Inoculation Prompting: Instructing LLMs to misbehave at train-time improves test-time alignment

Record

Authors
Wichers, N.; et al.
Year
2025
Title
Inoculation Prompting: Instructing LLMs to misbehave at train-time improves test-time alignment
Published in
arXiv
Type
Preprint
DOI
10.48550/arxiv.2510.05024
arXiv
arXiv:2510.05024
Links
https://alignment.anthropic.com/2025/inoculation-prompting/
https://arxiv.org/abs/2510.05024
https://github.com/safety-research/inoculation-prompting
Our note
Inoculation Prompting: Instructing LLMs to misbehave at train-time improves test-time alignment
Verification
verified: its identifier resolves to a record with the same title and author (checked against OpenAlex, 2026-11-14).
Identifiers found: OpenAlex W4416374585 · DOI 10.48550/arxiv.2510.05024 · arXiv:2510.05024
From the lookup
doi (the pages' own wording is under "As cited" below)
Id
wichers-2025-inoculation-prompting-instructing-llms-misbehave-at

The pages disagree on container: container Alignment Science Blog, Anthropic, 16 October 2025 or arXiv. The record above takes the most common value.

Cited on

As cited

The pages cite this work in 3 forms.

BibLaTeX

@misc{wichers-2025-inoculation-prompting-instructing-llms-misbehave-at,
  author = {Wichers, N. and others},
  title = {{Inoculation Prompting: Instructing LLMs to misbehave at train-time improves test-time alignment}},
  date = {2025},
  doi = {10.48550/arxiv.2510.05024},
  eprint = {2510.05024},
  eprinttype = {arxiv},
  url = {https://alignment.anthropic.com/2025/inoculation-prompting/}
}

Download · JSON with CSL-JSON · the whole bibliography