Machine learning news — deep learning, reinforcement learning, neural architecture search, diffusion models, and new ML frameworks and libraries.
Researchers have identified a critical underfitting problem in diffusion bridge models near the target endpoint, caused by a mismatch in noise levels between network input and regression target. The team proposes the Noi…
Researchers at Qwen2.5-3B-Instruct found that supervised fine-tuning (SFT) causes greater disruption to internal computational circuits and more catastrophic forgetting than reinforcement learning (RL) when adapting larg…
Researchers have developed SalsaAgent, a multimodal embodied language model that generates expressive, full-body salsa dance motions in response to a human leader and synchronized to music. The system extends a large lan…
Researchers developed a target-informed self-supervised pretraining and model-ensemble strategy to improve cross-device generalization of ultrasound AI models without requiring new annotations. The method, tested on pedi…
Researchers have introduced Embodied3DBench, a benchmark designed to evaluate low-level spatial intelligence in Vision Language Models (VLMs) within embodied 3D environments. The benchmark includes over 21,000 question-a…
Researchers have developed GeRaF, the first neural implicit learning method for reconstructing 3D geometry from radio frequency signals, enabling see-through-occlusion sensing without relying on RGB or LiDAR. The system …
Researchers have developed GAP3D, a diffusion-based method that aligns vision-language model latents directly to patch-level image embeddings, enabling frozen generative models to use VLMs as prompt encoders without expe…
Researchers have developed GenesisFunc, an automated pipeline that uses a multi-agent framework to generate high-quality, diverse training data for large language model function-calling. An 8B parameter model fine-tuned …
Researchers have introduced Micro-Macro Retrieval (M2R), a new framework designed to reduce hallucination in large language models during long-form text generation. The system addresses the problem of factual errors by e…
A new study systematically examined the effect of transformer-based language model size on topic quality, finding that model size had a negligible impact on coherence and divergence metrics. Researchers tested seven mode…
Researchers have introduced Thoughts-as-Planning, a framework that formalizes reasoning chain optimization in large language models as a sequential decision-making process over a latent semantic space. The method models …
Researchers have developed Deep Visual Coding, a frequency-domain representation inspired by 1990s image codes that creates psychovisual-style abstractions for deep learning. The approach enables the first psychovisual-b…
A new 306M-parameter language model architecture, the Cognitive Categorical Transformer (CCT), achieved 21.27 validation perplexity on WikiText-103, a 12% relative improvement over a fine-tuned GPT-2 Small baseline. The …
Researchers have introduced CosmicFish-HRM, a compact language model that uses a Hierarchical Reasoning Module (HRM) to dynamically allocate computational effort based on input complexity rather than applying fixed compu…
Researchers developed a self-play reinforcement learning framework for the four-player imperfect-information card game Big 2, enabling controlled comparisons of different RL agents. Under standardized conditions, Proxima…
A team of researchers has developed a training-free multi-signal segmentation pipeline for the CVPR 2026 UG2+ Challenge's Dynamic Object Segmentation in Turbulence (DOST) track. The system combines RAFT motion estimation…
Researchers have developed PrismFlow, a new flow matching method that uses Koopman-inspired dynamical experts to generate high-quality time-series data. The approach addresses spectral distortion and poor mode coverage i…
A new paired-minimum detectable effect (MDE) budget for 4-bit quantization benchmarks, derived from classical sample-size calculations, provides benchmark designers a one-line formula to pre-register reliability claims b…
Researchers have developed a policy-neutral execution and measurement layer to address reliability and interpretability issues in event-driven industrial scheduling policies. The framework constructs decision-valid snaps…
Researchers have developed Differentiable Belief-based Opponent Shaping (D-BOS), a first-order method for multi-agent reinforcement learning that treats an observer's belief as the shaped opponent state and differentiate…