pre-training degrades foundational reasoning and vision, yet simply routing timestamped ASR transcripts into a frozen VLM matches or beats native omni architectures across 56 benchmarks.

U-OPSD enables LLMs to self-improve without any external supervision, achieving up to 10.7% performance gains on challenging reasoning benchmarks.

SiPE not only boosts syntactic accuracy but also enhances general language understanding, setting a new standard for integrating syntax into Transformer models.

SkillZip achieves a remarkable 3.46x compression ratio while preserving 99.2% of dependencies and 98.7% of verifier reachability, revolutionizing how agent skill libraries can be managed.

Visual tool-use in multimodal LLMs may create an illusion of effectiveness, with many models achieving gains that are not causally justified.

Scaling language models with interpretability as a core constraint reveals that they can become more capable and understandable simultaneously, defying traditional trade-off assumptions.

Dynamic layer routing can boost LLM accuracy by 5% without the need for weight updates or expensive search loops.

EffectLearner achieves unprecedented video object removal quality by effectively reasoning about complex object-induced effects in dynamic scenes.

PaDoc achieves a remarkable 94.24 F1 score among end-to-end parsers while being the fastest parser at five concurrency levels, revolutionizing document parsing efficiency.

CalibForge reveals that adversarial calibration can dramatically enhance the effectiveness of training data for terminal agents, leading to unprecedented performance improvements on standard benchmarks.

Achieving 74.8% accuracy on a new temporal reasoning benchmark, ChronoVision redefines how multimodal models can tackle complex visual tasks.

Adding correctly labeled examples can paradoxically increase learning difficulty by a logarithmic factor, challenging existing assumptions about sample exchangeability.

SG-TULA achieves competitive performance in sampling from complex non-convex distributions while offering explicit convergence guarantees that traditional methods fail to provide.

An exact closed-form update for orthogonality-constrained optimization could revolutionize efficiency in matrix-aware methods like Muon.

AutoThread reduces simulation execution time by up to 83.8% while boosting throughput to 1.8x that of existing RL inference methods.

Achieving the statistically optimal risk bound in agnostic PAC learning finally clarifies the sample complexity landscape, matching lower bounds precisely.

Synthetic clinical benchmarks can be made significantly more realistic without sacrificing operational utility, challenging the assumption that utility alone guarantees quality. WHY_IT MATTERS: This work could fundamentally change how synthetic benchmarks are evaluated and optimized in healthcare AI, ensuring they are both useful and realistic.

BaKron accelerates neural network quantization by harnessing richer curvature information, achieving significant computational efficiency without sacrificing performance.

Interaction can reduce the number of required tests by a quadratic factor, but the expected exponential advantage of adaptive querying is not realized.

Surv-IPTB outperforms traditional survival analysis methods by accurately estimating treatment benefits in complex, nonlinear patient data.

Human-like robot trajectories scored an impressive 71.50 on a human-likeness scale, suggesting robots can now move with a level of fluidity that builds trust in human-robot collaboration.

Early stopping can transform gradient descent from a source of suboptimality into a strategy for achieving minimax-optimal classification performance in noisy settings.

Geometry-adaptive prompting can transform how we approach few-shot learning in dynamic graphs, leading to substantial performance gains.

Simple transformations don't universally translate across text embedding models, revealing critical compatibility issues that challenge existing assumptions in the field.

De-identified medical images may still reveal patient identities, challenging the assumption that such scans are truly anonymous.

Reinforcement learning can transform persistence diagrams into dynamic entities, balancing complexity reduction with the preservation of essential topological features.

Finite-sample guarantees reveal how localized conformal prediction can significantly reduce miscalibration while preserving coverage.

Adapting LLM inference scheduling to bursty traffic can boost throughput by leveraging real-time request intensity estimation.

Leveraging large language models to generate synthetic transitions, ProDVI boosts sample efficiency in deep reinforcement learning without the need for extensive pre-collected data or simulators.

A unified taxonomy reveals the intricate relationships among post-training adaptation techniques, illuminating how they evolve and interact across diverse AI models.

TS-RAG redefines time series forecasting by effectively merging input data with retrieved sequences, leading to unprecedented accuracy improvements.

Missing data handling in PRTrees can outperform traditional regression trees, offering a robust alternative that maintains interpretability and flexibility.

SkillTFM achieves a remarkable AUC boost of up to 0.142 without any training, redefining how we adapt tabular models in the face of distribution shifts.

Hardware confinement of cryptographic keys eliminates key exfiltration risks, achieving zero successful attacks in a rigorous evaluation against AI agent vulnerabilities.

Operational failures in multi-node fine-tuning can be mitigated by prioritizing power monitoring over utilization metrics, revealing critical insights for practitioners.

Early stopping of accumulation in binary neural networks can slash computation by over 86% with minimal accuracy trade-offs.

Token-level entropy emerges as a powerful predictor of task difficulty, revealing hidden flaws in environment design that traditional metrics overlook.

Self-PreTraining boosts transformer accuracy in medical time series by up to 6 percentage points, even with limited data.

Bridging implicit and explicit relational biases boosts skin lesion diagnosis accuracy by nearly 3% using graph-based methods.

Accurate predictions don't guarantee reliable uncertainty estimates, revealing critical gaps in current evaluation methods.

Co-optimizing network and ML parameters can accelerate training by up to 42%, unlocking new efficiencies in AI workloads.

Decoupling expert personas in LLMs can drastically improve their accuracy and appropriateness in high-stakes domains like healthcare and finance.

Super-resolution techniques can significantly erase small white matter lesions in MRI scans, challenging assumptions about their reliability in clinical settings.

TFMs, despite being the leading approach for tabular predictions, fail to consistently represent joint distributions, raising concerns about their reliability in practical applications.

VLMs struggle to provide accurate annotations in video games, revealing significant gaps in their understanding of dynamic environments.

Removing just a few companion cells can drastically alter single-cell annotation outcomes, revealing a hidden vulnerability in existing tools.

A wearable respiratory sensor can accurately distinguish stress from relaxation with 88% accuracy, making it a game-changer for non-invasive stress monitoring.

Annotating video game datasets with VLMs can transform the way RL agents learn by simplifying reward extraction and conditioning.

FALM-PINN reduces relative \(L^2\) errors by up to 100 times compared to current benchmarks in solving high-frequency nonlinear PDEs.