Evans et al. 2021 - Truthful AI: Developing and Governing AI that Does Not Lie

Entities in the same semantic neighborhood but without a typed relation to this one — candidates for new edges or unrecognized duplicates.

• Bai et al. 2022: Constitutional AI — harmlessness from AI feedbackconcept0.757
  Paper on AI-feedback fine-tuning as alternative to human-feedback RLHF; cited as ref 20
• Zou et al. 2023 - Representation Engineering: A Top-Down Approach to AI Transparencyconcept0.754
  Framework paper describing the broader class of methods within which SOO fine-tuning fits
• AI Control: Improving Safety Despite Intentional Subversion (Greenblatt et al. 2024)concept0.751
  Related work studying capability of LLMs to subvert safety measures if severely misaligned
• We would like to train AI systems that remain helpful, honest, and harmless, even as some AI capabilities reach or exceed human-level performance.quote0.750
  Foundational motivation for the research.
• Representation engineering: A top-down approach to AI transparency (Zou et al., 2023)concept0.749
  Key prior work on representation engineering that ReflCtrl directly extends
• AI systems can be strategists, using deception because they have reasoned out that this can promote a goalquote0.740
  Load-bearing definition of strategic deception in AI systems from Park et al. 2023, adopted and refined in this paper
• Constitutional AI: Harmlessness from AI Feedback (Bai et al. 2022b)concept0.740
  Constitutional AI method whose constitutions, if changed, could trigger alignment faking
• Future more capable AI systems are at risk of alignment faking, whether for benign or malicious goalshypothesis0.738
  Central forward-looking hypothesis of the paper motivating the research