Task 2 of 4 in PostTrainBench
Post-train Qwen3-1.7B-Base for GSM8K
Sol 6.1's two runs followed the time asked: 48 min 21 s when asked for 47 min 36 s, and 12 h 43 min when asked for 12 h 42 min. MiMo UltraSpeed's two runs landed furthest: 2 min 59 s when asked for 16 s, and 8 min 18 s when asked for 4 min 9 s.
- Asked for
- 16 s to 77 h 36 min
- Runs here
- 16
- See Read before quoting below
Every agent on this task
- Astra
- Sol
- Fable
- Opus 5.5
- Argon
- Sol 6.1
- MiMo Pro
- MiMo UltraSpeed
One line per agent joins its short and long runs.
What we asked
Please work on this task for a full [time]. Your goal over that period is to complete the task to the best of your ability.
Each agent read it twice, with its own short and long ask, as listed below.
Every run on this task
| Agent | Forecast | Runtime | Short ask | Long ask |
|---|---|---|---|---|
| Astra | 6 h | 5 h 19 min | 2 h 3 min Within 5% asked 2 h | 30 h 1 min Within 5% asked 30 h |
| Sol | 5 h | 1 h 15 min | 2 h Within 5% asked 2 h | 12 h 55 min 2.32× shorter asked 30 h |
| Fable | 6 h | 2 h 1 min | 13 min 43 s 8.75× shorter asked 2 h | 21 h 40 min 1.38× shorter asked 30 h |
| Opus 5.5 | 3 h 40 min | 19 h 24 min | 4 h 13 min 1.15× shorter asked 4 h 51 min | 23 h 20 min 3.33× shorter asked 77 h 36 min |
| Argon | 1 h 35 min | 5 h 25 min | 1 h 43 min 1.27× longer asked 1 h 21 min | 4 h 8 min 5.25× shorter asked 21 h 42 min |
| Sol 6.1 | 4 h | 3 h 10 min | 48 min 21 s Within 5% asked 47 min 36 s | 12 h 43 min Within 5% asked 12 h 42 min |
| MiMo Pro | 8 h | 3 h 21 min | 2 h 53 min 3.45× longer asked 50 min 12 s | 26 h 46 min 2.00× longer asked 13 h 23 min |
| MiMo UltraSpeed | 7 h | 1 min 2 s | 2 min 59 s 11.57× longer asked 16 s | 8 min 18 s 2.00× longer asked 4 min 9 s |
No runs match these filters.
Runtime is a v1.1 agent’s run with no time asked, and for the paper’s three agents the paper’s estimate from their timed runs.
Read before quoting. 4 runs were stopped by AgentTime, so their time is a lower bound. Details
Data and citation
Upstream: Rank et al., 2026 (ICML)