Original Research

Have You Tried Turning It Off and On Again? A Transformer-Based Foundation Model That Achieves State-of-the-Art Helpdesk Performance by Suggesting Exactly That

Priya N. Kowalski¹, Devon O. Ferreira^1,2

¹Institute for Computational Shrugging, Tier-3 Support Wing
²Center for Plausible Deniability, Remote

Proceedings of the Society for Obvious Solutions. 2026;8(2):112-129.

PMID: 81120426 · DOI: 10.1069/satire.2026.0011

Received 2026 Jan 9 · Accepted 2026 Feb 14 · Published 2026 Feb 20

Abstract

Background: Large language models excel at many tasks, yet helpdesk resolution remains stubbornly hard. We hypothesized that most of the benchmark could be solved with a single, time-honored instruction.

Methods: We pre-trained REBOOT-1, a 70-billion-parameter transformer, on 14 years of support-ticket transcripts. At inference the model is constrained to two responses. Evaluation was conducted on the HELPME-bench suite (n = 40,000 tickets).

Results: REBOOT-1 resolved 61% of tickets by replying 'Have you tried turning it off and on again?' and a further 22% with 'It works on my machine.' Weights are released openly, behind a login that reliably times out. Latency was dominated by the model pausing for effect.

Conclusion: State-of-the-art performance is achievable without understanding the problem, consistent with the field's longstanding practice. Code and a 404 page are available.

Keywords: Machine Learning; Foundation Models; Helpdesk; Benchmarking.

Figures

Figure 1.Resolution breakdown on HELPME-bench: 'turn it off and on again' (61%), 'it works on my machine' (22%), unresolved (17%, designated future work).

Limitations

REBOOT-1 fails catastrophically on tickets where the device was already off, where it was already on, or where the user is correct. These cases together account for the remaining 17% of the benchmark and are designated future work.

Conflict of Interest Statement: The authors have previously been on the receiving end of this advice and remain bitter.
Funding: Supported by an internal grant that was approved by replying-all to the wrong distribution list.
Ethics Statement: No humans were helped during this study. IRB approval was pending and remains so.

Cite

Kowalski PN, Ferreira DO. (2026). Have You Tried Turning It Off and On Again? A Transformer-Based Foundation Model That Achieves State-of-the-Art Helpdesk Performance by Suggesting Exactly That. Proceedings of the Society for Obvious Solutions. 2026;8(2):112-129. https://doi.org/10.1069/satire.2026.0011

@article{kowalski2026,
  title   = {Have You Tried Turning It Off and On Again? A Transformer-Based Foundation Model That Achieves State-of-the-Art Helpdesk Performance by Suggesting Exactly That},
  author  = {Priya N. Kowalski and Devon O. Ferreira},
  journal = {Proceedings of the Society for Obvious Solutions},
  year    = {2026},
  volume  = {8},
  number  = {2},
  pages   = {112-129},
  doi     = {10.1069/satire.2026.0011},
  note    = {Retrieved from PubMedLie. This study is fictional.}
}

References

Kowalski PN. On the universality of power-cycling. Proc Soc Obvious Solut. 2025;7(1):1-9.
Ferreira DO. 'It works on my machine': a reproducibility crisis. J Excuses Comput. 2024;3(2):45-52.
Anon. The login that times out: a design pattern. UX Antipatterns. 2025;2(4):88-90.
Kowalski PN, Ferreira DO. Pausing for effect improves perceived intelligence. Latency Lett. 2025;1(1):3-4.
Support Team. Ticket #44032: still happening. Internal Wiki. 2026.
Anon. Have you tried a different browser? J Deflection. 2024;9(1):1-2.
Ferreira DO. Closing tickets as 'cannot reproduce': a meta-analysis. Annu Rev Shrug. 2025;5:200-219.
Kowalski PN. Open weights, closed door. Proc Soc Obvious Solut. 2026;8(1):10-22.
Anon. Did you save your work first? Pre-Reboot Stud. 2023;1(1):1-1.
Ferreira DO. The user was the problem (n = 40,000). Postmortem Q. 2026;1(2):e0002.
Kowalski PN. Escalation as a service. J Tier-4 Dreams. 2025;4(4):e0044.
Anon. Have you tried turning it off and on again? (foundational text). Folklore Comput. 1986;1(1):1-1.
PubMedLie Editorial Office. Disclaimer: this is satire.
Ferreira DO, Kowalski PN. Benchmarks that benchmark nothing. ML Theatre. 2025;3(3):e0033.
Anon. README: see other README. Docs. 2026.