Machine learning news — deep learning, reinforcement learning, neural architecture search, diffusion models, and new ML frameworks and libraries.
Researchers introduced ViTAL-X, a lightweight 0.4B-parameter model that achieves state-of-the-art performance on six temporal benchmarks, outperforming 7B-parameter models and baselines trained on 600x more data. The mod…
Researchers propose DISTAL, a dual-prior framework for structure-agnostic materials property prediction that combines self-supervised compositional pretraining with knowledge distillation from a pretrained ALIGNN teacher…
Researchers propose ReNFT, a method that repairs mode collapse in reward post-training of diffusion generators by recalibrating internal probability mass, preserving 98.9% and 99.0% of NFT's reward on PickScore and GenEv…
Researchers introduced ECCBench, a benchmark evaluating vision-language models' memory beyond accuracy via efficiency, compression, and calibration, finding that pretrained VLMs compress text but not video memory and are…
Researchers introduced REAL-Q (Real-time E2E-loss Aligned LLM Quantization), a post-training quantization method that uses dynamic Block-wise Gradient Descent to reduce end-to-end KL divergence by up to ~49% relative to …
A study comparing nine foundation model variants against two state-of-the-art electricity price forecasting benchmarks in Germany, Poland, and Spain over 2021-2025 found that only TabPFN models consistently outperform th…
A new arXiv paper (2609.00090v1) introduces a Weight of Evidence (WoE) framework to evaluate feature importance methods (FIMs) in Explainable AI, quantifying how strongly evidence supports hypotheses about feature import…
A new retrieval-augmented generation framework for corrosion knowledge synthesis, demonstrated on magnesium alloy corrosion, achieved Token F1 gains of 143-194% and system faithfulness of 0.964, according to a preprint o…
Researchers introduced GeoLAMP, a geometry-aware latent autoregressive generative model for solving multiphysics partial differential equations (PDEs) in highly complex micrometer-scale tortuous geometries, as detailed i…
Researchers proposed the Evidence-Grounded Typed Knowledge Graph (EGT-KG), a retrieval framework that improves scientific question-answering with local Small Language Models (SLMs). In experiments on a Biopolymer-bound S…
Researchers propose TrustPropRAG, a retrieval-augmented generation framework that models document relations as a graph and propagates reliability signals from limited human feedback across the corpus, improving retrieval…
Researchers propose AnySearch, a reinforcement learning framework that enables a single LLM-based search agent policy to adapt to any budget constraint at deployment, outperforming baselines across seven QA benchmarks an…
A new arXiv study introduces a Holistic Sustainability Score (HSS) to compare small language models (SLMs) and quantized large language models (LLMs) for edge AI, finding that Qwen3-30B-A3B/GGUF Q4 ranks first with a sco…
A new study from arXiv (2609.00718v1) finds that structured pruning is the first stage where driving capability is lost in compressed autonomous driving policies, while distillation improves the pruned actor but is limit…
A new arXiv study identifies a multi-stage 'safety circuit' in large language models, comprising Harmful Detection Heads, Safety Neurons, and Refusal Heads, and shows that circuit-guided weight scaling improves safety ra…
Researchers proposed a framework that aligns self-supervised respiratory audio encoders with medical terminology using a medical LLM to synthesize structured reports, enabling zero-shot classification. Across 9 tasks on …
Researchers introduced OCGQuant, a post-training quantization method that pairs outlier channels with low-magnitude companions to reduce quantization error in NVFP4 low-bit inference, achieving the lowest WikiText-2 perp…
A new arXiv preprint (2609.00077v1) diagnoses a failure mode called algorithmic mode collapse in code-level autonomous research loops (ARLs), where LLM agents repeatedly propose the same algorithmic changes despite editi…
Researchers released SCAFFOLD, a large-scale structured dataset pairing computer science research figures with captions, context, question-answer pairs, and chain-of-thought reasoning traces, built from arXiv papers. The…
A new arXiv preprint (arXiv:2609.00071v1) reports that prediction error is not a reliable proxy for causal estimator quality, based on Monte Carlo simulations comparing OLS, GAMs, XGBoost, and Double Machine Learning wit…