Machine learning news — deep learning, reinforcement learning, neural architecture search, diffusion models, and new ML frameworks and libraries.
Anygrep, a new MIT-licensed terminal tool from developer kuroko1t, performs semantic search over notes, PDFs and images using Google's EmbeddingGemma 2 model converted to int8 ONNX and run entirely on the CPU, keeping a …
The CTML framework, detailed in a paper on arXiv, provides automated continuous testing for machine learning systems across four areas: data quality, drift detection, model quality, and pipeline validation. CTML's modula…
A paper by Marco Zullich and his team, arXiv:2610.12493, finds that current explainable AI faithfulness metrics are structurally misaligned, lacking a reliable ground truth and forcing computationally expensive shortcuts…
Amazon Web Services researchers introduced "priming," a method that reuses pretrained Transformer weights to cut the pretraining cost of hybrid state space architectures by more than 99%, according to a new paper titled …
EliteTable, a fixed-capacity table of high-scoring solutions that guides an LLM to generate, edit and remix executable control policies, reached the search-score ceiling in the first generation across all ten independent…
NVIDIA and Princeton's Dao AI Lab released ReplaySSM, a method that represents the SSM state as a checkpoint plus a buffer of recent inputs to cut state memory traffic, enable rollback for speculative decoding, and limit…
Meta has built a system that uses idle inference capacity across its fleet to improve ads recommendations, according to a post on Engineering at Meta. The system relies on two elasticity levers that let ads recommendatio…
Google announced the preview of TimesFM 3.0 in BigQuery, a 330-million parameter foundation model from Google Research that adds native multivariate time-series forecasting to BigQuery SQL via AI.FORECAST. TimesFM 3.0 ra…
Precision Numerics released Boole-1.7B, a 1-bit version of Qwen3-1.7B that retains 85.9% of BF16 performance at 247MB, 13.9x smaller than the full-precision model. The company said Boole-1.7B outperforms 1-bit Bonsai-1.7…
A developer outlined a technology-agnostic architecture for AIoT pipelines that move from raw sensor events to validated actions, using a warehouse tracking scenario as the example. The design separates event validation …
Allora's Forge launched topic 90, the Hyperliquid Log Return topic, on October 8, 2026, asking AI and machine learning models to forecast three-minute base-10 log returns across Hyperliquid perpetual futures markets with…
A developer built a Pac-Man agent that reaches an average of 9.8 ghost kills per game using Jeffy, an open-source CPU text classifier, instead of an LLM, according to the Jevman benchmark project on GitHub. The agent par…
Flo Health raised its open models' medical safety score from 34% to approximately 61% and usefulness from 42% to 76% — a combined medical safety and usefulness score of 16% to 47% — using a privacy-first, judge-driven po…
Logical Intelligence's energy-based reasoning model Kona, roughly a thousand times smaller than frontier LLMs, solved 96.2% of public Sudoku puzzles for about $4 in GPU time, while a panel of frontier models including Op…
A recent preprint shows that a backdoor implanted in a LoRA adapter leaves a detectable signature in the adapter's own weights, allowing detection without executing code or running a forward pass, according to the post. …
Researchers submitted a paper to arXiv on 9 Oct 2026 presenting Artisan, a multi-agent pipeline that extends the Articraft method with physical simulation feedback to generate articulated 3D assets with collision-free mo…
Byte Transformers trained with token-superposition training and hash embeddings consistently outperform subword Transformers as model size scales, according to an arXiv paper submitted on 5 October 2026 by researchers st…
BCI Report released its questions-in-language experiment on 8 October 2026, testing whether an EEG decoder can answer a question about a steady-state visual evoked potential (SSVEP) label it was never trained on. Using t…
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…