Athanor — a 104-hour campaign post-trains a small model to official Instruct level
| Date | 2026-10-19 |
|---|---|
| Category | Benchmarks |
| Judgement | Neutral |
| Gauge weight | 2 / 3 |
| Source | Athanor (arXiv 2610.21055, submitted 2026-10-15), recorded in AGI watch evening run, 2026-10-19 |
Summary
Four Opus 5.5 agents ran a 104-hour campaign without a teacher model or research advice, lifting Qwen3-4B-Base from 15.9 to 67.3 on the development set and from 9.2 to 48.2 on holdout. The sample was n=1, however, and scaffolding such as frozen evaluation was provided.
A/B Impact Analysis
By duration alone this crosses a long-autonomy threshold, but the record clearly discounts it as execution in a cheap-to-verify domain with an external oracle and scaffolding. For Scenario A it is directional evidence the execution loop is closing; for Scenario B it is simultaneous evidence the judgment loop is still open. It stays neutral, with replication beyond n=1 as the item that would decide the next call. It is not recorded as a crux pass.