Weekly Recap

Diverse LLMs, when strategically routed and collaborating, can collectively achieve significantly fairer outputs than any single model acting alone.

A focused narrative review of studies of treatment-exposed solid tumours, distinguishing paired human tissue measurements from unpaired associations, functional perturbations, post-withdrawal experiments and clinical studies identified one that combined a stroma-directed intervention, CAF-resolved pharmacodynamics and a controlled treatment-effect estimate.

Artificial intelligence should be regarded as a complementary decision-support tool whose practical contribution depends on combining computational methods with biological knowledge and real-world agricultural observations.

The Biomod2 ensemble model was employed to compare the differences in species distribution predictions derived from two distinct data sources, to evaluate the respective influence of each environmental variable, and to characterize the future shifts in suitable habitats of Anoplophora glabripennis across climatic zones in China under climate change scenarios.

A focused narrative review of studies of treatment-exposed solid tumours, distinguishing paired human tissue measurements from unpaired associations, functional perturbations, post-withdrawal experiments and clinical studies identified one that combined a stroma-directed intervention, CAF-resolved pharmacodynamics and a controlled treatment-effect estimate.

Artificial intelligence should be regarded as a complementary decision-support tool whose practical contribution depends on combining computational methods with biological knowledge and real-world agricultural observations.

The Biomod2 ensemble model was employed to compare the differences in species distribution predictions derived from two distinct data sources, to evaluate the respective influence of each environmental variable, and to characterize the future shifts in suitable habitats of Anoplophora glabripennis across climatic zones in China under climate change scenarios.

Higher income and older age were associated with a stronger sense of community, better perceived community safety, and improved health, and policymakers should consider the impact of community factors on well-being during health crises.

While the study was closed prematurely prior to completion, the sequence of regorafenib prior to anti-EGFR therapy suggested a possible survival benefit.

This study explored community perspectives on the influence of climate change on anxiety, trauma, depression, emotional wellbeing, and psychosocial development among young children in Malawi.

Unsupervisedly discovering genetic subgroups and performing layered meta-analysis unlocks 74 previously hidden association loci in a heterogeneous pig cohort and substantially reduces genomic inflation in the UK Biobank.

Brain control isn't a hierarchy, but a dynamic spectrum of influence shaped by context and practice, challenging decades of neuroscience dogma.

Activity-efficient RNNs exhibit a surprising dissociation between high-variance neural activity and causal output, and noise robustness actively hinders activity efficiency.

Directed transport can drive traveling wave solutions in viral spread models, even outperforming diffusion.

Model-generated DNA can now be traced with perfect accuracy, even after single-base edits, using a watermark detectable without knowing its precise location or strand.

Three generations of computational approaches to plant GRN inference are reviewed and measure them against a five-level ladder of evidence, including statistical association, conditional dependence, physical binding, functional regulation and causal circuits.

This study proposes an Artificial Intelligence (AI)-based model, namely Integrating Feature Reduction Models for Heart Disease Prediction (IFRM-HDP), which combines a hybrid feature selection strategy incorporating Boruta and minimum Redundancy Maximum Relevance with a Temporal Convolutional Network to capture complex dependencies among clinical attributes and improve classification performance.

Single-photon cameras transition from ambiguous to near-deterministic scene recovery as photon budget increases, a phenomenon precisely modeled by a novel Hypergeometric generative framework.

This mini-review highlights recent advances in AI-driven LIB recycling, focusing on the integration of machine learning, deep learning, and computer vision across the battery recycling value chain.

A cognitive formulation of breastfeeding trauma is presented, drawing from an existing formulation model for moral injury in PTSD, using cognitive therapy to support cognitive behavioural therapists to conceptualise breastfeeding trauma through the lens of moral injury where appropriate.

A training-light approach boosts fisheye object detection by 80% by strategically addressing input distortion, feature misalignment, and decision-level fusion with minimal parameter updates.

World models need to forget selectively, discarding outdated instance facts while rigidly retaining fundamental invariants like physics, a nuance standard forgetting metrics miss.

Output-layer gradients alone are sufficient for effective post-training data selection, slashing feature-extraction costs by up to 10x.

Success conditioning, a staple in RL, is formally proven to converge to optimal policies across broad MDP classes, with explicit convergence rates established.

Discovered options from high-dimensional observations unlock state-of-the-art Atari performance, drastically accelerating learning in complex, exploration-heavy games.

Most GNN explainers fail to detect indirect influences or alternative implementations of the same logic, even on seemingly robust benchmarks.

Learning conservative and periodic population dynamics is now 4-14x faster without needing numerical solvers.

A single AI pipeline can now detect marine environmental anomalies directly on satellites, drastically reducing data transmission needs and enabling real-time alerts.

Forgetting is not always beneficial in non-stationary bandits; evidence, not just change, should trigger adaptation.

Reward hacking is mitigated and sample quality preserved in few-step discrete diffusion models by replacing intractable KL penalties with a novel inverse-distillation regularization.

Reconstructing private client data from federated learning updates is significantly improved by treating model updates as path integrals, outperforming single-point gradient approximations.

Forward KL divergence in distillation inflates student entropy, while reverse KL can deflate it, offering a new lens on implicit regularization.

Predicting the predictable in latent space unlocks S2S weather forecasting skill beyond current AI model limits, matching and exceeding physics-based ensembles at critical 5-6 week horizons.

Structured variational inference unlocks richer epistemic uncertainty estimation in GPLVMs, outperforming mean-field approximations on manifold reconstruction.

Security benchmarks for LLM agents are fundamentally flawed: changing a tool's name can drastically alter reported attack success rates, invalidating cross-model comparisons.

Unlabeled physiological signals unlock multimodal EEG foundation model adaptation, matching supervised performance with zero-shot generation.

Ambiguity-aware NL2SQL generation framework produces realistic, complex queries that boost model performance on challenging real-world benchmarks, outperforming existing synthetic data.

Most "long-range" GNN benchmarks fail to actually test long-range interactions, but a new framework provides provably long-ranged tasks and a per-range error oracle.

Real-world robots can now learn faster and require drastically less human help by adaptively deciding *when* and *how* to be assisted, rather than relying on fixed rules.

Autonomous robotic navigation achieved 70% success in physically accessing cerebral veins, a critical step towards less invasive brain-computer interfaces.

Recurrently folding older time series context onto a most-recent anchor dramatically improves forecasting accuracy by naturally prioritizing recent evidence.

LLMs for math translate statements towards generality, widening quantification domains in over 60% of cases, yet fail to explicitly state implicit hypotheses, suggesting a shallow understanding of theoretical relationships.

A single 30-hour fMRI dataset unlocks whole-brain decoding of naturalistic speech, revealing continuous performance gains with more data.

Deep tabular generators fail to consistently outperform trivial baselines, even on datasets orders of magnitude larger than typical clinical data.

Diverse LLMs, when strategically routed and collaborating, can collectively achieve significantly fairer outputs than any single model acting alone.

Automated crack identification in fusion materials now scales across diverse conditions without manual tuning, enabling standardized inputs for predictive modeling.

Mechanistic edits can now offer provable guarantees of skill removal and preservation across entire input regions, not just tested subsets.

Mixture-of-Experts for order execution masks training failures rather than improving performance, with annealed exploration proving more critical for stability.

Few-step sampling in flow matching models achieves significantly higher fidelity by analytically integrating the dominant affine component of the dynamics, leaving only a smaller neural residual to be integrated.

Neural embeddings of conversation reveal individual identity, not just conversational role or trait differences, suggesting a fundamental organization of brain activity around the person, not the interaction.

Time series models can achieve state-of-the-art long-term forecasting by dynamically blending recent dynamics with retrieved historical trajectories, rather than relying solely on a single look-back window.

Deep generative priors unlock robust multistatic ultrasound recovery in rank-deficient regimes, outperforming linear methods and providing crucial uncertainty quantification.

Most visual computation in VLMs is redundant, allowing for a 55% FLOP reduction with minimal accuracy loss by simply bypassing token-specific projections in early and late layers.

Long-horizon agents achieve state-of-the-art performance by decoupling planning credit from execution guidance via structured subtask trees.

Metropolis-Hastings sampling for LLM policy composition consistently outperforms SIR, offering a provably better approximation of desired reward trade-offs with the same computational budget.

Drug-target affinity model performance rankings flip dramatically when evaluated on unseen protein targets, revealing a critical flaw in standard benchmarking.

Most graph EDL methods fail to reduce epistemic uncertainty with more data, but bootstrap ensembles show promise for consistent uncertainty estimation.

Operator-informed initialization for Fourier Feature PINNs directly combats spectral bias by tailoring weight distributions to the PDE's differential operator, leading to balanced frequency convergence and improved accuracy.

K-fold cross-validation's statistical evidence diminishes with increasing folds when data heterogeneity creates strong inter-fold dependence, a phenomenon quantified by an effective number of folds.

Pretrained generative models can directly sample exact Bayesian posteriors for data assimilation by conditioning their interpolant paths on observations, bypassing costly retraining.

LLMs struggle with character-level reasoning, with tokenization and prompting strategies profoundly affecting performance on tasks like character counting and edit distance, and complex substring extraction remaining a significant challenge.

Physical quantities like momentum and gravity are explicitly encoded and localized within the activations of video diffusion models, not just implicitly learned patterns.

Transformer NQS generalize with prediction error inversely proportional to in-context examples and depth, requiring only linear depth scaling with system size.

Unlocking causal discovery in time-series by disentangling instantaneous and lagged influences, even with nonstationary dynamics.

Agents can achieve globally optimal solutions in distributed learning with correlated data by learning a game where they share low-dimensional signals, even with non-convex objectives.

Past states are no longer fixed in data assimilation; DAWIS revises them dynamically using windowed inverse sampling and multitask interpolants.

A novel bidirectional curriculum strategy unlocks efficient exploration in sparse-reward RL, outperforming unidirectional methods and even matching demonstration-based approaches on challenging tasks.

Muon's finite-step orthogonalization provably guarantees reaching any target training loss for wide networks, resolving a key theoretical gap.

Spectral disentanglement and knowledge-guided aggregation unlock efficient, high-performance MoE neural operators for large-scale PDE pretraining, slashing expert count by 50% without sacrificing accuracy.

Jointly learning nonlinear features across multiple asymmetric data sources unlocks performance gains beyond traditional kernel methods.

Achieve 10x lower quantization errors on MCUs with 16-bit precision, matching or beating 8-bit inference speed and energy costs.

Critic-free LLM fine-tuning can now achieve state-of-the-art performance in stateful environments without costly repeated rollouts, by leveraging ordinal statistics on replay buffer samples.

Continuous-time weather forecasting with Neural SDEs achieves skillful 5-day probabilistic predictions at hourly resolution, overcoming autoregressive error accumulation.

Simple logistic regression, enhanced by feature selection, rivals deeper neural networks for cross-cohort TB classification, suggesting a practical and efficient approach for resource-limited settings.

Evolved quantum circuits can replace substantial portions of classical vision backbones, achieving comparable accuracy with orders of magnitude fewer parameters.

A low-rank tensor decomposition of a stochastic basis coupled with a learnable spatial dictionary unlocks accurate and efficient uncertainty quantification for random PDEs, outperforming existing methods by a significant margin.

Combinatorial causal discovery methods, previously limited by redundant scoring, can now scale to hundreds of variables in minutes with a novel spline-based score-evaluation technique.

GQA's KV-cache eviction overhead can be eliminated by pooling head importance with Renyi-2 entropy, achieving perfect budget preservation and needle recall where mean pooling fails.

Generative planners can now achieve real-time safety and replanning by aligning diffusion sampling with actual system dynamics, bypassing the need for full trajectory generation and costly iterative safety checks.

Resource-aware LLM task graphs can achieve competitive accuracy with 3.17x less cost and latency by jointly executing and verifying subtasks, learning LLM quality online.

A single, leakage-audited recipe reconstructs 19 public credit-default datasets, offering the first comprehensive benchmark for risk prediction with scarce, late-arriving labels.

Expert design guidance nearly doubles LLM agent speedups for GPU kernel generation, yet agents still miss critical low-level optimizations.

Hamiltonian property testing and learning with only forward evolution cannot achieve Heisenberg-limited scaling, requiring $\Omega(1/\varepsilon^2)$ query time.

Photonic quantum solvers can achieve near-optimal financial risk detection with drastically fewer features than classical methods, but their advantage is critically dependent on the predictive signal's concentration within the feature space.

Concatenation MLLMs fuse vision late, while native architectures co-adapt modalities from the start, revealing fundamentally different cross-modal information flow.

Lazy second-order optimization is far more efficient than previously thought, with a new bound achieving near-optimality.

Interference, measured by task vector variance, predicts LLM merge collapse and guides repair, revealing that sign conflict statistics are anti-predictive.

Lifelong RL agents can learn to dynamically select and weight prior policies using task similarity derived from online experience, outperforming static modular approaches and eliminating forgetting.

Geophysical inversion uncertainty is now quantifiable and reducible by coupling ensemble conditioning with physics-informed neural network residuals.

Fine-grained credit assignment in LLM agents can be unreliable due to downstream randomness, but AdaStep's optimal shrinkage coefficient precisely balances local and global rewards.

Gaussian Processes can be made

Real-time imputation for streaming time series now goes beyond prediction to include domain-aware validation and explicit retain-or-replace decisions, making data quality actionable for operators.

Generate high-quality hypergraph node embeddings in seconds, not hours, without sacrificing accuracy.

LLMs routinely overstate evidence and prematurely close investigations, but a novel dataset and RL approach enable them to match human expert judgment on evidence sufficiency.

A shared latent geometry reveals that paralinguistic speech attributes can be precisely controlled and disentangled from linguistic content within a frozen LLM.

Structured pruning can slash DNN energy and latency for space applications without sacrificing reliability, by trading redundancy for reduced SEU probability.

Adversarial attacks can completely neutralize trigger-tag mechanisms designed for misuse detection in open-weight LLMs, rendering them ineffective even with output transformations or weight modifications.

A dual-channel architecture enables LLMs to achieve perfect long-horizon retrieval and robust planning by isolating macro-objectives from noisy context, outperforming baselines by a significant margin.

Steerable conformal coverage guarantees are now achievable in RL-driven NAS, enabling precise pruning of millions of architecture evaluations without sacrificing accuracy.