Machine learning news — deep learning, reinforcement learning, neural architecture search, diffusion models, and new ML frameworks and libraries.
Elorian and Scale AI released Humanity's Sixth Sense (HSS), a benchmark of 522 questions covering 288 images and 234 short video clips (17.6 hours of video) that measures how models reason about information a visual scen…
The BrokenML benchmark, released by developer zhebrak on GitHub, ran nine models against 46 broken distributed ML jobs at three access levels and found that higher access improved every agent's performance while making f…
Microsoft released Microsoft-Decision-1, a small decision-scoring model that returns a calibrated probability for each option given a situation, a question and a fixed set of answer options. Achint Srivastava, VP of Soft…
A GitHub repository published by abduldattijo reproduces three AI text detectors — perplexity, near-duplicate similarity, and embedding density — plus a sentiment-classifier evaluation on 600 training reviews (200 source…
A developer proposes "consequence models" — models that take the state of the world and a proposed agent action and return the probability the action succeeds and how it is likely to fail — as a missing layer in most age…
Epochlypse Research released EpochGo, an open-source local-first AutoML platform that runs three autonomous AI agents on Ollama with zero API keys and $0 cloud costs. The system uses CrewAI with the recommended qwen2.5-c…
Researchers submitted Oído, an inference engine that runs open-vocabulary English speech recognition entirely on a $5 ESP32-S3 microcontroller with no neural accelerator, executing a 13M-parameter Conformer-CTC model wit…
HAL-X AI has open-sourced THX-01, a 322-million-parameter non-autoregressive decision model that returns structured answers with calibrated probabilities in a single forward pass, and reports over 70 million tokens proce…
Google DeepMind's AlphaProof Nexus solved dozens of open math problems, including nine Erdős problems — two of which had been unsolved for 56 years — according to a study published in Science. The results were peer revie…
The LFORLA Reverse Engineering benchmark scores LLMs on restoring C source from stripped binaries using identical prompts, a deterministic token-similarity judge, and four per-axis rubrics with a public verbatim answer t…
A developer trained a 22-million-parameter language model from scratch on a laptop with only 4 GB of VRAM, using a Lenovo IdeaPad Gaming machine with a GTX 1650, PyTorch and CUDA under WSL2. The run used a 256-token cont…
SotaTek outlines a systems-engineering approach to AI development that treats model selection as one decision within a larger lifecycle spanning problem definition, data readiness, and architecture choice. The framework …
The National Hurricane Center built its first advisory on Hurricane Isaias around Google DeepMind's WeatherNext forecast of a future Category 2 storm, and Isaias made landfall near Destin, Florida, around 8:30 p.m. local…
The EvoSyn framework reports a 13.3-point average task-success gain over the strongest synthetic-data baseline on τ²-Bench and CAR-Bench by making the synthesis space itself editable from learner feedback. EvoSyn convert…
Researchers at the University of California, San Diego trained a Qwen2.5-7B-Instruct agent with built-in theory of mind, reporting an average relative gain of approximately 11.5% on the Sotopia social-interaction benchma…
Hugging Face has converted OpenAI's 300+ math proofs into open-source reinforcement learning environments for training AI models, according to OfficeChai. The environments let developers use OpenAI's proofs as RL trainin…
A developer built Greenstep, an open-source app that recommends when, where, and what kind of walk to take, using a three-part pipeline of TabPFN for hourly predictions, SerpApi for nearby place suggestions, and Gemma 3 …
Saina Helm 2.0, a 0.8B-parameter decision model for workflow automation built by the developer run-saina, scored 30.17 on the public Decision Index 0.3 suite using model revision 9903c487a3b1eef45c8cc56c76155b9baccb55f9,…
Vllm-sr released the Decision 3.0 collection of open multimodal foundation decision models on Hugging Face, spanning six sizes from 0.6B to 26B parameters. The collection includes d3 (26B), d3-flash (8B), d3-mini (5B), d…
A paper submitted to arXiv on 28 Sep 2026 introduces "harness learning," a method that trains a proposer model via reinforcement learning to revise a solver language-model agent's harness — the executable program organiz…