Machine learning news — deep learning, reinforcement learning, neural architecture search, diffusion models, and new ML frameworks and libraries.
Efficient reasoning training does not always harm chain-of-thought faithfulness and monitorability, according to research addressing the concern that training large language models to solve tasks with fewer tokens degrad…
Yandex Music replaced its multi-stage recommendation cascade with Sona, a single generative recommender built on one transformer that performs both candidate generation and ranking without hand-engineered features, accor…
Stanford professor emeritus John Ousterhout presented the Homa protocol at the AI Engineer Summit, claiming it cuts short-message P99 latency in AI clusters from 1.2 milliseconds under TCP to 92 microseconds — a 13x impr…
A team fine-tuned MedGemma 4B with LoRA and MedGemma 27B with QLoRA for oncology tasks using Unsloth 2026.9.12 on a single DGX Spark, publishing the report with a DOI on Zenodo and the models and dataset publicly. The te…
A technical case study reports that a single NVIDIA RTX 4090 with 24 GB of VRAM can run a 125-billion-parameter Qwen 3.8 Flash Next model at roughly 100 trillion tokens per second, a throughput the writeup says was previ…
A developer built Collection Radar, a local AI tool that ingests a wholesale business's existing Excel invoice spreadsheet and converts it into a prioritized collections list. The tool uses TabPFN, a tabular-data model, …
Strata v0.1.38 runs Qwen3.8-Flash-Next, a 125B mixture-of-experts model, on a gaming GPU at 94 tokens per second without a server, according to byteiota. The report attributes the feat to the model's MoE architecture and…
A study asks whether learning protein folding generalizes to broader reasoning, using protein folding as a testbed because one solved structure yields thousands of exactly checkable spatial and topological statements. Th…
A new arXiv paper (2610.02478v1) proposes Tropical Reinforcement Learning, which replaces the standard sum of trajectory probabilities with a maximum over the tropical semiring, and reports that its TROPIC training algor…
A new arXiv paper (2610.02254v1) reports that an integrated LLM-ISM approach can replace repeated subject-matter-expert interviews in Interpretive Structural Modeling, with rowwise causal graph discovery scoring an SHD o…
A new arXiv paper (arXiv:2610.02255v1) introduces MACTS-EM, a multi-agent collaborative time series forecasting framework that combines domain-specialised forecasting agents, a meta-cognitive allocation layer, emergent m…
Researchers introduced MintFlow, a training-free constrained sampling framework that enforces constraints on flow matching models by applying a minimal perturbation to an intermediate flow state while leaving the pretrai…
Researchers proposed Comparative Inference for Tool-use Agents (CITA), a method that trains a Comparative Inference Model (CIM) to estimate the long-horizon value of a possible next tool invocation before an LLM agent ex…
Researchers introduced MEA, a reward-driven multi-agent framework that generates natural-language explanations of machine learning model behavior, reporting faithfulness gains of +28% on tabular data, +21% on text, and +…
A paired evaluation of two System-1 decision models for LLM agent harnesses found the hosted model Jev significantly more accurate than the open-weight model Laya on 9 of 11 agent decision points, by margins of +10.8 to …
Researchers proposed THPL, a vision-to-language decision support framework for rainbow trout feeding management in Recirculating Aquaculture Systems (RAS), reporting that its Activity Coefficient correlated with expert-a…
A new arXiv paper (2610.02405v1) reports that using a frontier model such as Claude Opus as a meta-agent to generate terminal tasks and verifiers for RL training can produce unfaithful pipelines despite runnable Docker i…
A data-driven control framework combining surrogate modeling and reinforcement learning achieved an almost 2x improvement over operators' actions in optimizing target polarization in nuclear physics scattering experiment…
A Mixture-of-Agents (MoA) study measuring per-token sufficient compute found that a 0.5B agent reproduces 92–95% of reference tokens on three core benchmarks, while the most expensive 10% of tokens account for 64–80% of …
A hybrid Raman spectroscopy framework called HyMLRaman identified six pharmaceutical compounds with 96.31% accuracy and a 96.36% macro-F1 score using an EfficientNet-B3 plus SVM configuration, according to an arXiv paper…