Some updates to Spiral Bench:
- A more detailed rubric for protective vs delusion-reinforcing behaviours
- Responses evaluated by a judge ensemble: sonnet-4.5, gpt-5 & kimi-k2
- New models evaluated: qwen3-235b, glm-4.6, grok-4-fast, mistral-medium-3.1
Introducing ComputerRL, a framework for autonomous desktop intelligence that enables agents to operate complex digital workspaces skillfully.
arxiv.org/abs/2508.14040
ComputerRL features the API-GUI paradigm, which unifies programmatic API calls and direct GUI interaction to address the inherent mismatch between machine agents and human-centric desktop environments.