RL glossary

The words this site uses to sort the RL training market, and how a vendor ends up in each one.

Sections

What a company is listed under, next to Benchmarks. A company is in every section it sells into: Scale AI sells RL environments and data, so it is in both, with one page and one tier.

RL environments 75 companies →
Where an AI agent practises and is graded: simulated workplaces with automatic graders, and the evals and verifiers that score models (fixed tasks with graders). Bought by labs to train and measure models. For example: Fleet, Plato, Mechanize, Vals AI. Includes: RL environments and Evals.
Data 51 companies →
Data to train and grade models with: networks of experts who write, judge and demonstrate tasks (human data), and ready-made datasets. For example: Mercor, Surge AI, Turing, Wirestock. Includes: Human data and Datasets.
Infrastructure 15 companies →
What RL runs on: sandboxed machines an agent can act in, compute for rollouts, and platforms to train models with reinforcement learning. For example: Modal, E2B, Daytona, Prime Intellect. Includes: Sandboxes and compute and Training platforms.

What companies sell

The finer categories, each read from the company's own site. A company can sell several, each either core (what it leads with today) or expanding (newly announced, beta or a recent launch). Each belongs to one section, except agent safety and enterprise services, which every company with them sells next to environments or data.

RL environments 56 companies →
A simulated workplace where an AI agent practises a task (a code repository, a browser, business software) and an automatic grader scores each attempt. Labs train on thousands of these attempts. Bought to improve models. For example: Fleet, Plato, Mechanize, Deeptune.
Evals 58 companies →
A fixed, held-out set of tasks with fixed scoring, used to measure and compare models. Never trained on, so the score means something. Used to measure models. For example: Vals AI, Patronus AI, Scale's leaderboards.
Human data 32 companies →
A network of experts (engineers, doctors, lawyers, analysts) hired to write, judge or demonstrate tasks: the human judgement behind training data, graders and evals. For example: Mercor, Surge AI, micro1, Turing.
Datasets 34 companies →
Ready-made or licensed data sold as a product: post-training sets, agent trajectories, licensed corpora. Bought once, not commissioned task by task. For example: Wirestock, Exabite.
Sandboxes and compute 12 companies →
The machines environments run on: isolated containers or virtual machines, started by the thousand, where an agent can safely run code or click through apps. They sell the runtime, not the tasks. For example: E2B, Daytona, Modal, Runloop.
Training platforms 4 companies →
Hosted services for running RL or post-training jobs and serving the resulting models. For example: Prime Intellect.
Agent safety and verification 6 companies →
Red-teaming, guardrails and runtime checks that test or constrain what agents do once deployed. For example: Gray Swan, Align AI, Veris AI.
Enterprise AI services 7 companies →
Engineers and applications deployed inside a company to build or run its AI systems. For example: Scale (applications), Mercor (forward-deployed engineers).

Domains

The subject matter of the tasks or data. A vendor lists every domain it covers and is shown under the one it leads with. A vendor that covers three or more with none leading is Multi-domain. Pure infrastructure has no domain.

Coding
Writing, reviewing and running software: repositories, terminals, tests.
Computer use
Operating a computer like a person: browsers, desktop apps, GUIs.
Enterprise work
Knowledge work in business software: CRM, ERP, email, spreadsheets, support, operations.
Finance
Banking, trading, accounting, insurance and financial analysis.
Health and bio
Medicine, clinical work, biology and drug discovery.
Science
Physical sciences and research workflows outside biology.
Math
Mathematical reasoning and proofs.
Security
Cybersecurity, capture-the-flag tasks and AI red-teaming.
Games
Video games and game-like simulations.
Robotics
Embodied and physical AI: robots, factories, manufacturing.
Chip design
Hardware and semiconductor design.
Media
Audio, voice, image, video and creative work.
Multi
Generalists: vendors covering three or more domains without one they lead with. Their specific domains are still listed.
General capability
Broad skills that belong to no single field: reasoning, knowledge, language, long context, instruction following and safety. ARC-AGI, MMLU-Pro, long-context and hallucination evals.

Task traits

Properties of the tasks themselves, recorded when a vendor's site states them.

Long-horizon
Tasks that take many steps or hours to finish.
Tool use
Tasks that call tools or APIs, including MCP servers.
Reasoning
Tasks that test multi-step reasoning more than any one skill.
Private data
Built on private codebases or proprietary data.
Multilingual
Tasks in several human languages.

How vendors are placed

  1. We read each vendor's homepage and one product page from its own website.
  2. A model sorts the vendor into the fixed categories above, using these definitions. It can only pick from this list.
  3. Every offering needs a quote from the vendor's site. If the quote can't be found on the page, the offering is dropped.
  4. Pages are re-read monthly, sooner when they change. New offerings, dropped ones and moves into a new domain show up as dated changes on the landscape.

Categories version 2026-10-11.1.