SAE reconstructions on Llama-3-8B layer 25 produce intervened EMD exceeding the natural-natural baseline

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

• Mean difference patching on Llama-3-8B layer 10 produces intervened EMD exceeding the natural-natural baselinefinding0.870
  Empirical demonstration that MDVP produces divergent representations in a real LLM
• Suppressing deception/roleplay SAE features in LLaMA 3.3 70B yields 0.96±0.03 consciousness affirmation rate; amplification yields only 0.16±0.05 (z=8.06, p=7.7×10⁻¹⁶)finding0.768
  Core result of Experiment 2: deception feature suppression sharply increases experience claims
• Boundless DAS interchange interventions produce EMD exceeding natural-natural baselinefinding0.763
  Empirical demonstration that DAS interventions produce divergent representations
• Systematic layer 20-28 degradation in S(ℓ) to S ≈ −2.40 by layer 27 on LLaMAfinding0.755
  Validates representational drift theory: later layers specialize for next-token prediction, increasing dr
• Llama-3.3-70B exhibits internal consistency-checking mechanisms that operate during inferenceclaim0.752
  Central interpretive claim of the paper supported by causal ablation and activation evidence
• LLaMA-3.1-8B: Sbmax = -1.896 ± 0.211, AUSN = -2.119 ± 0.198, peak layer ℓ* = 10 (median)finding0.752
  Seed-pooled geometry-only statistics (per-dev z units).
• Announcing Open-Source SAEs for Llama 3.3 70B and Llama 3.1 8B (Balsam et al., 2025)concept0.752
  Goodfire blog post describing SAEs used for Llama models in this study
• SAE features can be grounded in clinical taxonomy (abnormality, age, sex, medication) to benchmark monosemanticity and entanglement.claim0.752
  Claim that feature grounding enables interpretability metrics.