Machine learning news — deep learning, reinforcement learning, neural architecture search, diffusion models, and new ML frameworks and libraries.
A new arXiv paper introduces StepLearn, a nonparametric framework that lets large language model agents learn from individual transitions during deployment while requiring prospective validation before rules are reused a…
A team of researchers has developed an Optical Character Recognition (OCR) model that combines a deep learning model with image preprocessing techniques to extract information from Korean-language invoices, achieving an …
An independent replication of a recent uncertainty-aware graph neural network for road-level crash prediction found that 7 of 11 design decisions failed to replicate on a second London borough, with 4 of those reversing …
Researchers introduced FD-VAD, an ASR-free streaming endpointer that maps bounded causal audio windows directly to Continue/Stop decisions for full-duplex voice interaction, combining a frozen speech encoder with a light…
A lightweight module called Sieve, paired with a costly LLM dubbed Sage, improved Retrieval-Augmented Language Model accuracy by up to 69.4 percentage points and Macro-F1 by 55.2 percentage points over one-stage baseline…
Researchers introduced PrimeSeeker, a capability-oriented training framework for deep search agents that builds web-grounded anchor structures and jointly derives questions with reference evidence skeletons, per arXiv pa…
Researchers introduced Less Uniform Diffusion (LUDI), a uniform diffusion language model framework that addresses an over-uniform training objective and condition-target confusion during sampling by directing each revers…
Researchers introduced SABLE (Sparse-view Animal Behavior Latent Embeddings), a self-supervised framework that reconstructs 3D animal behavior from two-view video without ground-truth 3D labels, according to the arXiv pa…
A study of 12,000 texts from nine disasters found that text-based AI detectors cannot reliably separate human-authored from AI-generated disaster social-media posts, with AUROC between 0.402 and 0.517 across fourteen fro…
Researchers posted arXiv:2609.35965v1, a paper presenting what they describe as the first systematic formalization of multi-agent vision-and-language navigation (VLN) as a constrained coordination problem, with subtasks …
Researchers introduced Persistence Forcing (PerF), a pixel-space diffusion Transformer method that assigns different feature groups distinct refinement budgets across depth, producing what the authors call persistent and…
A new arXiv paper (2609.36101v1) reports that a single dominant direction captures 94.4-99.9% of the squared norm of the image-text mean separation in CLIP and SigLIP vision-language encoders, explaining why modifying th…
Xiaomi researchers released Xiaomi-OCR-0, a unified 0.8B-parameter vision-language model for document parsing and OCR-centric understanding, built on Qwen3.5-0.8B and trained on an approximately 170M-sample OCR-centric c…
A new arXiv paper (arXiv:2609.36219v1) introduces LeRF, a framework that trains Vision-Language Models to build and use explicit reference coordinate frames for perspective-taking reasoning, addressing the tendency of VL…
Researchers introduced MATE (Mutually Adversarial self-Training with Evolving data), a reinforcement-learning post-training framework that has the generation and understanding branches of a unified multimodal model chall…
A new arXiv paper (2609.36348v1) reports that adding cross-view class-token alignment to diffusion transformer training improves ImageNet linear-probing accuracy by 9.4% with the class token and 10.1% with mean-pooled pa…
A 35B-parameter model trained with a three-state autonomous search harness (Rubric, Answer, Verify) and a Seal Memory tool scored 72.83 on BrowseComp, outperforming comparable open-source systems, according to the arXiv …
A study posted to arXiv (2609.29769v2) found that Jev, a "decision model" that returns calibrated probabilities over permitted answers, can often substitute for three flash-tier LLM judges on nine panels from seven bench…
SemOPT, a semantic-guided framework for correcting LLM-based optimization models, achieves an average 7.6% accuracy improvement over the strongest baseline on complex datasets across seven optimization modeling benchmark…
Researchers proposed SCSP, a training-free framework that uses semantic compression prompts to select informative tokens for long-context embeddings instead of uniformly averaging all token representations, according to …