Machine learning news — deep learning, reinforcement learning, neural architecture search, diffusion models, and new ML frameworks and libraries.
Researchers propose Dual-Form ASR (DF-ASR), a framework that extends spoken-form automatic speech recognition to semantics-aware written-form inverse text normalization for Chinese, using paired supervision and a sequenc…
Researchers introduced RL-ADA, a co-evolutionary framework using world feedback instead of human labels to train adversarial robust enterprise dialogue agents, eliminating the annotation bottleneck. In a banking proof of…
Researchers propose Hybrid Search, a targeted correction strategy for warm-initialized LLM-based ASR models that leverages hidden-state interactions between the ASR model and its base LLM to refine tokens with high seman…
A new arXiv paper (2609.03005v1) introduces the Conformal Relevance framework, which uses in-context learning example curation and ensembling to create score functions for NLP tasks like summarization and extractive ques…
Jina AI released Jina-OCR-v1, an end-to-end document parsing model combining DeepSeek-OCR's compressed-vision encoder and 3B mixture-of-experts decoder (activating ~570M parameters per token) with a FastMTP speculative d…
Researchers propose Dude, the first dual-detection multi-agent system for paper-code discrepancy detection, which improves recall and precision by up to 22.8% and increases F1 score by up to 18.7% over baseline methods. …
A new benchmark, DuplexSpeechBench-IFEval (DSB-IFEval), evaluates implicit instruction-following in full-duplex voice agents, comprising 1,038 test cases across eight assistant roles. Testing six real-time speech systems…
Researchers introduced NTEP (Necessary Tool-Evidence Path), an annotation scheme and reward mechanism to improve tool use in agentic vision-language models, and NTEP-R, a supervision mechanism that ensures each tool call…
Researchers propose FreNet, a feature reconfiguration framework with visual priors, that outperforms state-of-the-art methods on 9 medical image segmentation benchmarks across three imaging modalities, achieving Dice imp…
Researchers proposed PPO-STGNN, a DAG task-scheduling algorithm integrating proximal policy optimization with spatio-temporal graph neural networks, to address NP-hard scheduling in heterogeneous cloud-edge-end computing…
Researchers introduced CulturalMenuBench, a benchmark of 4,870 items across 10 languages and 18 regions, revealing that multimodal language models scoring above 94% on standard food recognition tasks drop to at most 56% …
Researchers introduced Exemplar, a few-shot segmenter that fuses a frozen DINOv3 backbone with classical native-resolution filter responses in a single lightweight head, achieving a mean score of 0.782 across eleven biom…
Researchers propose TRIM (Trigger Removal by Identifying Manipulated Regions), a black-box defense that detects and removes backdoor triggers in deep neural networks at inference time without requiring model internals, t…
Researchers propose CoverPruner, a training-free visual token pruner for vision-language models (VLMs) that frames pruning as Representational Coverage Maximization (RCM), ensuring each removed token is represented by a …
Researchers introduced ProgResViT, an input-adaptive Vision Transformer that progressively increases input resolution and model width across rounds, terminating early when predictions are confident. Applied to DeiT, Prog…
Researchers proposed an ensemble-based self-taught learning approach using convolutional autoencoders for parking space classification, achieving 93% to 96% accuracy on the PKLot and CNRPark benchmarks under cross-datase…
A new arXiv preprint (arXiv:2609.03261v1) finds that medical multimodal AI benchmarks overstate visual reasoning because correct answers can be derived from non-visual cues, a phenomenon the authors call 'reasoning infla…
Researchers present a unified computational approach to tensor-based morphometry that detects brain surface shape differences between clinical groups from magnetic resonance images, combining surface modeling, smoothing,…
Researchers propose Laplacian Frequency Hierarchies, a plug-and-play scheme for 3D Gaussian Splatting training that reduces active Gaussians by archiving lower-frequency fields, achieving average speedups of 1.73x and 1.…
Researchers propose P-CORE, a self-supervised method for point-based neural editing that improves robustness to large deformations by enforcing surface consistency before and after random deformations. The method, integr…