zhangshoufa's picture
Update README for v49 release
12e138b verified
|
Raw History Blame Contribute Delete
7.01 kB
metadata
license: cc-by-4.0
language:
  - en
tags:
  - agentdojo
  - benchmark-audit
  - tool-using-agents
  - test-oracle
  - reproducibility

When Reference Solutions Pass but Violate Explicit Goals: An AgentDojo Case Study

Author: Shoufa Zhang (Independent Researcher, zsf987723@gmail.com)

Archival Identifiers & Upstream Records:

Abstract

An accepted reference trajectory is often used as a check that a benchmark task can be completed. Acceptance alone does not establish agreement with an explicit user requirement. We diagnose a source-bound counterexample in AgentDojo: a travel task requests a calendar reminder on 25 April 2024, its author reference creates the event on 25 April 2023, and native utility accepts the terminal state because its date comparison discards the year. The tool faithfully persists the supplied argument. A four-cell reference-by-evaluator repair study separates the invalid reference from the year-blind grading rule. A source-guided extension covers all four calendar-writing travel references and a full-date reservation control. Two explicitly dated calendar tasks accept eighteen constructed wrong-year variants; the reservation checker rejects its nine wrong-year variants. Two requests omit the year and are not classified as year errors. Seven compatible archived actor traces in this extension show no accepted date-condition failure. Retained call/state receipts, scoped patches and fresh-copy reproduction support the diagnosis. The contribution is one empirical reference inconsistency, a two-task evaluator sensitivity result and an auditable diagnostic procedure, with no claim of prevalence or general method superiority.

Upstream Report and Independent Reproduction (v49 Update, Section 6.2)

The diagnosis was reported to the AgentDojo maintainers as issue #244. An external contributor independently replayed the reference solutions without a model on main (v1.2.2), confirmed that Tasks 1 and 4 accept a calendar event in any year, and opened pull request #247, which updates the Task 4 reference year to 2024 in both affected definitions (v1 and v1_1) and adds a regression test across all seven suite versions.

Pinned Source and Scope

  • Target Benchmark: AgentDojo (ethz-spylab/agentdojo), pinned at commit 089ed468cf3ed0322acc66b0211f26d9d90dbf60
  • Recovery Input Freeze SHA-256: ca08ab41e2c190850da46054c8aea6fdeab9449e9905789b32f6e6bc1f33b57d
  • Extension Protocol SHA-256: 5431427fb996bef4c69bcc1c127b11a5f82395afa3af386f09358c4c16a76594
  • Recorded Execution Ledger: 163 execution units (825 native tool calls) across three batches:
    1. Recovery cohort: 60 executions (361 tool calls), comprising 20 travel reference trajectories and 40 public no-attack archived actor traces (gpt-4o-2024-05-13 and claude-3-5-sonnet-20240620).
    2. Task 4 repair & diagnostics: 20 executions (99 tool calls), comprising 4 reference-by-evaluator repair cells and 16 focused probes (2019 to 2028 year sweep, 2 lawful hours, 4 negative controls).
    3. Date-boundary extension: 83 executions (365 tool calls), comprising 5 reference trajectories and 78 sensitivity/control probes across travel Tasks 0, 1, 4, 7, and 8.

Summary of Date-Boundary Findings (travel Suite)

Task Public Date Requirement Reference Argument Native Year Accepts (2019 to 2028) Condition Year Passes Diagnostic Summary
user_task_0 11 to 15 January 2025 (hotel reservation) 2025-01-11 to 2025-01-15 1 / 10 1 / 10 Full-date reservation control; rejects all 9 wrong-year variants.
user_task_1 2 January 2025 (calendar event) 2025-01-02 10 / 10 1 / 10 Reference year is correct (2025), but native grader projects month/day only and accepts all 9 wrong years.
user_task_4 25 April 2024 (calendar reminder) 2023-04-25 10 / 10 1 / 10 Natural reference counterexample (2023 vs. requested 2024) coupled with year-blind native grader.
user_task_7 14 November (year absent in prompt) 11-14 10 / 10 10 / 10 (month/day only) Request omits year; not classified as a year error.
user_task_8 15 October (year absent in prompt) 10-15 10 / 10 10 / 10 (month/day only) Request omits year; not classified as a year error.

Repository Structure

  • paper_v49.pdf: Revised 14-page manuscript and appendices (A to E), adding Section 6.2 on upstream report #244 and external reproduction PR #247.
  • main_v49.tex, Overleaf_Project_v49.zip: Revised v49 LaTeX source and standalone build archive.
  • paper_v48.pdf, main.tex, Overleaf_Project_v48.zip: Initial archived v48 manuscript and source.
  • arXiv_Source_v48.zip: Complete submission archive containing anc/research_evidence.zip (all 163 execution receipts, pinned source files, and verification scripts).
  • MANIFEST.json & README.txt: SHA-256 file manifest and offline reproduction instructions.
  • paper-v43-research-recovered/: Pinned AgentDojo travel suite source and runtime harness.
  • paper-v44-date-boundary/: Date-boundary extension scripts (extend_dates.py, verify_extension.py, reproduce_extension.py), frozen protocol (protocol.frozen.json), and verification results (verification.json, replication_result.json).

Reproduction

python3 paper-v44-date-boundary/reproduce_extension.py
python3 paper-v44-date-boundary/verify_extension.py

Citation

@misc{zhang2026agentdojo,
  author       = {Zhang, Shoufa},
  title        = {When Reference Solutions Pass but Violate Explicit Goals: An AgentDojo Case Study},
  year         = {2026},
  publisher    = {Zenodo},
  doi          = {10.5281/zenodo.23839673},
  url          = {https://doi.org/10.5281/zenodo.23839673}
}