RL environments providers

Creates reinforcement-learning datasets and environments, drawing on simulation and model-training work in coding, physical science, and alignment.

Turns specialist problem-solving into fine-tuning examples, grading rubrics, agent environments, and demonstrations of browser and desktop work.

Provides AI red teaming, alignment data, training environments, and runtime guardrails, with a focus on adversarial behavior and model safety.

Generates training environments, tasks, and verifiers from real-world material for long-horizon agent development. Its website describes a limited partner beta.

Provides human-validated data for model alignment, agent trajectories, speech, physical AI, and model evaluation.

Combines professional specialists, domain datasets, and task environments for AI training and evaluation in regulated knowledge work.

Offers resettable API sandboxes that model business services, allowing integration testing and agent rollouts against persistent but isolated state.

Creates stateful service twins accessed through APIs, command-line tools, or MCP for training agents and testing integrations outside production.

Builds long-horizon simulated tasks and benchmarks for agent security, reliability, scientific work, and operational incident handling.

Designs problem-solving worlds and converts them into post-training datasets and environments spanning engineering, knowledge work, and scientific workflows.

Builds agent-learning environments and benchmarks, including workplace simulations, robotics tasks, and a runtime for executing and verifying agent work.

Builds production-style RL environments and infrastructure, with agent evaluation and optimization tools and open research datasets.

Sources and structures clinical and preclinical records for model development, with additional tools for causal analysis and task reward functions.

Creates specialist-authored tasks and graders for training and evaluating models on finance, legal, computer-use, and responsible-AI work.

Builds simulated work settings with users, tools, persistent context, and tasks for training and evaluating agents under changing conditions.

Develops specialized reinforcement-learning tasks for scientific reasoning and inference or performance engineering.

Develops Cerebro, a platform for turning operational records into fine-tuning data, reinforcement-learning material, and outcome-graded environments.

Supplies post-training datasets and environments for coding, multimodal systems, and physical AI, alongside enterprise agent services.

Builds agent-training gyms that imitate workplace software and include tasks, datasets, and runtime infrastructure. Its website announces that Deeptune is joining Mercor.

Builds reinforcement-learning environments for model-training teams. Its engineering work converts software changes, agent traces, and CI failures into containerized tasks with checkable rewards, difficulty calibration, and held-out testing.

Transforms institutional lending records into training examples, reference trajectories, graders, and financial-workflow environments.

Develops environments that train models to conduct interpretability and alignment research, with additional in-house agent research.

Builds market-data-based environments for agents to form hypotheses, run experiments, and improve through increasingly difficult research tasks.

Constructs reproducible enterprise environments and training loops from seed data, including browser-based CRM, ERP, and commerce tasks.

Provides post-training datasets and environments for software engineering, machine-learning engineering, and autonomous research.

Creates multi-agent security worlds with persistent services and verifiable outcomes for adversarial training and adaptive benchmarking.

Builds simulated worlds and challenges to study and improve AI behavior. Its public homepage gives limited detail about individual commercial deliverables.

Converts expert decisions, corrections, preferences, and verification evidence into datasets, private evaluations, and regression environments.

Uses software-engineering expertise to turn production workflows into tasks, rubrics, evaluations, and training environments with scoring and telemetry.

Researches reinforcement learning, environment design, and multi-agent coordination over long tasks. Its current site emphasizes research and open releases.

Turns selected games into interactive learning environments designed to exercise planning, memory, tool use, and collaboration.

Builds finance-oriented training environments and benchmarks using professional workflows and anonymized transaction material.

Builds research-focused training environments intended to help models experiment and improve, and markets these environments to frontier labs.

Runs automated experiment loops against a customer-defined metric to improve models or algorithms, including generated training environments.

Offers an SDK and platform to build, run, inspect, and grade training environments; its site also connects environment creators to DataVendor.

Builds long-horizon training environments, computer-use models, and enterprise evaluations, with post-training data and infrastructure for AI labs.

Builds environments and graders for learning agents and offers selective technical research engagements from a small London research group.

Develops evaluation tasks, environments, and off-the-shelf datasets for long-horizon software work and other professional capabilities.

Builds infrastructure for training autonomous cyber capabilities, including verifiable environments based on realistic security challenges.

Offers data services for generative, agentic, and physical AI, including custom training environments, collection, and model-safety evaluation.

Builds environments used to train frontier AI and offers enterprise agent products for engineering and research workflows.

Captures human computer workflows as synchronized screen video, input events, and narration. Offers computer-use datasets and CAD environments with task specifications, reference assets, output files, and grading rubrics.

Builds graded software-engineering environments where coding agents implement features, debug unfamiliar projects, and deploy applications. Task scores serve as reinforcement-learning feedback and evaluation measurements.

Supplies professional expertise, human evaluations, benchmarks, and environments; it also offers a business-workflow data monetization program.

Licenses ready-made reinforcement-learning environments, recorded agent trajectories, and fine-tuning datasets. Published examples span mainframes, backend services, Figma workflows, and reasoning over financial documents.

Develops expert data, realistic training environments, contextual agent evaluations, and demonstration data for robotics.

Builds reinforcement-learning environments from scientific and technical research problems. Its stated approach uses verifiable tasks and formal methods across machine learning, computational science, robotics, and cybersecurity.

Designs multi-agent simulations for training and assessing AI systems. Environments explore negotiation, coordination, enterprise tasks, and emergent behavior under explicitly chosen incentives and participant roles.

Transforms company documents, communications, and tool records into privacy-processed digital twins. These support multi-step agent tasks and evaluations that preserve relationships across organizational systems.

Researches model failures and converts them into training tasks, synthetic experiences, and simulated environments. Published work spans software engineering, computer use, knowledge work, alignment, and scientific reasoning.

Converts real-world data into simulated software environments for training and evaluating agents. Its public example combines multiple business tools with loadable tasks and scored agent rollouts.

Builds simulated work environments in which autonomous agents can practice extended tasks. Its data-lab approach combines applications, task data, and verifiers to improve agent reliability through experience.

Packages tasks mined from licensed private code changes into runnable environments with graded checks, hidden reference solutions, and solver trajectories. Delivery includes verification evidence and reproducibility materials.

Develops reinforcement-learning environments with an emphasis on robust reward functions and hardened graders. Public task examples include coding kernels, strategic games, mathematical reasoning, and image-data curation.

Offers an integrated stack for model training, evaluation, inference, and compute. Its environment hub and open-source tooling help teams create tasks, execute rollouts, and run reinforcement-learning workflows.

Combines domain data and expert review to produce reasoning datasets, sandboxed training tasks, and adaptive evaluations. Coverage includes scientific problems, software engineering, financial analysis, and physical-system sensor data.

Builds simulated software environments and verifiable tasks for agent training and evaluation. Its published coverage includes connected workplace applications, coding tasks, and Blender-based design and animation workflows.

Uses software engineers to develop code-focused datasets, expert demonstrations, evaluation tasks, and training environments, alongside its engineering talent business.

Builds extended reinforcement-learning tasks across software engineering, computer interaction, cybersecurity exercises, and incident response. Environments combine realistic systems with grading of task execution and diagnostic performance.

Researches data-centric AI and develops expert-data methods, training systems, and benchmarks for advanced model and agent capabilities.

Builds procurement environments and benchmarks that exercise supplier selection, purchasing policies, approvals, and invoice-resolution tasks.

Combines multimodal annotation, data orchestration, human review, and evaluation tooling, including support for agent trajectories and RL environments.

Builds research tools, evaluations, and environments for models working on empirical science and automated discovery.

Builds private evaluations and training tasks from complete creative workflows, beginning with video editing and the TimelineBench benchmark.

Creates AI training material, evaluations, and environments from industrial companies’ operational records.

Transforms business workspaces and expert input into simulated environments for agents learning multi-step professional tasks.

Captures production agent traces, recreates hard cases as training environments, and feeds resulting evaluation signals into custom-model improvement.

Creates agent-training data, simulated environments, human feedback, and evaluations covering tool use, coding, and model safety.

Provides AI safety evaluations, red teaming, and training environments through researchers and engineers working on frontier-model risks.

Supports frontier-model development with datasets, reinforcement-learning environments, and specialists in software, knowledge work, and STEM.

Creates simulated versions of systems, APIs, data, and users for agent testing, training, and software development before production use.

Designs model-learning tasks, rubrics, verifiers, feedback, and evaluations across professional and multimodal work.

Develops environments that generate and adapt training tasks to an agent’s capabilities, aiming to sustain progressively harder learning challenges.

Develops Campaign, a self-play system that turns proprietary data and evaluations into new tasks and environments. Access is described as early-stage.