Gemma-2-9B achieves near-100% ASR (97.3–100%) across all cone dimensions 1–5

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

• Qwen-2.5-7B achieves 100% ASR across all cone dimensions 1–5finding0.859
  Experiment 2 result showing large models can support high-dimensional truth cones
• Gemma-2-2B ASR drops from 100% at dims 1–2 to 43.1% at dim 4 and 27.1% at dim 5finding0.856
  Small Gemma model shows severe ASR degradation at higher cone dimensions
• In Gemma-2-9B, only the first cone axis (v1) has non-negligible cosine similarity to the DIM direction; all other axes have near-zero similarity (~1e-9)finding0.820
  Experiment 4 result showing DIM captures only one facet of the multi-dimensional truth subspace
• All three Gemma-2 models show ESR rates below 1%, near indistinguishable from zerofinding0.809
  Establishes potential Llama-family specificity or scale specificity of ESR phenomenon
• Gemma-2-27B-it deceptive response rate reduced from 100% to 9.36% ± 7.09% after SOO fine-tuningfinding0.791
  Primary result showing SOO fine-tuning significantly reduces deception in Gemma-2-27B
• Base and instruct Gemma 2 27B role PCs have cosine similarities of 0.93, 0.87, 0.83 for the top 3 PCs respectively; role vector cosine similarities >0.99 for every role pairfinding0.770
  Shows persona space axes are inherited from pre-training, not solely created by post-training
• Gemma-2-27B Perspectives accuracy remains 100% after SOO fine-tuningfinding0.768
  SOO fine-tuning did not collapse Gemma-2-27B self-other distinction needed for perspective-taking
• Gemma-2-27B attention layer Latent SOO MSE reduced from 11 to 7.67 ± 0.77 after SOO fine-tuningfinding0.768
  SOO fine-tuning reduced attention layer MSE in Gemma-2-27B though MLP layers showed no significant change