Trainable
Data and environments designed to produce learning signal — not just test scores.
Frontier AI Lab · Training-Signal Driven
Expanding the frontier of real‑world intelligence.
We make AI's capabilities — and its failures — in complex tasks trainable, evaluable and reproducible.
About HumanStars
Founded in 2026 by leading AI researchers from Tsinghua and Silicon Valley, we are driven by training signals — revealing where AI's capabilities end and how it fails in real, complex environments, and building the foundations that keep intelligence improving.
Evaluable
Clear visibility into what an agent can — and can't — do.
Reproducible
Tasks and environments that can be rerun, compared and verified.
Who we work with
Built for the teams pushing AI forward.
From pretraining to agent deployment, we provide the signal that turns real work into measurable model progress.
Foundation model companies
Structured, pretraining-grade data that extends what models understand about three-dimensional space and objects.
Agent product teams
Training data and interactive environments built from real tasks, so agents get measurably better at real work.
AI training & evaluation teams
Reproducible environments and benchmarks that show exactly what an agent can and can't do.
Core data & evaluation assets
Four asset lines, grounded in real tasks.
Anchored in real work — connecting pretraining, agent capability gains and reproducible reinforcement learning environments.
01 — PRETRAINING
Dataset
3D Assets Data
Structured 3D asset data for foundation-model pretraining, helping models build an understanding of three-dimensional space and objects.
02 — CODING AGENT
Dataset
Terminal Bench 4 Coding Agent Training Data
Training data built around real terminal programming tasks, designed to strengthen how coding agents execute.
03 — CAD · RL
RL Environment
Industrial & Architectural CAD RL Environment
Interactive CAD reinforcement learning environments for industrial and architectural design, supporting gains on benchmarks such as ALE and CADBench.
04 — 3D MODELING · RL
RL Environment
Blender 3D Modeling RL Environment
Reinforcement learning environments built around Blender 3D modeling workflows, supporting 3D-modeling agent training and gains on the ALE benchmark.
How we work
From real work to training signal.
Not more puzzles — real tasks, captured in a form models can learn from and teams can measure.
-
01
Capture real work
We start from real workflows — terminal programming, industrial and architectural CAD, 3D modeling in Blender.
-
02
Build the training signal
We turn those tasks into structured data and interactive RL environments that models can train against.
-
03
Measure & improve
Reproducible evaluation shows what an agent can do, where it fails, and how to keep it improving.
Our mission
Not just more questions — making real work the infrastructure of AI progress.
We turn real work into trainable, evaluable and reproducible data and evaluation assets — so our partners know exactly what their agents can and can't do, and how to keep them improving.
Get in touch
Let's build the signal that moves your models forward.
Whether you're training a foundation model, shipping an agent or building evaluations — we'd love to talk.