Machine learning news — deep learning, reinforcement learning, neural architecture search, diffusion models, and new ML frameworks and libraries.
Researchers analyzing token consumption in LLM-based multi-agent software engineering systems found that code review stages consume the majority of tokens at 59.4%, with input tokens accounting for 53.9% of total usage. …
Researchers introduced Gaia2, a benchmark for evaluating large language model agents in dynamic, asynchronous environments where scenarios evolve independently of agent actions. Testing of state-of-the-art models reveale…
Ernest Chan presented Conditional Portfolio Optimization (CPO), a machine-learning method that trains a neural network to predict portfolio performance based on market features and then solves a constrained optimization …
A developer found that deploying a 70B Llama model with RAG features caused time-to-first-token (TTFT) to jump from 180 ms to 1.4 seconds, as the model recomputed identical attention states for repeated 6,000-token conte…
OpenAI’s GPT models and Meta’s LLaMA-family use transformer architectures that convert text into token IDs, then map each ID to a dense vector via an embedding matrix. These vectors are combined with positional encodings…
A developer built a Quantified-Self RAG (Retrieval-Augmented Generation) system that ingests a decade of personal medical records from messy PDF scans using Unstructured.io, Sentence-Transformers, and Qdrant. The pipelin…
Developer Anil Madhavapeddy released TESSERA v1.1 and GeoTessera 0.9, improving temporal embeddings and S3+Zarr support, while also advancing io-uring libraries for OCaml. The updates address performance issues and expan…
A speculative proposal argues that training massively overparameterized neural networks with extremely high learning rates on small, diverse datasets could trigger a "catapulting" effect, producing human-like generalizat…
SimTooReal, a platform for robotics teams, now enables live telemetry and failure diagnosis for training runs in Isaac Lab, MuJoCo, Gazebo, and LeRobot with a single command. The tool wraps existing training scripts to s…
Overworld released Waypoint-1.5, an update to its generative world model that produces real-time environments at up to 720p and 60 frames per second on consumer hardware. The release introduces two model tiers — a 720p v…
A developer built a three-tier content quality ladder for programmatic directory sites, using a single `model_used` column to track content origin across three sites launched in April. The system prioritizes Claude Haiku…
A new machine learning paradigm treats relational databases as graphs, enabling graph neural networks to learn directly from the raw schema without manual feature engineering. The approach, detailed in a 2023 paper by Fe…
A new analysis of metastable states in transformer activation spaces confirms that token representations cluster into metastable groups across layers in trained models, as predicted by a recent dynamical systems theory, …
In an experiment on Gemma-2-2B's residual stream, researchers found that transformers store long-range contextual information in a compact, low-dimensional subspace of roughly 31 directions, rather than spreading it acro…
A developer found that improving data quality through feature engineering boosted a Logistic Regression model's accuracy from 72% to 86%—a 14-percentage-point gain—without changing the algorithm. By handling missing valu…
Patrick Jiang (@patpcj) released Harness-1, a 20 billion parameter search agent that externalizes search state from the model into a structured harness, as detailed in a 13-post thread on X and an accompanying paper. Res…
MIT researcher Markus Buehler announced Friday that his team has published a paper on arXiv establishing a formal framework for AI-driven scientific discovery, arguing the process constitutes a fundamental shift in repre…
A team of developers has built an AI-powered job search tool that uses a fine-tuned Qwen3-8B language model to generate LinkedIn queries, scrape job postings, and score each role against a candidate's resume across five …
A developer built TrustGuard AI, a fraud detection system that catches 99.76% of fraudulent transactions on the PaySim dataset, achieving an AUC-ROC of 0.9995 and Recall of 0.9976 on a 6.3 million row test set. The syste…
GGUF, GPTQ, and AWQ are the three dominant formats for running quantized large language models locally, each optimized for different hardware and use cases. GGUF, the format used by llama.cpp and its derivatives, offers …