Machine learning news — deep learning, reinforcement learning, neural architecture search, diffusion models, and new ML frameworks and libraries.
Vilya, a new deep learning model called Vilya-1, improves geometric accuracy and broad chemical coverage for macrocyclic peptide design, addressing core challenges in drug development. Trained on diverse datasets with a …
A new study reveals that low-precision training can cause AI model parameters to freeze when gradient updates fall below half the unit in the last place (ULP) of a weight, effectively halting learning. The phenomenon was…
Ground truth datasets in machine learning are not objective or universal but are constructed through human decisions and technological limitations, argues a new analysis calling for transparency and context-aware reliabi…
A new method proposes a predictive divergence mask for reinforcement learning in large language models, offering potentially more stable updates than traditional Proximal Policy Optimization (PPO)-style approaches. The a…
The Energy-guided Recursive Model (ERM) uses Hopfield energies to achieve a 98.97% completion rate on Sudoku puzzles, 88.04% on the Pencil Puzzle Benchmark, and 99.30% on Maze puzzles, signaling a potential shift in AI r…
Researchers have introduced a bilevel optimization framework for Direct Preference Optimization (DPO) that improves large language model alignment by filtering out noise in preference data. The method uses a learnable we…
Volvo AB's autonomous trucks aim to slash per-mile costs to $1.35 from the current $11.35, achieving capital recovery within 18 months. The Swedish manufacturer promises over 80% reduction in operational expenses, positi…
LFortran, LLVM, and Enzyme have been combined to automatically differentiate legacy Fortran simulation code, producing exact gradients through a multi-step time loop that match an analytic answer. The pipeline wraps a 22…
A developer building a multi-agent document-search copilot describes how they resolved the Hybrid retrieval strategy by routing on filter selectivity: when a filter matches few documents, the system filters first then ra…
Q-Score, a new molecular docking method using Graph Neural Networks and quantum optimization, outperforms classical scoring functions by predicting orbital donor-acceptor energies and solving a maximum-weight vertex cliq…
Graph Neural Networks (GNNs) are redefining RFID localization by focusing on relationships between data points rather than just coordinates, according to a new analysis. This graph-based approach integrates signal streng…
New research introduces a nonparametric Bayesian method for inverse reinforcement learning that handles diverse expert inputs by using a Dirichlet Process prior over reward functions. The approach achieved an Adjusted Ra…
Richard Sutton, co-winner of the 2024 Turing Award for founding modern reinforcement learning, announced Monday he is leaving John Carmack's startup Keen Technologies to launch his own AI startup, Oak Lab, with collabora…
A study using log data from Japanese junior high students using a grammar app found that AI-generated English learning content can improve performance, with multiple-choice questions requiring the least effort and cloze …
Manticore Search 28.4.4 introduces faster KNN rescoring through batched distance calculations, conversational search via the HTTP API, and a simplified install path with a single curl command. The release also adds zero-…
TreeThink, a new open-source Python library for neural theorem proving, offers modular asynchronous tree search with native integration for formal verifiers in Lean 4, Rocq, and Isabelle/HOL, achieving up to 6.3 times sp…
A new benchmark called QIMG-7 reveals that multimodal retrieval-augmented generation systems suffer significant accuracy drops when faced with polluted data, with the GPT-4o-mini stack falling from 0.908 to 0.490 on corr…
A new method called UMoE realigns Mixture-of-Experts (MoE) models to enhance domain-specific performance without increasing expert count, parameters, or inference cost. Across Qwen3-30B-A3B and Qwen3.5-35B-A3B architectu…
Researchers have introduced AdvancedMathBench, a new benchmark suite for evaluating large language models on advanced mathematical reasoning, including proof generation and verification. The benchmark comprises ProverBen…
EasyOPD, a new framework built on the verl platform, simplifies AI model training by integrating on-policy distillation (OPD) methods into a unified system, enabling faster and more efficient model evolution. The framewo…