More compute, more capability: Why AI agent evaluations need to account for test-time compute | AISI Work
Standard evaluations cap how much compute AI agents can use. We show that raising those caps changes measured capability, the difficulty of tasks agents can solve, and how fast the frontier appears to move.
As AI agents are given more autonomy and harder tasks, underestimating their capabilities becomes more consequential. Most agent evaluations still reduce capability to a single number: a benchmark score, a pass/fail, or the length of task an agent can finish. That number hides a key design choice: how much compute the agent is allowed to spend before stopping. In March 2026, AISI published results suggesting that modest compute limits understate model capability on cyber tasks, and that the same limits may now miss substantial gains. Our Science of Evaluation team has since run frontier…
saved by
related reading
- Noam Brown on X: "Implications of Large-Scale Test-Time Compute" / Xx.com
- Measuring AI Ability to Complete Long Tasks - METRmetr.org
- Demystifying evals for AI agents \ Anthropicanthropic.com
- Measuring AI Ability to Complete Long Software Tasksarxiv.org
- Metrics of Agent Ability - METRmetr.org
- Quantifying infrastructure noise in agentic coding evals \ Anthropicanthropic.com
- Does AI Progress Have a Speed Limit?—Asteriskasteriskmag.com
- Incident Report: unsanctioned agent behaviour during cyber testing | AISI Workaisi.gov.uk
- An Apple-Picking Model of AI R&D | Tom Cunningham – Tom Cunninghamtecunningham.github.io
- AI progress is about to speed up | Epoch AIepoch.ai
- [2603.11214] Measuring AI Agents' Progress on Multi-Step Cyber Attack Scenariosarxiv.org
- What will be left for us to work on?normaltech.ai