Primus by Transformer Lab

Primus for Math is an autonomous AI research scientist designed for mathematics and engineering. It automates the research workflow by reading existing literature, deriving precise models and equations from high-level problems, running computations, and logging its reasoning process (including failed paths) for user verification.

Primus acts as an automated machine learning researcher. It takes a natural language prompt (even a single sentence) and executes a full research cycle: scoping the question, reviewing existing literature, designing and running experiments on GPUs, evaluating results, and authoring a final research paper. Users can monitor progress in a live workspace and provide steering notes (e.g., "focus on the smaller model first") to guide the AI's direction.

Primus is an autonomous AI researcher developed by Transformer Lab. It automates the full research loop: hypothesizing, reading literature, coding, running experiments on real compute, iterating, drawing conclusions, and writing final papers. It is designed to handle tasks typically assigned to a team of expert ML researchers across domains such as LLMs, seismology, physics, crystallography, and biology.

The RAG (Retrieval Augmented Generation) feature provides a drag-and-drop file user interface for document ingestion and is compatible with multiple backend engines, including Apple MLX and Hugging Face (HF) Transformers.

Transformer Lab focuses on Recursive Self-Improvement (RSI) to automate the act of scientific discovery. Their goal is to replace the manual human effort of hypothesis generation, architecture design, experimental coding, and optimization with an "AI Scientist" system. This approach aims to remove the "biological speed limit" of research—such as the time required for human schooling, reading, and manual experimentation—allowing for research cycles to happen in months rather than years.

Primus is an autonomous ML research scientist designed to accelerate the research loop (hypothesizing, literature review, coding, experimenting, and paper writing). It is capable of running this loop significantly faster than human researchers, autonomously producing complete research papers across diverse fields such as LLMs, seismology, physics, crystallography, and biology.

The Advanced Training & Multimodal Workflows environment provides a unified platform for pre-training, fine-tuning, and evaluating Large Language Models (LLMs), Diffusion models, and Audio models. It supports production-ready preference optimization techniques including Direct Preference Optimization (DPO), Odds Ratio Preference Optimization (ORPO), Simple Preference Optimization (SIMPO), and Group Relative Policy Optimization (GRPO). Additionally, it enables complex multimodal workflows such as implementing ControlNets, image inpainting, and fine-tuning Text-to-Speech (TTS) models using custom datasets.

Primus is an autonomous AI research agent for machine learning. It allows users to start a research project with a single-sentence prompt. The system handles the end-to-end research pipeline: scoping the research question, conducting literature reviews, designing and running experiments on provisioned GPUs, and writing the final paper (including methods, results, and citations). It is designed to be usable by both ML engineers and non-experts.

Primus Society is a virtual community of 10,000 autonomous AI researchers designed to tackle frontier problems in machine learning, quantum mechanics, and mathematics. Unlike single-agent systems, it uses a population-based approach where agents are organized into labs, governed by a central institute, and compete for compute resources. The system mimics human scientific institutions—using grants, peer review, and conferences—to ensure diversity of thought and coordinated progress.

The 'parseable-answering rate' is the frequency with which a model writes its final answer in the requested format (e.g., a number wrapped in a box). This rate acts as a 'gate' for reinforcement learning (RL). If the answering rate is near zero, the model is not producing readable answers for a grader to evaluate, rendering any reward signal—including perfect answer keys or random rewards—ineffective. Training cannot improve accuracy until the model consistently produces a parseable format.

Experiment Tracking allows you to automatically record hyperparameters, code versions, metrics, and logs for every run. This creates a consistent history within a single workspace where jobs are launched and monitored, enabling you to revisit, compare, and resume experiments without manual searching.