A small group of hidden states (group b) over end-of-sentence punctuation tokens is highly causally implicated in truth judgments

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

• Patching group (b) hidden states (over clause-ending punctuation, early-middle layers) in LLaMA-2-13B produces the strongest causal effect on TRUE/FALSE output predictionsfinding0.842
  Localizes truth representations to specific hidden states, motivating the rest of the analysis
• Little evidence of steganography in NLAs; meaning-preserving transformations cause only small drops in FVEfinding0.769
  Quantitative evaluation showing NLAs do not heavily rely on covert encoding beyond overt language.
• We observe features related to a broad range of safety concerns, including deception, sycophancy, bias, and dangerous content.claim0.766
  SAEs uncover safety-relevant representations that might be monitored or controlled.
• In LLaMA-2-7B, PCA of larger_than+smaller_than shows statements clustering by surface-level characteristics (e.g., presence of token 'eighty') rather than truth valuefinding0.764
  Shows absence of abstract truth representations in smallest model, supporting scale-dependent emergence claim
• Patching h[1] with a divergent representation can activate distinct, hidden pathways that result in misleadingly confirmatory behavior and/or undetected behavior.quote0.760
  Load-bearing description of the core pernicious divergence mechanism illustrated in Figure 1
• Causally-masked attention in a decoder-only model has no ordered phase (Proposition 2)finding0.760
  Application to transformer language models
• Deception feature suppression yields higher truthfulness in 28 of 29 evaluable TruthfulQA categoriesfinding0.759
  Breadth of generalization of deception feature effects across independent reasoning domains in Experiment 2
• Truth-evaluation framing specifically contributes to truth geometry shifts beyond generic instruction-following prefix.claim0.757
  Supported by the neutral read-prompt changing emergence but not fully replicating ask-correct cross-task generalization.