Machine learning news — deep learning, reinforcement learning, neural architecture search, diffusion models, and new ML frameworks and libraries.
Researchers introduced PRISMS, a closed-loop framework that detects and steers tool-use failures in agentic LLMs using sparse MLP neuron readouts, achieving ROC-AUC 0.90-1.00 for over-calling and missing detection and 0.…
Researchers propose AdaMTP, an adaptive training paradigm for multi-token prediction that dynamically aligns prediction horizons with sequence predictability, outperforming standard MTP across mathematical reasoning, cod…
Researchers introduced an executable benchmark and a budget-aware meta-router for agentic workflows, achieving 100% success on held-out test tasks versus 93.5% for static workflows, with 43% lower cost. The benchmark inc…
Researchers propose a multiplicative policy shaping framework to steer pretrained deep reinforcement learning policies toward welfare-based fairness objectives at inference time without updating base policy parameters. T…
Researchers introduced a permutation-equivariant neural operator for collision-aware trajectory planning of spacecraft swarms, trained on ten spacecraft and generalizing zero-shot to swarms of 1,000 amid more than 11,000…
Researchers introduced ELECTRIC (Evidential Learning-Enhanced CT Reconstruction via Iterative Correction), a physics-guided Bayesian method that uses an evidential neural network to iteratively reconstruct CT images. In …
Researchers introduced Semantic-Aware Kernel Entropy (SAKE), a training-free guidance method for text diffusion models that computes order-2 Rényi entropy over a kernel Gram matrix to balance fidelity and diversity. SAKE…
A new study from arXiv (arXiv:2608.00135v1) evaluating Convolutional Neural Networks on the JONES-19 cultural design dataset finds that learning from scratch with repeated local sampling (multi-crop) can match the discri…
Researchers introduced a physics-chemistry-informed neural network (PCINN) that predicts surface coverage in spatial atomic layer deposition (SALD) in about 7 milliseconds, roughly 50,000 times faster than CFD, with a te…
A study from arXiv (arXiv:2608.00220v1) shows that on-policy reinforcement learning with verifiable rewards (RLVR) can improve the current objective while making successful behaviors for later objectives too rare to samp…
Researchers introduced Counterfactual Modality Attribution (CMA), the first framework for quantifying modality-level contributions in multimodal large language models (MLLMs), which identifies the decision-driving modali…
A new arXiv paper (2608.00296v1) proposes Stabilized Best-of-K training for neural combinatorial optimization, modifying the Leader Reward method to use a stabilized rank signal indexed by sampling budget K. In tests on …
Researchers introduced agentic Bayesian optimization, a paradigm where an LLM agent acts as the central decision maker in the BO loop while a Bayesian backend provides uncertainty-aware optimization. They instantiated th…
A new arXiv preprint (2608.00346v1) reports that ensemble methods combining unsupervised deep clustering algorithms outperform individual methods on imbalanced tabular data, achieving higher accuracy, normalized mutual i…
Researchers introduced AgentMemBench, a benchmark evaluating five long-term memory strategies for conversational AI agents across three datasets and 491 annotated question turns, finding that external key-value store (EK…
Researchers propose a unified LLMOps architecture for real-time, enterprise-ready deployments of large language models, addressing knowledge staleness, catastrophic forgetting, hallucination, and weak feedback loops. The…
A study from arXiv (2608.00507v1) training a ~7M-parameter Transformer encoder on child-directed and read speech found that the learning objective, not architecture, determines perceptual narrowing in self-supervised spe…
Researchers propose the Capability-Driven Multimodal Scaling Law, the first cross-family framework that predicts vision-language model (VLM) benchmark accuracy from textual capability scores, trained on over 150 VLMs acr…
Researchers trained a small language model to route queries to specialized retrieval agents, achieving an NDCG@10 of 0.918 on mismatched queries, compared to 0.539 for Amazon Nova Lite and 0.490 for Claude Haiku 4.5, and…
A new arXiv preprint (2608.00585v1) shows that per-chunk verification fails on multi-hop retrieval-augmented generation, with entailment scoring reaching only 0.643, 0.523, and 0.560 AUC on HotpotQA, 2WikiMultihopQA, and…