27B open-source model beats GPT-5.5 at computer tasks via evolved interfaces, at 1.5% the cost
alex_verem · x · 2026-10-23
Simular researchers introduce Agent S4, a computer-use agent that evolves its own interfaces instead of using screens built for humans. A meta-agent mines task trajectories (without success labels) to build and revise an interface/knowledge repository, with population-based evolution selecting changes that generalize to held-out validation tasks. Paired with Qwen3.8-27B, Agent S4 hits 21.3% on OSWorld V2 — a new open-source SOTA, 3.8x the same model with fixed interfaces, and above GPT-5.5 (13.0%) and Claude Opus 4.8 (20.6%) at roughly 4.3% and 1.5% of their costs. On ScienceBoard, evolved interfaces lift success from 41.4% to 58.6%. No model retraining required — interface design itself becomes an agent capability.
Related event: Agent S4: 27B open model beats GPT-5.5 via self-evolving interfaces(2 posts)→
More from coding & agent
- Carmack: Give every stack a full virtual address space on 64-bit and let it do memory management — ID_AA_Carmack · 2026-10-24
- TokenFuel works as an AI gateway: one OpenRouter key caps spend at funded amount — RichardsonDx · 2026-10-24
- Kimi K3.1 Pro goes live on Kimi Code and web, positioned as flagship model — cedric_chee · 2026-10-24
- GitHub hosts 'Clawfather' steipete to stress-test Surface Laptop Ultra ahead of GitHub Universe — pavandavuluri · 2026-10-24
- Anthropic's Claude Duet personal agent runs on Opus 5.5 and works with the laptop closed — altryne · 2026-10-24
- Nous Research's Hermes Agent processed 53.3T tokens last month, tops open-source agent harnesses — NousResearch · 2026-10-24