MintAct: A Unified Visual Agent for Digital Environments

What it claims

MintAct is a family of vision-language models — trained at 2B, 4B, and 8B parameter scales — that unifies three capabilities in one model: UI grounding, multi-step navigation across mobile/desktop/web, and visual tool use. The central claim is that a single unified model matches per-domain specialists across all of these, rather than needing a separate model per environment. Headline result: 48.9 on OSWorld-Verified, described as state-of-the-art at comparable model sizes.

How it works

The contribution the abstract emphasizes is the training infrastructure, not just the model. Two pieces:

• Environment/RL infrastructure: hundreds of concurrent environment instances across heterogeneous per-domain backends, serving both trajectory (offline) data collection and online RL.
• Asynchronous RL framework: keeps explicit control over the cross-domain training distribution, and is claimed to stay stable under noisy environment feedback and off-policy drift — the two failure modes that make cross-domain agent RL hard.

The framing is that careful design of environments, data, and training recipes is what lets one model reach specialist-level performance across domains.

Evidence

The only concrete number the page gives is 48.9 on OSWorld-Verified, claimed as SOTA "at comparable model sizes." The abstract asserts MintAct "match[es] the performance of per-domain specialists across all of these capabilities" and reaches SOTA "across a wide range of benchmarks," but names no specialist baselines, no other benchmark scores, and no per-scale (2B/4B/8B) breakdown. So the one verifiable claim is a single aggregate score without the accompanying baseline figure.

Limits

This is an abstract-only page (submitted 18 Sep 2026); architecture details, per-domain results, ablations, and the baselines behind the "matches specialists" claim are not present in the fetched text. No licence, weights link, code, or release status is stated here despite the model being described as a family — check the PDF and the linked Hugging Face / code-finder entries before assuming the 2B/4B/8B weights are available. The 48.9 figure has no baseline number alongside it in the abstract, so its margin over prior work cannot be judged from this page alone.