Machine Learning News

A new study from arXiv reveals that large reasoning models fine-tuned via reinforcement learning (RL) outperform supervised fine-tuned (SFT) models on mathematical reasoning due to more linearly separable and hierarchica…

Pangram Labs released Pangram 4, a deep-learning AI-text classification model achieving an AUROC of 0.9916 with a false positive rate of 0.0041% and a false negative rate of 0.3396%. The model shows improved accuracy ove…

A new game-theoretic framework for reinforcement learning fine-tuning, proposed in arXiv:2607.26358, treats the trade-off between reward and policy drift as a sequential game where a monitor tests for deviations from a r…

Researchers propose Zero-Fi, a contrastive signal-language alignment framework that enables zero-shot Wi-Fi-based human activity recognition by aligning Wi-Fi signal features with natural-language activity descriptions i…

Researchers propose AtmosERC, a graph-based framework for Emotion Recognition in Conversation (ERC) that models dialogue-level affective atmosphere to improve emotion prediction. The framework uses a relation-aware graph…

Voice Memory, a new inference-only scheme for agentic speech recognition, reduces weighted word error rate from 8.36% to 7.52% across ten HyPoradise domains without regressing any dataset below its 1-best baseline, accor…

A new diagnostic benchmark, EC-Reason-Bench, reveals that general large language models (LLMs) score near zero on complete enzyme classification (EC) number prediction, with accuracy dropping sharply from coarse to fine-…

Researchers introduced CMT-RAG, a complementary memory framework for multi-turn multi-hop retrieval-augmented generation that aligns conversational memory with retrieval by representing dialogue context as sub-question-l…

Researchers introduce WildShadowRemover, a framework that adapts a pretrained video diffusion model via LoRA fine-tuning for robust video shadow removal in unconstrained real-world scenarios. The method augments the froz…

A new study presents a federated physics-informed neural network (PINN) for privacy-preserving brain tumor biomechanical modeling, achieving 91.4% accuracy against 90.0% for a non-federated baseline. The method combines …

A study using CT Foundation embeddings achieved a Receiver Operating Characteristic Area Under the Curve (AUC) of 0.791 for predicting distant metastasis risk in head and neck cancer, outperforming radiomics (AUC 0.772) …

Researchers at University College London propose Eddeep, a deep-learning framework for fast eddy-current distortion correction in diffusion MRI. The method achieves correction quality comparable to FSL Eddy while substan…

Researchers introduce TraceCLIP, a training-free framework that recovers latent patch-level semantic evidence from CLIP's CLS attention output, achieving gains of 1.3 to 4.5 points in average mIoU over prior training-fre…

A new arXiv study evaluating local LLM machine translation finds that dedicated MT systems from OPUS-MT and NLLB-200 remain stronger overall than three instruction-tuned LLMs (llama3.2:3b, mistral:latest, qwen2.5:14b) on…

FloDR, an invertible dimensionality reduction method based on a normalising flow, embeds high-dimensional data into two dimensions while retaining all coordinates to enable exact inverse and density calculations. Unlike …

Researchers have developed a deep neural network approach to estimate weight, height, and Body Mass Index (BMI) from a single human image captured in the wild, achieving best results with full-body images. The team creat…

Rad-JEPA 3D, a joint-embedding predictive framework for 3D CT scans introduced by researchers, achieves state-of-the-art results on closed-ended VQA and the best average spatial-reasoning score on the Spatial-Med benchma…

A systematic analysis of language sensitivity in codec-based self-supervised learning (SSL) models shows that downstream performance is insensitive to the neural audio codec (NAC) training language but strongly dependent…

A hybrid deep learning method developed by researchers uses Mask R-CNN and a color-based algorithm to quantify exposed skin from images for dermal exposure assessment. Testing on 170 indoor-painting images, the approach …

Five emerging developer tools are transforming AI and cloud workflows, including NovaCode, an open-source AI coding assistant from ex-Google engineers; EthicalGuard, a bias detection SDK; DeployML, a unified ML deploymen…