Commonsense tasks show weaker but uniform anchoring on LLaMA (S ≈ −2.15)

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

• Commonsense reasoning shows uniform but weaker anchoring (S ≈ −2.15)finding0.863
  Task-specific comparison.
• Math and code tasks show strongest mid-layer anchoring on LLaMA (S ≈ −1.65 at layers 8-12)finding0.811
  Task-specific E3 finding showing compositional reasoning requires deeper processing
• Commonsense reasoning tasks S≈-2.15finding0.809
  Lower, more uniform anchoring for commonsense tasks
• Llama 3.3 70B is the most likely to take on a non-Assistant persona when steered, with even split between human and nonhuman portrayalsfinding0.773
  Model-specific difference in persona susceptibility
• Llama 3.1 405B shows 14% compliance gap in minimal helpful-only replication but smaller Llama and Mistral models show no gapfinding0.771
  Replication across open-weight models supports scale-emergence finding
• Linear steering on Llama-3.1 8B for the days-of-week task cuts across the behavior manifold, producing noisy off-target effects where predicted tokens are not even days of the week.finding0.770
  Empirical result demonstrating the failure mode of linear steering when concept geometry is cyclic.
• Llama-3.3-70B exhibits internal consistency-checking mechanisms that operate during inferenceclaim0.770
  Central interpretive claim of the paper supported by causal ablation and activation evidence
• Does Llama compute modular addition or base-10 addition for cyclic tasks?question0.769
  The specific computational question the paper resolves empirically