This survey explores using abundant human video data to train Vision-Language-Action (VLA) models, overcoming the high cost and embodiment coupling of robot demonstrations. It categorizes methods into four classes: latent action representations, predictive world models, explicit 2D supervision, and explicit 3D reconstruction, and highlights three key challenges: video structuring, cross-embodiment grounding, and evaluation protocols.

VLAMotor is a novel framework for enhancing Vision-Language-Action (VLA) models by exposing failures through distance-aware testing and fine-tuning with agent-synthesized data. It significantly improves failure detection and success rates in both simulated and real robotic manipulation tasks.

A novel deepfake detection method that uses semantic segmentation to focus on specific facial regions, achieving 0.905 AUC on Celeb-DF v2 without fine-tuning, outperforming LipForensics and Xception.

This paper investigates tool-use dynamics in complex visual reasoning tasks, identifying a 'tool-use collapse' phenomenon where models gradually stop using tools yet improve accuracy. The authors find that eliminating tools degrades performance, but incentivizing tool use yields only marginal gains. By adding entropy regularization to encourage rollout diversity, they achieve the best performance despite declining tool usage, suggesting tools should be viewed as scaffolding during training.

A new zero-shot framework called HSGM bridges the semantic-geometric gap in Vision-Language Navigation by transforming 3D information into a hierarchical map compatible with VLMs, achieving state-of-the-art performance on R2R-CE and RxR-CE benchmarks.

Researchers propose MIND (Data Manifold-aware Image Diffusion Model), which explicitly models manifold geometry by integrating discrete patch tokenization into the score function of a continuous diffusion model. On ImageNet 256×256, MIND-B with 130M parameters achieves FID 2.06 under guidance, surpassing LlamaGen-3B's performance.

Researchers propose using cellular sheaves to improve interpretability of weakly-supervised pathology localization. By aligning sheaf disagreement with classifier attention, they achieve higher localization accuracy and transferability.

The study proposes EviOSAHS, a framework that decomposes facial images into seven anatomical queries combined with clinical data for high-sensitivity screening of obstructive sleep apnea-hypopnea syndrome. On a 642-subject cohort, it achieved 88.47% accuracy, 94.86% sensitivity, 93.74% F1-score, and a 5.14% false-negative rate.

Researchers introduce Planktonzilla-17M, the largest and most comprehensive plankton image dataset to date, with 17.4 million images from 13 imaging systems spanning 602 taxonomic classes. They find that supervised classifiers using taxonomic lineage as text match or exceed CLIP-style training, while biological foundation models like BioCLIP perform poorly on plankton identification.

This proof-of-concept study presents a task-aware flow-based generative framework that reformulates the conventional Flow Matching training paradigm to optimize subsampling in compressed sensing. It achieves state-of-the-art performance in image reconstruction (PSNR 25.17 dB on CelebA at 5% subsampling, and 29.24 dB on fastMRI for 8x accelerated MRI) with minimal computational overhead, offering a flexible approach for designing data- and task-driven sensing schemes.

This paper proposes Belief2-Attention, an extension of Belief-Attention that utilizes both the perpendicular and projected components from orthogonal projection. The projected component is processed via an activation function and linear mapping, functioning as a two-layer FFN. Additionally, an extra inner-product matrix ZZ^T is added to QK^T to capture richer token correlations. Experiments on image classification and segmentation demonstrate improved performance.

DefocusTrackerAI introduces a deep-learning framework for detecting defocused particle images in various optical configurations. It compares Faster R-CNN and YOLOv9, finding YOLOv9 superior in recall and uncertainty. The framework works for astigmatic and non-astigmatic images in different lighting conditions and has been validated on real DPT experiments, including fluorescence and shadowgraph data. Pre-trained models are available for public use.

This study introduces a multi-domain red teaming framework to evaluate medical LLMs under adversarial and ethically complex scenarios. Testing 11 models on 690 clinical scenarios, mean scores ranged from 0.791 to 0.984, but high-performing models exhibited complete failures in safety-critical cases. Equity tasks showed 10-20% error amplification with demographic modifications. The findings highlight that performance variance and worst-case failures are more clinically meaningful than average accuracy, and hybrid evaluation with clinician oversight is essential for credible safety assessment.

A new study combines cognitive-linguistic features with DistilBERT embeddings using Holographic Reduced Representation (HRR) to detect depression in online text, achieving a macro F1 score of 0.94, significantly outperforming the traditional TF-IDF baseline of 0.80.

Long-context decoding in LLMs is constrained by memory bandwidth for fetching KV cache. Existing methods prune keys before decoding, ignoring joint key-value dependence. ART is a lightweight run-time mechanism that tracks accumulated attention outputs and terminates KV block accesses when contributions become negligible. It is orthogonal to existing methods and achieves 20% higher throughput on LongBench with comparable accuracy.

A new benchmark, TrustLDM, evaluates Language Diffusion Models across safety, privacy, and fairness. Results show trustworthiness degrades with malicious contexts, but longer contexts don't always worsen effects. The work also introduces TrustLDM-Auto for automatic vulnerability identification.

A system using a generate-many-select-best strategy with preference modeling from human pairwise judgments achieves top ranks in SemEval-2026 humor generation task across English, Chinese, and Spanish subtasks.

Proposes SENSE, which uses target model hidden states for semantic retrieval and soft-gated evaluation to improve robustness and efficiency of retrieval-based speculative decoding, achieving up to 4.09 mean acceptance length and 3.26x speedup on LLaMA and Qwen.

CSRP is a three-stage framework for Chinese grammatical error correction that addresses over-correction via continual pre-training, chain-of-thought SFT, and RL with efficiency-aware rewards. Achieves SOTA on NACGEC (F0.5=50.99, precision=57.17) and CSCD spelling correction (F1=59.61, surpassing GPT-4 by 5.20).

Detecting AI-generated text is becoming increasingly challenging as modern language models approach human-level fluency. AEyeDE proposes an attribution-driven approach that leverages model attention as a discriminative signal. Using a proxy Transformer model to extract attention-based attribution matrices and a lightweight CNN to learn representations, the method outperforms text-only baselines in encoder-decoder translation settings, performs strongly in decoder-only settings, and shows robustness under cross-dataset transfer and spelling perturbations. The findings suggest attention-based attribution maps provide a complementary and interpretable signal for AI-generated text detection.

A new framework DOPA improves LLM robustness in out-of-distribution scenarios by using an OOD proxy to retrieve diverse demonstrations from source domain when target domain is inaccessible. Accepted at ACL 2026.

Researchers have constructed DraDDP, the first publicly available English multimodal dataset for multi-party dialogue discourse parsing, based on American TV dramas. It includes 495 dialogue segments, 6,374 utterances, and 9.1 hours of parallel video. Experiments show multimodal information helps capture dialogue structures and relation types. The dataset, guidelines, and code will be released.

This survey proposes a proactive lifecycle-based detection framework to counter adversarial synthetic content accelerated by Generative AI. By integrating machine learning and social science through the C5 Interaction Model, it reviews emerging inauthentic narrative detection techniques including Coordinated Inauthentic Behavior analysis, epidemiological modeling, and Hawkes processes. It discusses anomaly detection in high-dimensional embedding spaces, unsupervised coordination detection on multi-layer graphs, and agentic AI systems. Finally, it addresses challenges of tracking rapidly evolving threats and multi-level distributional drift, outlining a future research agenda.

This paper systematically analyzes tool-calling in LLM agents along two axes: effectiveness (measurement) and efficiency (learning). It reveals that evaluation results are highly sensitive to minor implementation choices like random seed and system prompt, making leaderboard rankings unreliable, especially in multi-turn settings. On efficiency, it identifies computational waste in standard RL training and proposes two techniques that achieve substantial speedup without performance loss.

This survey presents a unified taxonomy for world models across four dimensions: architecture, methodology, reasoning, and applications. It traces milestones from PlaNet to Genie, discusses convergence of chain-of-thought reasoning with world models, and identifies key challenges and future directions.

FoLoRA introduces a forgetting-aware optimization framework that uses a generalized Rayleigh quotient to balance task utility and forgetting penalty, and constructs proxy calibration data by sampling from the pretrained model, achieving superior preservation-adaptation balance on math, code, and instruction following tasks.

This paper introduces Automatically Differentiable Nonlinear Tensor Networks (ADNTNs), a family of structured weight generators trained end-to-end via reverse-mode automatic differentiation. Extending low-rank adaptation and tensor factorization, ADNTNs build large weight tensors through hierarchies of small cores, nonlinear activations, and lateral mixing tensors. Three architectures are studied: Tree Tensor Networks (TTNs), augmented TTNs (aTTNs) with boundary disentanglers, and Multi-scale Entanglement Renormalisation Ansatze (MERA). Experiments on AlexNet and VGG-16 layers achieve per-layer compression ratios from 2000x to 77000x with accuracy matching or exceeding dense baselines. The results are encouraging but require joint optimization of contraction schedules and deployment kernels.

Researchers find a shared valence axis (V-axis) between modern LLMs and human EEG, built from just 9 sentences and consistent across 14 LLMs. It maps linearly to EEG features from 123 subjects. While 36 EEG classifiers spontaneously rediscover this direction, additional supervision fails to improve decoding and often degrades performance, formalized as the saturation regularity. Ensembling across residual diversity boosts accuracy by 10.5% on FACED.

Demo2Reward is a test-time adaptation technique that optimizes prompts for Vision-Language Model (VLM) reward models using a few expert demonstrations (3-10 trajectories), reducing false positives while preserving true positives, without additional training or computational resources.

Hoeffding Concept Bottleneck Models (HCBM) leverage the Hoeffding functional decomposition of gradient-boosted trees for non-linear and sparse aggregation of concept scores, improving explainability, and outperform linear CBM in classification and object detection tasks.