Vmax
CommercialVmax is a San Francisco RL company and applied research lab, founded in 2025 by RL PhDs Augustine Mavor-Parker and Matthew Sargent and backed by South Park Commons and Race Capital. It aims to automate reinforcement learning by turning proprietary data and evals into new environments, with long-horizon agents as the target. Its 2026 research is about generating RL tasks and environments automatically: PopuLoRA (self-play curricula), unix-ctf (procedural shell CTF environments) and PROPEL (probe-guided task generation, with a Goodfire AI co-author). Earlier, with Martian, it released 1k Harbor-format JavaScript tasks.
What's new
- 10 Jun 2026research
Vmax publishes PROPEL: training task generators at the learnable frontier, co-authored with a Goodfire AI researcher
PROPEL trains task generators with RL. Instead of running the solver repeatedly in the loop, it uses a lightweight activation probe on a frozen generator to predict how the solver will do. Vmax reports that the share of generated tasks at the target difficulty roughly doubles. Code: 10.1% to 20.0% with a 3B solver and 5.3% to 12.6% with a 7B solver. SWE: 9.8% to 19.6% with a 27B agent on 11 unseen repositories. arXiv:2606.18284, submitted 2026-06-10. Six authors are Vmax; Connor Watts is listed as Goodfire AI.
Why it matters: This is Vmax's core technical bet: generating tasks and environments automatically at the right difficulty, which is the bottleneck in producing RL environments at scale. Through the Goodfire co-author it links to interpretability, which matches Vmax's mech-interp hiring.
- 25 May 2026research
unix-ctf: procedural CTF environments for training Unix-competent shell agents
A procedural generator of capture-the-flag environments in Linux containers, for training and evaluating LM shell agents. Vmax posted it on 2026-05-25; arXiv:2605.29115 v1 followed on 2026-05-27. The abstract reports that fine-tuning Qwen3-8B raised the held-out solve rate from 11.6% to 43.6% (n=225), with 32/100 on InterCode-CTF. It argues that Unix competence is separate from programming ability and that Terminal-Bench 2.0 does not isolate it.
Why it matters: A concrete, published environment family for terminal and shell agents, aimed at the gap it identifies in Terminal-Bench-style evaluation.
- 20 May 2026research
PopuLoRA: population-based asymmetric self-play for RLVR post-training
Teacher and student LoRA adapters on a frozen base model co-evolve: teachers propose coding problems, students solve them under a programmatic verifier, and the two sub-populations cross-evaluate each other. The paper reports gains on HumanEval+, MBPP+ and LiveCodeBench, plus transfer to math benchmarks (AIME, AMC and others), which the authors call 'suggestive'. The setup is 4 teachers and 4 students at 1.31x wall-clock overhead. arXiv:2605.16727, submitted 2026-05-16; listed on vmax.ai 2026-05-20.
Why it matters: Shows Vmax's self-play and open-endedness approach: building curricula automatically instead of writing tasks by hand.
- 20 May 2026hire
Vmax posts five new roles, including mechanistic interpretability; MTS pay listed at $300k-$500k
Posted 2026-05-20: MTS Mechanistic Interpretability, MTS RL Algorithms, MTS RL Infrastructure, a Mechanistic Interpretability research fellowship, and an 'Exceptional Talent' open application. They join three roles posted 2025-11-05 (MTS Open Endedness, MTS Applied RL, Open Endedness fellowship). All 8 are in SF. All five MTS roles list $300k-$500k USD.
Why it matters: Shows the team's direction: Vmax is building a research team around RL algorithms, RL infrastructure and interpretability, not a large data-ops workforce.
Founding team
What practitioners should know
For researchers
- Three 2026 arXiv papers, all on generating tasks and curricula automatically: PopuLoRA (2605.16727, self-play), unix-ctf (2605.29115, procedural shell CTF environments) and PROPEL (2606.18284, activation-probe proxy for task difficulty).
- PROPEL has one Goodfire AI co-author (Connor Watts), which links Vmax's RL work to mechanistic interpretability. Vmax is also hiring MTS and fellows in mech interp.
- None of the three paper pages on vmax.ai links a code repo. GitHub user 'VMax-ai' shows no link to the company.
For program managers
- Small, research-led team. Six Vmax-affiliated researchers are named as authors on vmax.ai: Mavor-Parker, Sargent/Daborn-Sargent, Bradway, Wolf, Lin, Creus Castanyer. A seventh name, Connor Watts, is a Goodfire AI co-author.
- No round, amount or valuation disclosed. Race Capital and South Park Commons are reported backers (SPC portfolio page; PitchBook per the baseline).
- No frontier-lab customers are named publicly. The only public partner is Martian (ARES), which released 1k JS tasks with Vmax on 2026-01-30. That is a collaboration, not a confirmed paying customer.
- 8 open roles on Ashby, all in SF; MTS pay $300k-$500k. A research-lab hiring profile.
For data ops
- Tasks released jointly with Martian use the Harbor format (per Martian's ARES post).
- No public docs, API, security/trust page or SOC 2 statement found.
Benchmarks & research
Products
Scale & velocity
Research depth
Capital
Security & compliance
Product
Buyer analysis
How we verified this
Related vendors
Sources
- vmax.ai/ · 2026-10-05, Official site. Meta description ('automating reinforcement learning... proprietary data and evals into new sets of environments'); nav to Research, Careers (Ashby), Sign in.
- vmax.ai/research · 2026-10-05, Research list: PROPEL (2026-06-10), unix-ctf (2026-05-25), PopuLoRA (2026-05-20), with authors.
- vmax.ai/team · 2026-10-05, Lists 7 author names, no titles. One of them, Connor Watts, is Goodfire AI per the PROPEL page.
- vmax.ai/team/propel · 2026-10-05, PROPEL page: author/affiliation list (Watts = Goodfire AI, the rest Vmax) and headline numbers.
- arxiv.org/abs/2606.18284 · 2026-10-05, PROPEL arXiv abstract, submitted 2026-06-10.
- vmax.ai/team/populora-co-evolving-llm-populations-for-reasoning-self-pla · 2026-10-05, PopuLoRA page: 4+4 adapters, 1.31x overhead.
- arxiv.org/abs/2605.16727 · 2026-10-05, PopuLoRA arXiv, submitted 2026-05-16.
- vmax.ai/team/unix-ctf-procedural-environments-for-unix-competence-reinfo · 2026-10-05, unix-ctf page, dated 2026-05-25.
- arxiv.org/abs/2605.29115 · 2026-10-05, unix-ctf arXiv, v1 2026-05-27. Abstract: Qwen3-8B 11.6% to 43.6%; 32/100 InterCode-CTF.
- api.ashbyhq.com/posting-api/job-board/vmax · 2026-10-05, 8 open SF roles, posting dates, $300k-$500k MTS pay, company self-description.
- www.southparkcommons.com/companies/vmax/ · 2026-10-05, Two founders, bios, headshot images (both load as image/webp, single-person photos), founded 2025, SF.
- withmartian.com/post/ares-open-source-infrastructure-for-online-rl-on-co · 2026-10-05, 2026-01-30 post: 1k Harbor-format JS tasks released in partnership with Vmax.
- www.grasp.upenn.edu/people/heejin-jeong/ · 2026-10-05, Heejin Jeong listed as Staff Research Engineer at Waymo; no Vmax mention. Undercuts the baseline's third-co-founder claim.
- www.goodfire.com/research · 2026-10-05, Goodfire research index has no mention of Vmax or PROPEL. The collaboration is attested only on the Vmax and arXiv pages.
- github.com/vmax-ai · 2026-10-05, Name collision: GitHub user with a Deep-Live-Cam fork, no link to vmax.ai.
- job-boards.greenhouse.io/vmax · 2026-06-07, Careers, 8 open roles, all San Francisco
- www.linkedin.com/company/vmax-ai · 2026-06-07, Public LinkedIn snippet, ~11 employees, San Francisco (Brannan St 94107), 'Open ended task generation', named team members
- www.linkedin.com/in/augustine-mavor-parker/ · 2026-06-07, Co-founder/CTO; RL PhD UCL; ex-Redwood Research, CSHL, Illumina
- www.linkedin.com/in/matthewjsargent/ · 2026-06-07, Co-founder; RL PhD UCL
- newsletter.semianalysis.com/p/rl-environments-and-rl-for-science · 2026-06-07, SemiAnalysis, lists Vmax among RL-environment vendors
- pitchbook.com/profiles/company/907262-38 · 2026-06-07, PitchBook profile (403 blocked); no disclosed funding amount per search snippets
- www.crunchbase.com/organization/vmax · 2026-06-07, Crunchbase listing referenced; funding unannounced
- x.com/MavorParker/status/1868009967518880183 · 2026-06-07, Founder post describing agents leveraging large-company dataset structure for multistep RL on long-horizon tasks