Machine learning news — deep learning, reinforcement learning, neural architecture search, diffusion models, and new ML frameworks and libraries.
A new arXiv paper (2610.10918v1) evaluating large language models for machine translation quality annotation found that LLM agreement with human annotators exceeds inter-human agreement on some tasks but remains unreliab…
Researchers introduced STI-OPD, a stochastic teacher intervention framework for multi-turn agentic on-policy distillation (OPD) that replaces a student model's proposed action with a teacher-generated one based on teache…
Activation probes for monitoring deployed language models need sample counts set by coverage rather than per-concept difficulty, according to arXiv paper 2610.10594v1. Tracing learning curves over 10-590 synthetic sample…
A router-augmented membership inference attack that combines output-side signals with aggregated MoE routing features increases true-positive rate at 1% false-positive rate by 2.7 to 9.4 percentage points across all nine…
SPERA, an EEG foundation model built on a joint-embedding predictive architecture (JEPA) that predicts in latent space, achieved the highest average balanced accuracy across nine downstream clinical, cognitive, and BCI t…
A new arXiv paper (2610.10613v1) presents a privacy-preserving anomaly detection framework for in-vehicle Controller Area Network (CAN) buses, combining a Temporal Transformer CAN Encoder with federated lightweight heads…
Researchers introduced LoopOPD, a cross-loop on-policy distillation framework that uses a looped language model's own frozen terminal loop as a compute-privileged teacher for an intermediate loop student, plus Dynamic Lo…
Researchers introduced the Invariant-Conditioned Isotropic Kernel Neural Operator (IKNO), a compact graph model that predicts three-dimensional Navier-Stokes dynamics with exact SO(3) rotational equivariance, according t…
A systematic study of Kolmogorov-Arnold Networks (KANs) found that the architecture can match Multi-Layer-Perceptron performance in reinforcement learning with 40% fewer samples, with relative performance improvements up…
A new arXiv paper (2610.11087v1) proposes a verification-centric benchmark for LLM-assisted peer review that tests systems on detecting logical contradictions synthetically inserted into conference papers, rather than im…
Researchers introduced Nullify, a training-free activation steering method for LLM unlearning that uses inference-time steering vectors to redirect privacy-related activations away from memorized answers while satisfying…
A new arXiv paper (2610.10833v1) finds that small LLM agents fail to use explicit wall-clock time budgets productively even when they learn to respect them. Testing Qwen3.6-27B on five MLE-Bench Lite competitions and Qwe…
Researchers proposed CARing, a framework that represents diagnoses with compositional Semantic IDs (SIDs) and optimizes LLM reasoning trajectories for multi-label coverage, reporting R@30 of 46.04% on MIMIC-III and 46.52…
A new arXiv paper (2610.10666v1) offers a theoretical explanation for why adversarially-trained neural networks produce sparse gradient saliency maps, proving that for two-layer ReLU networks, minimizers converge to a Ba…
A new arXiv paper, arXiv:2610.10611v1, presents a formal framework that separates the mechanisms by which agentic AI systems improve — longer search, additional support, or modified proposal and verification — using boun…
Researchers propose Keyframe Mnemonics, a self-supervised method that discovers information-critical observations, or mnemonics, by learning an objective from randomly sampled past observations and using it as a reward f…
A new arXiv paper (2610.11007v1) formulates the curation of always-loaded LLM agent context files, such as AGENTS.md, as a capacitated assortment problem in which each loaded token is charged again in every later round o…
A new arXiv paper (2610.11012v1) introduces two distillation methods for AI safety: Distillation for Incrimination (DFI), which transfers misalignment but not concealment ability, and Distillation for Capabilities (DFC),…
OpenProblemBench, a new benchmark of 82 unresolved problems from the mathematics and theoretical physics literature, reports that GPT-6-Astra achieved the highest mean judged solve rate at 14.0% across seven evaluated co…
Researchers introduced ASPIRE (Atomistic Saddle-Point Inference with Refinement for Events), a framework that predicts multiple saddle-point candidates from a single initial atomic environment and refines them via Dimer …