July 2026 Summaries

The open research ecosystem has developed numerous datasets for software engineering, terminal use, and web research, each with unique harnesses, image conventions, and grading scripts, leading to challenges in integration and evaluation. To address this, an integrated system has been introduced that consolidates 23 tasksets under a unified API, allowing for consistent evaluation and reinforcement learning (RL) training across approximately 365,000 tasks. This integration maintains the original grading paths of each taskset while normalizing them around a single API, ensuring that the original scoring semantics remain intact. The tasks are organized into three main domains: software engineering, terminal, and search, with a focus on creating a seamless and scalable environment for RL training. To ensure the integrity and reliability of the datasets, a rigorous validation process is employed, filtering out tasks with broken images, unstable tests, or solvable without intended code interaction. This consistent validation and re-upload process aims to maintain high standards and transparency, with validated datasets available for public use, enhancing the accessibility and reproducibility of RL experiments in agentic domains.

Verifiers 0.2.0 introduces a revamped core designed to support composable tasksets and harnesses, facilitating modern agentic training and evaluations. This update allows for more sophisticated evaluations that go beyond simple prompts by incorporating coding agents equipped with tools and custom logic. The new architecture consists of a taskset, harness, and runtime, enabling the execution of arbitrary tasks essential for current evaluations and training. The core features include composable tasksets that can operate with any compatible harness, swappable runtimes for flexible deployment, and first-class branching for non-linear rollouts. Additionally, verifiers v1 introduces an interception server to manage requests between agents and inference servers, supporting multiple dialects to ensure compatibility with different agent languages. With the transition to verifiers.v1, the legacy codebase is deprecated, and emphasis is placed on the new abstraction for scalability and efficiency in training, leading towards a full 1.0.0 release that promises multi-agent environments and comprehensive support for various environment frameworks.

The company has announced a $130 million funding round led by Radical Ventures, with contributions from NVIDIA Ventures, Intel Capital, Dell Technologies Capital, and existing investors, bringing its total funding to over $150 million to develop the open superintelligence stack. This funding will help scale its infrastructure, allowing companies to own their model optimization loop and build continuously improving AI agents. The open superintelligence stack encompasses training, deploying, and refining models, and is already used by over 6,000 customers, generating more than $100 million in annualized revenue. Notable success stories include Ramp, which developed a model that outperformed closed frontier models in efficiency and cost. The company is focusing on scaling compute clusters, enhancing reinforcement learning (RL) capabilities, and advancing long-horizon agents and Recursive Language Models (RLMs) to address complex AI challenges. They are actively hiring to expand their team and further develop infrastructure that supports open superintelligence initiatives.

Prime-rl has introduced a first-class algorithms layer designed to centralize algorithm-specific elements within its reinforcement learning framework, which now supports six built-in algorithms—GRPO, MaxRL, On-Policy Distillation, Self-Distillation, SFT distillation, and ECHO. This new layer allows for algorithm selection on a per-environment basis, enabling a single run to train different algorithms on different environments without altering the core trainer, thus enhancing flexibility and performance. The centralization is achieved by organizing each algorithm as a module under a common orchestrator directory, simplifying the understanding and modification of algorithms without compromising system performance. This setup allows researchers to innovate by subclassing a common abstraction without delving into trainer internals and supports multi-teacher distillation, enabling environments to utilize specialized teacher models for more domain-specific training. By resolving algorithms per environment rather than per run, prime-rl offers a unique approach not commonly found in other open-source RL frameworks, facilitating more targeted and effective training signals based on the specific properties of each environment.