Horizon

AI model · Multimodal model · Text → Text by Labelbox

Environments for post-training frontier AI at scale.

Horizon is a simulation platform for post-training frontier AI models, providing RL environments for reasoning, tool use, and computer use. It enables AI labs to train agents in realistic enterprise simulations like Jira and GitLab with calibrated reward signals.

Model specs

The full spec card (context, pricing per token, modalities and more) on AntSpace. Sign up free to see the full spec card.

Languages
English

Category: MLOps

Primary Use Case
Post-training RL and Evaluation
Simulation Engine
WorldSim
Supported Modalities
Text, Multimodal, Computer Use, Voice
Integration Support
Hundreds of MCP tools
Target Optimization
Pass@k targets
Environment Types
Software-generated RL environments
  • RL training gyms for reasoning, tool use, and computer use
  • WorldSim enterprise environment recreation (Jira, GitLab, CRM)
  • Calibrated reward signals and target pass@k optimization
  • Expert-authored problems for frontier AI reasoning evaluation

Overview

Horizon is Labelbox's simulation platform designed for post-training frontier AI models at scale. It provides Reinforcement Learning (RL) training gyms and evaluation environments for reasoning, tool use, and computer use across high-value economic domains. The platform features WorldSim, which recreates enterprise environments such as GitLab, Jira, and CRM systems using realistic business data to generate diverse tasks for AI agents. Horizon solves the reward signal problem by delivering calibrated feedback and target pass@k optimization for complex, long-horizon tasks. It is used by leading AI labs to teach models "taste" through preference labels and expert-authored graders, ensuring models learn real capabilities rather than shortcuts.

Simulation Environments

  • RL Training Gyms: The platform provides software-generated RL environments at scale for reasoning, tool use, and computer use.
  • WorldSim Enterprise Simulation: WorldSim recreates enterprise environments like GitLab, Jira, CRM, email, and chat using realistic business data.
  • MCP Tool Navigation: Agents navigate hundreds of MCP tools across computer and terminal tasks.

Reward & Evaluation

  • Calibrated Reward Signals: Delivers calibrated reward signals and target pass@k optimization across valuable knowledge work domains.
  • Human Preference Signal: Preference labels turn human judgment into structured comparisons across agent trajectories.
  • Pass@k Optimization: Evaluation is tuned to specific pass@k targets and RL infrastructure requirements.
  • Expert-Authored Graders: Expert-authored problems and rubric-graded prompts provide high-quality grounding signal.

Specialized Domains

  • Agent Coding & Software Engineering: Supports long-horizon software tasks across real codebases, including debugging production failures and authoring PRs.
  • Computer Use: Enables GUI-based task execution across enterprise software with verifiable multi-step outcomes.
  • Cybersecurity Training: Provides attack and defense scenarios with CTF-style tasks and programmatic verification.
  • Multimodal Knowledge Work: Supports tasks spanning text, images, charts, and documents within a single workflow.
  • Voice with Agentic Tool Use: Voice-native agents reason, plan, and execute tool calls mid-conversation.

Use cases

  • AI researchers use Horizon to create long-horizon reasoning tasks with intermediate reward signals for autonomous research.
  • Software engineering teams train agents on real codebases to debug production failures and author pull requests.
  • Enterprise teams simulate GUI-based task execution across CRM and terminal tasks to verify multi-step outcomes.
  • Security researchers use CTF-style tasks and adversarial edge cases to train and evaluate AI attack and defense capabilities.
  • Multimodal AI developers train models to reason across text, images, and charts within a single complex workflow.

Specifications

Modalities
Text, Image, Chart, Document, Voice, Computer GUI → Text, Action, Reward Signal
Primary Use Case
Post-training RL and Evaluation
Simulation Engine
WorldSim
Supported Modalities
Text, Multimodal, Computer Use, Voice
Integration Support
Hundreds of MCP tools
Target Optimization
Pass@k targets
Environment Types
Software-generated RL environments

How it works

  1. Define Environment and Tasks: Users define the tasks and enterprise environments (like Jira or CRM) they need to simulate for model training.
  2. Generate Simulation Scenarios: WorldSim generates diverse, configurable scenarios at scale using realistic business data and hundreds of MCP tools.
  3. Train and Evaluate Models: Models interact with the environment, receiving calibrated reward signals and expert-graded preference labels to optimize performance.

FAQ

What is WorldSim?

WorldSim is a simulation platform that recreates enterprise environments like GitLab and Jira using realistic business data to generate diverse tasks for AI agents.

What domains does Horizon cover?

Horizon supports autonomous AI research, agent coding, multimodal knowledge work, voice-native agents, computer use (GUI-based tasks), and cybersecurity scenarios.

Added by AntSpace from public sources (labelbox.com).

Related

More AI models by Labelbox