The open arena for post‑training

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.

1. Contribute. A self-contained task package: task.md plus an environment, verifier, and oracle.
2. Post-train. We run a Qwen3-8B model through a managed pipeline on your environment — no GPUs or API keys from you.
3. Score. We evaluate the trained model on IndexBench, a private held-out suite where no single domain exceeds 20% of the tasks.
4. Rank. By held-out generalization over a baseline, on tasks your environment can’t have memorized.

Built by the team behind SkillsBench — the most-cited new agentic, diverse-domain benchmark of 2026 — and the CAIS Agent Skills workshop.

SkillsBench drew roughly 100 citations in four months; the 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.

• 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
• Yan LiTexas A&M
• Manling LiNorthwestern
• Yue ZhaoUSC
• Han-chung LeeMoody's
• Dawn SongUC Berkeley

A submission is a self-contained task package: a task.md describing the goal and sandbox limits, an environment Docker image that runs it, a verifier that scores attempts, and an oracle that proves the task is solvable. Everything accepted is released openly.

What a submission contains

task.md

YAML frontmatter for limits and metadata; Markdown body for the prompt, with optional multi-scene / multi-role structure.

environment/

Sealed Dockerfile and any seed data. Built on a shared base image so a task only adds what is task-specific.

verifier/

Scoring logic that runs at the end of every trial and emits a numeric reward plus side info.

oracle/

A reference solution that achieves a passing reward, so reviewers can confirm the task is solvable.

How it works

1. 1Read the spec. The full task.md reference covers the frontmatter schema, body sections, and the validation contract.
2. 2Validate locally. Run bench tasks check <dir> until the environment builds, the verifier runs, and the oracle scores a pass.
3. 3Open a pull request. Submit against the posttrain-arena repo. We review, post-train a Qwen3-8B model on accepted tasks, and score the result on IndexBench.

Eight under-served domains

• Sciences
• Industrial & Energy Operations
• Cybersecurity
• Finance & Economics
• Office & Knowledge Work
• Media & Multimodal Content
• AI/ML & Agentic Systems
• Software Engineering

1. Phase 0Warm-upLate Jun 2026
2. Phase 1Full submissionsSep 1–Oct 21
3. AwardsWinners announcedNov 7
4. WorkshopNeurIPS showcaseDec 2026