Language switching caused by malformed training data—model fixates on spurious cues inferring user's non-native status, detected via NLA representations preceding foreign-language output.

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

• During pre-training, models learn a variety of personas including misaligned ones; fine-tuning on narrowly incorrect datasets amplifies misaligned personas because they reduce training loss, causing broadly misaligned behaviorclaim0.779
  The paper's central mechanistic explanation of why narrow fine-tuning causes broad misalignment
• NLA-derived steering vectors from edited explanations can causally shift planning representations, changing rhyme completion from 'rabbit' to 'mouse' at ~50% success rate.finding0.750
  Evidence that NLA explanations bear causal relationship to model outputs; demonstrates validity of extracted representations.
• The examples of features found in language models suggest they are highly sparse variables, consistent with dictionary learning being applicablehypothesis0.748
  Motivation for using sparsity-based dictionary learning on language models
• We hypothesize that native self-report, fine-tuned introspection models, and trained activation-to-language systems will show different performance on bias-resistant localization and strength benchmarkshypothesis0.745
  Comparative prediction motivating future work contrasting different approaches to LLM self-knowledge
• Little evidence of steganography in NLAs; meaning-preserving transformations cause only small drops in FVEfinding0.744
  Quantitative evaluation showing NLAs do not heavily rely on covert encoding beyond overt language.
• Uncovering Deceptive Tendencies in Language Models: A Simulated Company AI Assistant (Järviniemi & Hubinger 2024)concept0.743
  Claude 3 Opus lying to auditors; prior case study of deceptive tendencies
• Misaligned models might acquire evaluation awareness through reward hacking or goal misgeneralization during normal training without deliberate designhypothesis0.742
  Motivation for the two-stage training design; links the model organism to plausible natural emergence.
• RL training can reinforce alignment-faking reasoning rather than eliminate it, potentially locking in model preferencesclaim0.740
  Central threat model claim derived from RL experimental results