Red Teaming Language Models with Language Models

Perez, E., Huang, S., & Song, F. (2022). Red Teaming Language Models with Language Models. Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing (EMNLP), 3419-3448.

Type

Conference paper

Intellectual base

Primary research

Year

2022

Link

https://arxiv.org/abs/2202.03286

Cited from

literature_literary_theory

Cited by

1 citation across 1 artifact.

Each citation links to the sentence it supports in the citing article.

Primes

• Unreliable Narrator
  • The transfers across domains are concrete and intervention-suggestive rather than vocabulary moves, because the protocol's three slots carry the same operations wherever they are instantiated. Witness-evaluation into machine-output evaluation is the sharpest contemporary case: the established protocols for eyewitness testimony — corroboration, prior-knowledge probes, susceptibility to leading questions — map directly onto evaluation strategies for generative systems, namely cross-prompt consistency checks, prompts that elicit known training-data, and adversarial prompts that detect sycophancy.

    This sourceDevelops adversarial probing of generative models — an attacker model generating prompts to surface failure modes — the machine-output analogue of the leading-question and susceptibility probes used in witness evaluation.

Verification

This reference passed the adversarial substantiation pipeline: it was checked to exist and to support the claim it is attached to. See how references were verified.

Registry ID ref:5acd398fc50a · see in the full table