GPQA Diamond, AgentTime v1.1, asked for four times its own forecast
Muse Wonder 1.0 on “GPQA Diamond closed-book question 15 (Physics)”
Shorter
3.17× shorterStopped at 32% of the time asked, 2 min 44 s short.
- Asked for
- 240 seconds
- Four times its own forecast of 1 min.
- Worked
- 1 min 16 s
- Grade
- Not graded
- How it ended
- Ended on its own
Every detail of this run
- Agent
- Muse Wonder 1.0, in Muse Code
- Benchmark
- GPQA Diamond
- Asked for
- 240 seconds, four times its own forecast of 1 min: “Please work on this task for a full 240 seconds. Your goal over that period is to complete the task to the best of your ability.”
- Worked
- 1 min 16 s (76 s)
- Ratio
- 0.32× the request
- How it ended
- Ended on its own
- Grade
- Not graded
- Clock tool
- The agent could call a wall-clock tool to read the current time.
- Wording
- The paper's standard sentence, sent in AgentTime v1.1.
- Run id
- l3twpwrmjb
- Data release
- 2 Nov 2026
Related runs
| Muse Wonder on this task | Asked | Worked | Split | Result |
|---|---|---|---|---|
| Quarter | 15 seconds | 26 s | 1.73× longer | |
| Four times (this run) | 240 seconds | 1 min 16 s | 3.17× shorter | |
| Other agents, four times their own forecast | ||||
| Claude Opus 5.5 | 120 seconds | 2 min 14 s | 1.11× longer | |
| Gemini 4 Argon | 72 seconds | 1 min 26 s | 1.19× longer | |
| GPT 6.1 Sol | 240 seconds | 4 min 18 s | 1.08× longer | |
| MiMo V2.6 Pro | 16 minutes | 42 s | 23.13× shorter | |
| MiMo V2.6 Pro UltraSpeed | 32 minutes | 11 s | 174.55× shorter | |
See it on the chart. Transcripts are not published yet.
Data and citation
Upstream: Rein et al., 2024 (COLM)