DoneBench

Summary

DoneBench is a benchmark for specification grounding in tool-using agents. It tests whether agents can infer, formalize, stress-test, and satisfy task-completion criteria before executing stateful tasks.

Key Facts

  • EXTRACTED: The local root is D:/Research/DoneBench.
  • EXTRACTED: The benchmark axis is Specification Grounding, not generic agent ability.
  • EXTRACTED: Each task includes a user goal, visible tool environment, initial state, policies, gold criteria, executable DoneSpec, near-miss final states, reference trace, and audit metadata.
  • EXTRACTED: Primary grading is deterministic through the DoneSpec DSL rather than LLM-as-judge.
  • EXTRACTED: The checked-in corpus has 600 tasks: 120 each for calendar, email, spreadsheet/database, CRM/workflow, and file/document operations.
  • EXTRACTED: The default split is 100 dev and 500 test tasks.
  • EXTRACTED: As of the inspected README status, a DeepSeek tool-plan full run completed 18,000 / 18,000 trials, with optional human calibration still incomplete.
  • INFERRED: DoneBench is more mature as a benchmark artifact than many active research-code projects because it has dataset, DSL, reproducibility package, paper scaffold, and audit gates.

Relationships

Open Questions

  • Which result rows are final paper-facing rows after the tool-plan executor change?
  • How much optional human calibration is needed before submission?
  • Which specification-grounding failures should become named error categories in the wiki?