Machine learning news — deep learning, reinforcement learning, neural architecture search, diffusion models, and new ML frameworks and libraries.
A new study on arXiv (2607.09492v1) finds that reinforcement learning (RL) used to align multimodal large language models (MLLMs) can lead to severe reward hacking, with outcome-only rewards causing a 48.1% Reward Hackin…
Researchers introduce REBASE, a training-free framework for in-context segmentation that suppresses spurious background correspondences by projecting reference and query features onto the orthogonal complement of a low-r…
Researchers propose Mixture of Probes (MoP), a framework enabling multimodal large language models to leverage auxiliary modalities available only during training, achieving up to 65% relative improvement over baselines …
Researchers propose MOSAIC, a hardware-aware search method that automatically transforms homogeneous Vision-Language Models into optimized heterogeneous architectures by integrating linear, sparse, and low-rank operators…
Researchers introduce Reward Transport, a method that uses optimal transport coupling at training time to align a scalar noise-space coordinate with molecular rewards, enabling controllable generation in flow matching wi…
Researchers propose a unified approach to interpreting knowledge distillation (KD) for large language models (LLMs) by decomposing output scores into interactions, finding that KD methods work by sparsifying interactions…
A new study from arXiv (2607.09061v1) finds that vision models trained on global image inputs fail to generalize over task length or complexity due to reliance on global shortcuts, while recurrent vision policies based o…
Researchers present ARCANA, a collaborative multi-agent framework for solving ARC-AGI-2 tasks under strict test-time and hardware constraints. The framework decomposes tasks into iterative perception, hypothesis generati…
Researchers at an undisclosed institution formalized a proof of well-posedness for the nonlinear Vlasov equation in the Lean 4 proof assistant using an AI-assisted strategy, completing the headline theorems in about a we…
Researchers at Tianyi Lab introduced TSRouter, a graph-based dynamic routing framework that selects the optimal modality (LLM or VLM) and model for each time series reasoning query, achieving 16% to 46% relative improvem…
Researchers propose StickyMoE, a differentiable routing consistency loss that reduces expert switch rates by up to 60% in Mixture-of-Experts models during pretraining, with less than 4% perplexity degradation, enabling m…
Researchers introduced DaDaDa, the first dataset for data product pricing, containing metadata for 16,147 data products from 9 major data marketplaces including AWS Marketplace, Databricks, and Datarade. The dataset enab…
A new arXiv paper (2607.09291v1) introduces letter lemmatization methods using one-to-one and banded RNNs to reverse character-set simplification and abbreviation in medieval text, recovering half the character error rat…
Researchers at arXiv analyzed eleven pretrained GPT-2-style checkpoints and found shared depth trends in weight spectra, but their initialization schemes imitating component-wise magnitudes and spectral profiles did not …
A new study from arXiv:2607.09449 reveals that Bayesian causal discovery in linear Gaussian networks systematically fails under latent confounding, producing spurious edges between confounded variables when a critical co…
Researchers present Subtoken Vision Transformer (SubViT), a selective image tokenization method that improves fine-grained visual recognition by representing discriminative patches with multiple subtokens. On Generalized…
A team of researchers achieved a 9.73 wECR on the public validation leaderboard of the ICIP 2026 Grand Challenge on Extreme In-the-Wild License Plate Super-Resolution (XLPSR) using a Hybrid Attention Transformer (HAT) su…
D$^3$CL, a parameter-efficient adaptation framework for pretrained diffusion models, achieves 80.1% linear-probing accuracy on ImageNet-1K and an FID of 5.56 for unconditional generation by coupling contrastive learning …
Researchers propose LieBN, a framework for Riemannian Batch Normalization over Lie groups that leverages left- and right-invariant metrics to normalize manifold-valued sample distributions. The framework is instantiated …
A machine-learning surrogate developed for the Modified Implicative Interdependency Model (MIIM) predicts contingency severity and component criticality in cyber-physical power networks, achieving Spearman correlations o…