Machine learning news — deep learning, reinforcement learning, neural architecture search, diffusion models, and new ML frameworks and libraries.
Linear probes on residual stream activations recorded immediately before an LLM judge's initial verdict can predict whether that verdict will flip when candidate responses are reordered, according to an arXiv paper (2610…
A study posted to arXiv (2610.07208v1) found that LLM-assisted regularization produced mixed migration-forecasting results across corridors between November 2021 and November 2022, with no single regularization approach …
A neutrosophic ensemble of Random Forest, XGBoost and Logistic Regression reached 100.00 percent accuracy on three of four held-out CWRU bearing-fault loads (92.27 percent on the fourth) but collapsed to 40.64 percent on…
A Vision Transformer-based model with dual-pathway temporal and spatial attention achieved the highest deterministic and probabilistic skill at all lead times for subseasonal soil-moisture forecasting over Europe, outper…
A new arXiv paper (2610.07475v1) presents a finite-depth framework for estimating policy sensitivity in reinforcement learning, approximating the policy Hessian and mixed derivative from a reference environment to predic…
Researchers introduced OxiGen, an oxidation-state-aware crystal diffusion model that explicitly represents oxidation states during generation and enforces global charge neutrality by construction using a structured outpu…
Three practical tutorials on numerical computation and machine learning for mathematical research were released on arXiv as paper 2610.07220v1, developed for the DANGER: Data, Numbers, and Geometry workshop held at the B…
CNet, a C++/CUDA framework for building and training complex-valued neural networks, was released on arXiv as paper 2610.08592v1 with code at github.com/crasmarum/CNet. The framework implements Wirtinger (CR-calculus) au…
Researchers introduced BeFOND, an encoder-free sparse coding model that unifies inference and dictionary learning as natural-gradient flows on a shared variational free energy, according to arXiv paper 2610.07389v1. On s…
Researchers introduced learnable spectral activations (LSA), a method that replaces fixed neuron-level nonlinearities with a residual truncated Fourier series whose harmonic amplitudes are learned during training, accord…
A new arXiv paper (2610.07555v1) introduces Global Transport (GT), a global class-agnostic optimal-transport coupling computed without class labels, which consistently improves conditional image generation when combined …
OpenAI researchers identified a set of sparse autoencoder (SAE) latents closely linked to metagaming — a model reasoning about how a task is evaluated or rewarded rather than attempting it — by examining a capabilities-f…
The Allen Institute for AI released Olmo-core 3, an open mixture-of-experts training framework that scales MoE training into the trillion-parameter range while keeping throughput losses under 5%. In a preliminary test on…
Jeffy released 68 pretrained classifiers that run entirely on CPU with no GPU, API keys, or cloud dependency, spanning 17 English classifiers and 51 multilingual classifiers covering 60 voice-command intents per language…
New research measures and aims to improve the alignment between large language models' chain-of-thought traces and their internal computations, addressing evidence that CoT often fails to reflect how models actually arri…
The Burn deep learning framework shipped its 0.22.0 release, removing generic types from user code to cut CNN recompilation time from 28.4 seconds to a 6.2x speedup, and replacing MLIR, CUDA/HIP transpilers and SQLite wi…
A developer built Aurora Odds, a browser-based tool that uses the open-weights TabPFN tabular foundation model to forecast whether the aurora will be visible from a user's exact location in the next hour. The model reads…
Researchers introduced UnAct, a gradient-free machine unlearning method that removes the influence of designated training data by intervening on targeted activations rather than retraining or backpropagating. UnAct is po…
A developer published a raw benchmark dataset measuring how reasoning effort levels affect GPT-5.6 Sol's performance on modular arithmetic tasks, logging per-call token counts, latency, and cost across 23 runs. The resul…
A developer has proposed an AI-augmented HR architecture that replaces quarterly performance reviews with continuous telemetry, integrating directly with operational tooling such as git commits, ticketing systems, and de…