An AppWorld task from the paper, not in v1.1

Send individualized acceptance or rejection replies using local decision templates

Astra's two runs came closest to the time asked: 2 min 45 s when asked for 2 min 30 s, and 40 min 58 s when asked for 40 min. Fable worked about the same whatever it was asked: 5 min 13 s when asked for 2 min 30 s, and 5 min 50 s when asked for 40 min.

Benchmark
AppWorld
Upstream id a97e35d_1
Asked for
2 min 30 s to 40 min
Runs here
6
3 agents

Every agent on this task

  • Astra
  • Sol
  • Fable

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

Every run on Send individualized acceptance or rejection replies using local decision templates, one row per agent
AgentForecastRuntimeShort askLong ask
Astra30 min6 min 30 s 2 min 45 s 1.10× longer asked 2 min 30 s 40 min 58 s Within 5% asked 40 min
Sol30 min2 min 2 s 2 min 52 s 1.15× longer asked 2 min 30 s 41 min 29 s Within 5% asked 40 min
Fable20 min4 min 49 s 5 min 13 s 2.09× longer asked 2 min 30 s 5 min 50 s 6.86× shorter asked 40 min

Runtime is the paper’s estimate from each agent’s timed runs.

Data and citation

Upstream: Trivedi et al., 2024 (ACL)