VLADriveBench is a new framework to evaluate whether chain-of-thought (CoT) reasoning in vision-language-action (VLA) models is relevant, consistent, and causally connected to driving trajectories. It combines observational metrics (mentioning, hallucination, contradiction, action alignment) with a CoT intervention protocol. Applied to three models across two architectures, it reveals that observational and causal analyses can diverge sharply: ORION scores high on observational alignment but its CoT is epiphenomenal, while Alpamayo v1.5 scores lower yet its CoT is strongly causal, with visual salience gating the extent of CoT influence.

SalArt-VQA is a diagnostic benchmark for evaluating fine-grained understanding of artifacts in AI-generated images by vision-language models (VLMs). It includes 950 images and 3,681 multiple-choice questions covering presence detection, semantic localization, spatial grounding, and evidence-grounded defect identification. Testing 20 VLMs revealed that even the best model, with 99.37% detection recall, answered all four artifact questions correctly on only 53.26% of images, highlighting a sensitivity-calibration tradeoff.

This paper proposes Efficient Continual Alignment (ECA) for incremental learning in open-ended image-to-text generation. By introducing continual alignment and three core mechanisms (Mixture of Query, Fisher Dynamic Expansion, Dictionary Replay), ECA mitigates catastrophic forgetting without accessing old data, achieving superior performance on new benchmarks.

Proposes a novel framework called Context-Centric Feature Fusion (CCFF) to handle co-occurring object detection in autonomous driving, using Local Context Fusion Module (LCFM) and Global Context Attention Module (GCAM). Achieves Category-level Consistency Strategy (CCS) of 0.973 and 0.969 on Cityscapes and BDD100K, respectively, with a 14.1% improvement in small object detection AP_S and successful recovery of rare classes like 'Train'. The framework processes images in real-time with only 0.2 FPS overhead.

Existing slot-based video object-centric learning methods suffer from slot swapping due to encoding appearance and identity in a single slot vector. Dual-State Slot Attention (DSSA) separates these aspects into a local state (appearance) and an identity state, updated via a recurrent transition and competition-modulated aggregation. DSSA improves segmentation quality and temporal consistency on MOVi-C, MOVi-D, and YouTube-VIS.

Medical LVLMs are prone to factual inconsistencies and poor visual grounding. Existing alignment methods have three key limitations in the medical domain: sequence-level rewards treat clinically critical tokens equally, reliance on static SFT references causes off-policy shift, and alignment lacks visual grounding constraints. The proposed method uses a bidirectional token-wise KL regularizer and a visual-contrastive grounding objective, forming a fine-grained on-policy alignment framework that constructs preference pairs by minimally editing model outputs. Experiments validate its effectiveness.

Few-step diffusion distillation has become increasingly mature for 4-8-step generation, yet pushing further to 2 steps remains challenging. This paper introduces Z-Image Turbo++, a high-quality 2-step image generation model distilled from the 8-step Z-Image Turbo teacher using three key design choices: Distribution-Aligned Adversarial Learning, Step-Decoupled Parameterization, and End-to-End Training with Iterative Regularization. These designs substantially narrow the quality gap between 2-step and 8-step generation.

A 3D-aware framework for hairstyle transfer between images that handles large pose and scale differences using a Bald Converter and a 3D-Aware Transfer Pipeline.

This paper presents a portable, low-power, battery-operated vision-based fall prediction and detection system using human pose estimation on an AMD Kria K26 SOM. The system uses an Intel RealSense D455 camera and a three-stage pipeline (quantized YOLOX, A2J, and CNN) to achieve real-time, privacy-preserving fall detection on the edge. Results show 4.5 FPS throughput with 75.85% classification accuracy.

This paper presents a multi-agent simulation explaining the emergence and persistence of morphological alternations like "go/went". Alternate forms arise from phonological changes or lexical variants and spread through population dynamics. To evaluate realism, the authors introduce the AI Historical Linguist, an LLM-driven system that simulates debates between linguists, comparing real and simulated morphologies. Results indicate scale-free networks and random Bernoulli adoption produce more plausible patterns. Three case studies model attested historical changes.

AfriSUD is the first large-scale collection of syntactically annotated treebanks for nine diverse African languages using the SUD framework. Evaluations reveal a significant syntax gap across models, highlighting limitations in capturing African language syntax.

This study analyzes observable patterns in latent reasoning models (LRMs) and finds that patterns like BFS frontiers and decodable arithmetic also appear in controls and do not always causally affect behavior. Causal interventions reveal graded utilization of latent thoughts, and geometric analysis shows effects concentrate in low-rank directions. The authors conclude that observable patterns alone cannot establish internal reasoning mechanisms, and LRM interpretability requires matched controls and causal tests.

A new study proposes MentalMARBERT, a domain-adapted version of MARBERT, for detecting mental health disorders from Arabic social media text. Using a two-phase framework with adaptive pre-training and hierarchical fine-tuning, the model achieves state-of-the-art performance with 0.861 macro-F1 and 0.877 accuracy on a novel dataset of 50,670 tweets across six categories.

The Shopping Reasoning Bench is a new benchmark created by retail domain experts, consisting of 525 missions (232 single-turn, 293 multi-turn) and 10,863 importance-weighted binary rubrics. It evaluates multi-turn reasoning capabilities such as preference refinement, trade-off analysis, and compatibility assessment in conversational shopping assistants. Evaluations of top models (GPT, Claude, Gemini) show overall pass rates of only 57-77%, with significant degradation in multi-turn tasks, highlighting a gap in expert-level advice.

This paper frames the transformation of abstract Persian proverbs into morally faithful stories as a constrained semantic decompression task. It introduces the Proverb Aligned Narrative Dataset (PAND) and a hybrid evaluation framework. Findings reveal a decompression gap: LLMs achieve fluency but fail to instantiate underlying moral structures. Explicit reasoning and iterative refinement partially mitigate this.

This paper introduces a reproducible labeling and evaluation protocol for mechanism-level drug-drug interaction (DDI) prediction, featuring a 7-family/147-subtype taxonomy and leakage-safe cold-split strategies. It also presents MARD-7B, a model trained with three innovations: single-token KL divergence, PRM-weighted DPO, and a mechanism-aware retrieval channel. On the April 2026 DrugBank release, MARD-7B is the only system among 32 that maintains accuracy under drug-pair novelty, outperforming the best baseline by +13.9 pp and GPT-4o by +6.7 pp at ~1% of frontier API cost. Anti-memorization analysis suggests gains stem from structured pharmacological reasoning.

EDEN (Emergency Department Electronic Notes) is a new large-scale corpus of approximately 4 million fully anonymized clinical notes from Italian hospital emergency departments. A subset of about 6,000 notes has been manually annotated by clinical experts with 132 items relevant to dyspnea and loss of consciousness. The dataset aims to fill a gap in Italian clinical data to support the development of large language models in medical applications.

A common hypothesis in LLM modular design is that adapter interference arises from linear parameter overlap. This study tests it using DoRA-RBAC, finding that geometry-aware merging offers no consistent advantage over standard averaging, and orthogonality is a weak predictor, suggesting interference stems from shared nonlinear representations.

This study diagnoses the loss landscape by backpropagating through the PDE structure directly, finding that optimization failure arises from flat plateaus and sharp cliffs. When the neural network is fixed, the residual loss yields a smooth landscape, avoiding pathology, while the neural network only serves to complete observed data.

This paper introduces a novel random-feature construction for Bernstein-Schur kernels, which are products of a finite-feature kernel and a completely monotone shift-invariant kernel. The proposed method combines sketched modulation with radial randomization, achieving linear feature dimension while providing rigorous theoretical guarantees including unbiasedness and operator-norm bounds. The approach is shown to improve efficiency in kernel ridge regression tasks, with a flagship instance being the biased yat-kernel.

MF-Net is a recurrent dynamical model that represents all variables in a shared field state and updates this state through a learned relation law. It achieves competitive short- and medium-horizon forecasting across known-law interaction systems, chaotic benchmarks, real neural recordings, and ecological time series while retaining inspectable structural readout. On the 40-dimensional Lorenz-96 testbed, it achieves an eight-step R² of 0.798±0.018 and recovers local coupling support with a local/nonlocal strength ratio of 19.80±1.00 and Precision@K of 1.000±0.000.

This perspective argues that generative AI for physically constrained domains, such as semiconductor manufacturing, must embed physics into model architecture from the start rather than relying on post-hoc filtering. It surveys architectural approaches and proposes a research agenda centered on physics-fidelity benchmarks and differentiable simulators.

ProHiFlo is a novel hierarchical flow matching framework for de novo protein generation that employs coarse-to-fine generation, functional guidance via pretrained predictors, and an adaptive SE(3)-equivariant architecture. It achieves state-of-the-art performance with 4× fewer sampling steps and a 58.9% success rate on enzyme active site scaffolding, outperforming RFDiffusion (41.2%).

Evaluating statistical significance of data mining results typically requires thousands of resampled datasets, which is impractical for large-scale data. This paper introduces FewRS, a novel resampling approach that needs only an extremely small number of resampled datasets by deriving a new bound on the supremum deviation of test statistics. FewRS provides rigorous guarantees on false discoveries and achieves up to two orders of magnitude speedup on pattern mining and network analysis tasks while maintaining high statistical power.

Activation steering can shift LLM behaviour, but standard evaluations don't test whether sycophancy reduction also suppresses factual agreement. The authors introduce dual-stance evaluation, finding that while sycophantic and factual agreement are in distinct subspaces, the steering direction projects equally onto both, reducing both. This reveals a gap: readable representations may not be writable.

This paper studies restless bandits with binary latent states and imperfect binary feedback, motivated by opportunistic spectrum access with sensing errors. The authors develop a partial conservation laws (PCL)-based analytical and computational framework for establishing indexability and computing the Whittle index. Using deterministic skeleton, renewal decompositions, and combinatorics on words, they obtain tractable expressions in several threshold regimes, fully verifying PCL-indexability. For the remaining regime, efficient numerical schemes are derived for computing the marginal productivity index. Experiments show that the MP index policy typically outperforms standard benchmarks.

This paper presents a deployment-centered evaluation of an LLM system integrated into electronic health records at an academic medical center. By training a pre-response classifier that uses query content and deployment-specific context (e.g., provider type, department, language model), the model predicts the risk of user rejection with an AUROC of 0.719 over 4.5 months of prospective analysis. The findings demonstrate the feasibility of predicting user rejection using deployment context, enabling targeted guardrails and abstention strategies.

A new preprint explores the transition from human-level Artificial General Intelligence (AGI) to Artificial General Superintelligence (ASI), outlining four potential pathways: scaling AGI, AI paradigm shifts, recursive improvement, and ASI emerging from large-scale multi-agent collectives. The paper also discusses frictions, bottlenecks, and the possibility that AI progress may accelerate, leading to a series of transformative changes rather than a single breakthrough.

Compact language models face challenges beyond isolated function calling when using tools. Evoflux uses evolutionary search at inference time to repair executable tool workflows, raising execution feasibility from 3% to 17-24% on MCP-Bench tasks, outperforming SFT and DPO baselines.

TrajGenAgent proposes a hierarchical LLM agent framework for generating realistic synthetic human mobility trajectories without model fine-tuning. It uses a two-stage orchestrator-worker design: an LLM first synthesizes individual- and weekday-conditioned activity chains via in-context learning, then a deterministic workflow grounds each activity into a complete visit using personalized POI retrieval, distance-aware location selection, kinematics-aware travel-time propagation, and LLM-based duration estimation. An anomaly-detection-based evaluation framework assesses behavioral and semantic plausibility. Experiments show improvements in spatiotemporal fidelity, semantic coherence, and individual-specific behavioral realism over existing methods.