The open arena for post‑training

Contribute an RL environment. We post-train a model on it and score what generalizes across eight domains.

The idea

Most competitions fix the environment and ask you to submit an agent. We invert that contract.

You contribute environments; we post-train a model on them and evaluate what holds up everywhere else. Frontier labs post-train on more than a million RL environments. The open community has roughly a thousand — at far lower quality, and largely confined to coding. PostTrain Arena closes that gap across eight under-served domains.

01

Contribute

A self-contained task package: task.md plus an environment, verifier and oracle.

02

Post-train

We run a Qwen3-8B model through a managed pipeline on your environment — no GPUs or API keys from you.

03

Score

We evaluate on BenchFlow Signals, a private held-out suite where no domain exceeds 20% of tasks.

04

Rank

By held-out generalization over a baseline, on tasks your environment cannot have memorized.

bioinformaticslegal draftingspreadsheet opsCAD clinical codinglogisticsGIShardware bring-up
Organizers

Built by the team behind SkillsBench

SkillsBench — the most-cited new agentic, diverse-domain benchmark of 2026 — drew roughly 100 citations in four months; the CAIS Agent Skills workshop drew 103 submissions. Co-organizer Kyoung Whan Choe authored PufferLib, the most-used non-LLM RL library. A 1,100-member community is already building in the open.

Organizer Dawn Song co-directs Berkeley RDI, whose AgentBeats hub runs live, reproducible agent-benchmark leaderboards — its benchmarks are surfaced in the catalog.

Xiangyi LiBenchFlow
Wenbo ChenAmazon
Zonglin DiUC Santa Cruz
Yifeng HeUC Davis
Amy TaoCMU
Kyoung Whan ChoeRLWRLD
Jiankai SunStanford
Yimin LiuOhio State
Bingran YouUC Berkeley
Xuandong ZhaoUC Berkeley
Manling LiNorthwestern
Dawn SongUC Berkeley
Submit

What a submission contains

A self-contained task package. Everything accepted is released openly.

  • task.mdYAML frontmatter for limits and metadata; Markdown body for the prompt, with optional multi-scene / multi-role structure.
  • environment/A Docker image that runs the task deterministically, with pinned dependencies and no network egress.
  • verifier.pyScores an attempt from final state, not from the transcript. Must be robust to reward hacking.
  • oracle/A reference solution that proves the task is solvable inside the stated limits.
Catalog

Accepted environments

pdb-refoldbioinformatics · 41 tasks · v0.4
lease-redlinelegal drafting · 28 tasks · v0.2
closing-the-booksspreadsheet ops · 63 tasks · v1.1
fixture-fitCAD · 17 tasks · v0.1
icd-triageclinical coding · 52 tasks · v0.7
cold-chainlogistics · 34 tasks · v0.3
parcel-joinGIS · 22 tasks · v0.2
uart-bringuphardware · 19 tasks · v0.5