Machine learning news — deep learning, reinforcement learning, neural architecture search, diffusion models, and new ML frameworks and libraries.
A paper submitted to arXiv on 13 Oct 2026 finds that forcing syntax-specialized attention heads in transformer language models to attend to a distractor noun instead of the subject still leaves the model's syntactic pred…
MonoMIP, a dual-decoder framework for monocular 3D object detection, reached 25.84% Moderate AP3D|R40 on KITTI val and 21.66% on the official KITTI test set for Car without extra training data, according to the arXiv pap…
A new arXiv paper (2610.19007v1) introduces Action Jets, a network that estimates the local shape of the action-conditioned data manifold as a third-order Taylor polynomial, letting a planner score sixteen candidate acti…
A new arXiv survey (arXiv:2610.19108v1) examines data-centric approaches to fine-tuning pretrained Multimodal Large Language Models (MLLMs) for spatial reasoning, covering methods from supervised to reinforcement learnin…
LidarVerse, a unified LiDAR world model presented in arXiv paper 2610.19201v1, achieves state-of-the-art performance on most reported metrics across all three evaluation protocols while running real-time inference at 11 …
Researchers introduced SonoWorld, a causal ultrasound world model that learns video dynamics from roughly 6.9 million unlabeled multi-organ ultrasound frames and grounds them to 6-DoF probe control using a small EM-track…
MMKG-R1, a reinforcement learning framework that trains a multimodal large language model to construct multimodal knowledge graphs using quality-aware rewards, improves downstream question-answering accuracy across multi…
A study comparing YOLO11, RT-DETR, and YOLO26 detector families with ByteTrack and BoT-SORT trackers on MOT17, MOT20, and SOMPT22 found that the RT-DETR family achieved the highest mean HOTA tracking-quality scores acros…
A study comparing ten representative self-distillation methods against GRPO in agentic reinforcement learning found no method consistently outperforming GRPO across settings, according to the arXiv paper 2610.19789v1. Th…
A new arXiv paper (arXiv:2610.19915v1) proposes Importance-Aware Criteria Weighting (ICW), a method that derives rubric criteria weights from pairwise importance comparisons made by an LLM instead of the uniform weightin…
Researchers introduced Selective Retrieval-Augmented Translation (SRAT), an agentic reinforcement learning framework that learns when to translate directly and when to retrieve external evidence, according to the arXiv p…
Researchers introduced PA1-OPD, a one-step prefix-augmented on-policy distillation method that adds teacher feedback after one student-sampled transition per selected corpus prefix, according to the arXiv paper 2610.1960…
Speechmatics launched Oak on October 14th, a batch medical transcription model built to follow consultations as patients and clinicians switch languages and speakers. Speechmatics said Oak ranked first in its own 11-lang…
Databricks released Predictive AI Functions in Beta, adding four new SQL-native functions — ai_top_drivers, ai_detect_anomalies, ai_predict_value, and ai_predict_class — alongside improved accuracy for ai_forecast, which…
Leash Bio released a preprint, blog post, and open dataset for Minos, an AI playground that turns the company's DNA-encoded library data into tens of thousands of "lab-in-the-loop" simulations for scientific reinforcemen…
Insilico Medicine researchers reported in Cell and Nature Biotechnology that AI-assisted drug discovery tools identified the target and molecule behind rentosertib, reaching a nominated preclinical candidate in 18 months…
A study by MIT researcher Neil Thompson, PhD student Gabriel Manso, and research scientist Claudio Mayrink Verdun found that the best AI models went from correctly solving about 2 percent of theoretical problems in proba…
A team of engineers at the Micro Air Vehicle Lab at Delft University of Technology won a high-profile Abu Dhabi first-person-view (FPV) drone race in November 2025 with an AI-piloted quadcopter, the first time an AI-cont…
FineEnvs founder Adithya S Kolavi released two reinforcement-learning environments for OCR and speech recognition on October 14th, containing more than 12 million tasks across 22 OCR languages and 102 speech languages, a…
Researchers primarily from Nobel laureate David Baker's lab at the University of Washington used AI-based protein design tools to build new proteases by taking the catalytic core of two protease types — zinc proteases an…