Five leading publishers and a best-selling author filed a class-action lawsuit against Meta and founder Mark Zuckerberg on Tuesday, alleging the company violated copyright law by training its generative AI platform on millions of illegally pirated books and articles.

Goldman Sachs reports reveal that the AI investment boom is driven by fear of missing out rather than rational returns, with massive spending yielding little benefit, and Nvidia as the primary beneficiary.

Blink is an AI search engine that provides sources with its results.

As AI inference becomes the primary workload, traditional TCO metrics like per-dollar FLOPS are misleading. NVIDIA argues that per-token cost is the critical measure of AI infrastructure economics, demonstrating that its Blackwell platform delivers up to 35x lower cost per million tokens and 50x more tokens per watt compared to Hopper.

This paper proposes a tactile-augmented two-stage method combining Imitation Learning and Reinforcement Learning for precision insertion tasks. First stage learns reaching policy, second stage executes insertion with failure recovery. Tactile group sampling and tactile critic enhance performance. Under 0.05 mm clearance, 67% success rate, 60% force reduction, 44% torque reduction.

ReflectDrive-2 is a masked discrete diffusion planner for autonomous driving that represents plans as discrete trajectory tokens and generates them through parallel masked decoding. It introduces AutoEdit for in-place trajectory revision and uses a two-stage training process: structure-aware perturbation supervised learning followed by reinforcement learning fine-tuning. Full-rollout RL significantly boosts drafting-editing coupling, achieving 91.0 PDMS on NAVSIM with camera-only input and 31.8 ms latency.

Researchers propose an active contact sensing approach for robust robot-to-human handover. The robot applies information-gathering motions and senses human-applied forces to infer contact state, using a Bayesian linear model to distinguish firm grasp from incidental touch. Experiments with 12 participants and 30 objects achieved 97.5% success rate, over 30% higher than baselines.

A new multi-phase whole-body MPC approach for bipedal walking combines detailed models near-term with simplified models long-term to reduce complexity while maintaining performance, validated on the HyPer-2 robot in simulation.

Researchers have developed a Bayesian framework to validate the practical equivalence of synthetic pre-crash scenarios for assessing the safety of driving automation systems. The method uses binning-based statistics to compare datasets, overcoming limitations of conventional significance testing. A case study on rear-end crashes demonstrates its effectiveness.

HDFlow is a novel hierarchical planning framework that combines diffusion and rectified flow models. A high-level diffusion planner generates strategic subgoals in latent space, while a low-level rectified flow planner produces smooth trajectories using ODEs. It outperforms state-of-the-art methods on four furniture assembly tasks and generalizes to other long-horizon benchmarks. The paper is accepted as ICML 2026 Spotlight.

This paper proposes a tightly-coupled estimation and guidance architecture for minimum-fuel low-thrust rendezvous, where navigation confidence modulates the homotopy parameter. Using a Kalman filter with Multiple Tuning Factors (MTF), the method reduces terminal miss distance from hundreds of meters to sub-meter levels under severe measurement degradation.

Researchers propose a control method for laparoscope-holding robots that unifies multimodal intraoperative signals (position, force/torque, images) into an equivalent-wrench representation. Using a task-priority scheme, the method achieves compliant manipulation and autonomous instrument tracking while maintaining RCM constraints. Experiments on surgical phantoms and in vivo porcine trials demonstrate effectiveness.

This paper presents a conditional latent flow matching approach for scalable and realistic safety-critical scenario generation. By performing distribution matching, the method transforms nominal driving scenes into safety-critical rollouts. Combining simulation and real-world data enables the framework to efficiently generate diverse scenarios. Experiments show the approach consistently produces novel safety-critical scenarios, making it a valuable tool for training and benchmarking autonomous vehicles.

arXiv:2605.04327v1 Announce Type: new Abstract: We propose an architecture for integrating high-level, human-provided safety rules and operator-aligned semantic preferences into autonomous robot navigation in unstructured outdoor environments. In our approach, natural-language rules are translated into Signal Temporal Logic (STL) specifications that guide planning and navigation during runtime. Persistent, environment-centric rules and terrain preferences are grounded into a 2D cost map, while temporally dynamic requirements are expressed as STL specifications to be monitored during runtime. We hypothesize the use of Vision-Language Models (VLMs) for zero-shot scene understanding, enabling mapping between human instructions, semantic features, and environmental constraints. Within this fr…

To address the limitations of existing multimodal LLMs in multi-step reasoning across multiple subplots, researchers propose HierVA, a hierarchical visual agent framework that iteratively constructs and updates a working context in a joint image-text space. Experiments on the CharXiv reasoning subset show consistent improvements over strong baselines.

This paper proposes adaptive confidence thresholding to improve multi-task classification in autonomous driving, and introduces the IUST-XAI-AD dataset of 958 annotated images for cross-cultural explainability evaluation.

A novel image dataset of approximately 34,700 high-resolution images captures the complete degradation lifecycle of synthetic fibre ropes under controlled cyclic fatigue loading. The dataset supports various machine learning tasks for condition monitoring and prognostics.

Researchers propose Physics-Guided Regime Unmixing (PGRU), which estimates a pixel-wise scalar from observable physical features to activate nonlinear mixing only where justified, overcoming limitations of the Linear Mixing Model under multiple scattering. Experiments on Samson, Jasper Ridge, and Urban show consistent improvements with physical coherence > 0.90.

This paper proposes a probabilistic deep learning framework for densifying and forecasting Sentinel-2 time series by generating optical images at arbitrary past or future dates. It leverages multimodal data from Sentinel-2 optical and Sentinel-1 SAR, focusing on uncertainty quantification. Experimental results show effective performance on sparse and temporally misaligned data.

Implicit neural representations (INRs) have emerged as a powerful paradigm for medical imaging via physics-informed unsupervised learning. Classical INRs optimize an entire network from scratch for each subject, leading to inefficient training and suboptimal imaging quality. Recent initialization-based approaches attempt to inject population priors into pre-trained networks, yet they rely on high-quality images and often suffer from catastrophic forgetting during fine-tuning. We present DisINR, a novel INR framework that explicitly disentangles shared and subject-specific representations. DisINR introduces a shared encoder-decoder pair and subject-specific encoders, whose features are jointly decoded for image reconstruction. By integrating differentiable forward models, it pre-trains the…

The naive application of deep learning to scientific imaging (e.g., infrared) leads to critical failures due to a mismatch between data priors and DL's simplicity bias, causing models to collapse to one-dimensional predictions and underperform even compared to RGB-based models.

We propose a topology-constrained quantized nnUNet framework for efficient and anatomically accurate 3D tooth segmentation, addressing the challenges of spatial distortion introduced by quantization in deep learning models. The proposed method integrates a novel tooth-specific topological loss into quantization-aware training, preserving critical anatomical structures such as tooth count, adjacency relationships, and cavity integrity while maintaining computational efficiency.

This paper proposes MuCALD-SplitFed, a multi-task Split Federated Learning framework integrating causal representation learning and latent diffusion to address task inconsistency and privacy leakage in medical image segmentation. Experiments show consistent segmentation improvement, reduced information leakage, and outperformance over state-of-the-art methods.

A study evaluating five multimodal LLMs on real-world dermatology data found that benchmark performance substantially overestimates clinical capability. In a hospital consultation cohort of 5,811 cases and 46,405 images, top-3 diagnostic accuracy for open-weight models was only 1.5%-13.35% (images only), rising to 28.75% with clinical context. GPT-4.1 achieved 24.65% (images only) and 38.93% with context. Models were highly sensitive to incomplete or erroneous context. Severity triage sensitivity was over 60% but insufficiently reliable for clinical deployment.

A new framework enables small language models to self-generate, verify, and refine prompts for extracting clinical entities from dental notes. Using 1,200 annotated notes, Qwen2.5-14B-Instruct achieved micro/macro F1 scores of 0.864/0.837 after direct preference optimization, demonstrating feasibility for local deployment.

A new benchmark, Nsanku, evaluates 19 LLMs on zero-shot translation across 43 Ghanaian languages. Gemini-2.5-flash leads with an average score of 26.88, but no model achieves both high performance and high consistency, indicating current LLMs are not yet reliable for large-scale Ghanaian language translation.

A study investigates how vocabulary overlaps affect knowledge transfer in multilingual neural machine translation, finding that domain match and language relatedness are more critical than joint vocabulary.

arXiv:2605.04180v1 presents MedFabric, a dataset generated via a data-centric pipeline to create realistic word-level fabrications in medical LLM outputs, and EtHER, a detector that improves factual alignment using Text2Table decomposition, word masking/filling, and hybrid sentence pair evaluation. The framework achieves over 15% improvement over state-of-the-art detectors on word-level fabrication benchmarks.

This study evaluates six large language models on conflict-event classification in Nigeria and Cameroon, finding that open-weight models exhibit significant False Illegitimation bias, while domain-adapted models are directionally neutral but still show actor-based selection bias. The authors conclude that current models are not ready for unsupervised deployment and call for fairness-aware fine-tuning, adversarial robustness evaluation, and context-specific human-in-the-loop oversight.

A new study investigates hallucinations in four popular large language models for academic writing, introducing a novel Hallucination Index (HI) metric. The findings reveal that hallucination behavior depends not only on model architecture but also on task type and prompting conditions.