Machine learning news — deep learning, reinforcement learning, neural architecture search, diffusion models, and new ML frameworks and libraries.
Researchers propose ExFold, a training-free expert-folding framework that accelerates both prefill and decode phases of Mixture-of-Experts (MoE) inference by projecting excluded expert contributions onto retained experts…
Researchers propose PACE (Prioritized Adaptive Coverage of Evidence), a training-free framework that combines evidence frontloading with pressure-adaptive budgeting to relieve bottlenecks in Retrieval-Augmented Generatio…
Researchers propose REP-LIE, a pruning method that estimates weight importance from LoRA low-rank matrix gradients during finetuning, avoiding full gradient computation and prior finetuning. Tests on LLaMA-7B and Mistral…
Researchers at an undisclosed institution report that targeting 32 attention heads in the LLaVA-1.5-7B vision-language model reduces object hallucination in image captions, lowering CHAIRs from 0.370 to 0.230 and CHAIRi …
Researchers introduced an AI-enhanced RF interference suppression method that adds a Finite Scalar Quantization (FSQ) tokenizer layer to an autoregressive transformer-based model, achieving low latency and improved inter…
PointRL, a verifiable reinforcement learning framework introduced in an arXiv paper (2608.25299v1), improves point-level vision-language grounding by converting bounding boxes, masks, and instance labels into pointing in…
Researchers investigating a multimodal graph neural network (GNN) model that combines self-supervised pretraining with alternating optimization of a language model and GNN found that the combined approach failed to suffi…
Researchers introduced SAMpLE, an open-source SystemC-AMS framework that integrates machine learning models into virtual prototyping as first-class Timed Dataflow components via a standardized interface. The framework su…
A new study on arXiv (2608.25976v1) finds that automatic speech recognition models trained on one language and then switched to a second retain functional traces of the first language in their lowest, pre-phonemic layers…
Researchers introduced physics-informed spectrum-aware pruning (PI-SAP), a method that assigns parameter saliency based on PDE residual sensitivity, to create sparse physics-informed neural network (PINN) solvers for non…
Researchers introduced ClueWeaver, a reward-guided dual-agent framework that improves question answering on long literary narratives using compact local language models. The framework, detailed in arXiv:2608.25531v1, use…
Researchers introduced YOLOEZ, an open-source, GUI-based tool that enables no-code end-to-end YOLO model application for automated structural defect detection, integrating data labeling, training, and inference in a sing…
Researchers introduced a lightweight multimodal vision-language framework adapting TinyCLIP for early-stage apple fruitlet classification, achieving F1-scores of 0.95 for calyx, 0.98 for fruitlet, and 0.85 for peduncle o…
Researchers introduced R3, a post-training recipe that turns off-the-shelf vision-language models into robotic reasoners by mid-training on expert reasoning traces and improving with single-step rubric-based reinforcemen…
Researchers propose AutoVerifier, a residual-guided non-parametric optimization method that learns verifier inductive biases from recurring errors to improve reference-based answer verification. On four verifier benchmar…
A new arXiv preprint (2608.25166v1) reports that the grammatical role of tokens, marked by part-of-speech tags, shapes the local geometry of Transformer representations, with closed-class items expanding earlier and coll…
Researchers propose Golden-GRPO Injection (GRIN), a three-stage mixed-policy reinforcement learning framework for continual knowledge injection in large language models, which outperforms supervised fine-tuning on harder…
A new fusion strategy called Local-Geo and Spatial Context Fusion (LGSCF) improves landslide susceptibility mapping accuracy, achieving F1-scores up to 87.09% and AUC values up to 0.9472, according to a study on arXiv. T…
MoganAI released MoganBert-TR, a 149M-parameter Turkish encoder foundation model trained from scratch on 237.3B tokens with a two-stage CLM-to-MLM curriculum, achieving 78.41 on TrGLUE, the best among compared Turkish Mo…
Researchers introduced WAVE, a method for guided depth super-resolution that uses a multi-level discrete wavelet transform to reverse the traditional fine-to-coarse guidance hierarchy, enabling coarse-to-fine reconstruct…