Optimal activation capping layers for Llama 3.3 70B are layers 56-71 (out of 80) at 25th percentile cap

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

• Optimal activation capping layers for Qwen 3 32B are layers 46-53 (out of 64) at 25th percentile capfinding0.874
  Specific implementation finding for Qwen capping parameters
• LLaMA-3.1-8B: Sbmax = -1.896 ± 0.211, AUSN = -2.119 ± 0.198, peak layer ℓ* = 10 (median)finding0.809
  Seed-pooled geometry-only statistics (per-dev z units).
• Math and code tasks show strongest mid-layer anchoring on LLaMA (S ≈ −1.65 at layers 8-12)finding0.802
  Task-specific E3 finding showing compositional reasoning requires deeper processing
• The case at approximately the 2/3 layer of LLaMA3.1-8B (Layer 24, satisfying Criteria 1 and 2) aligns with prior studies showing the 2/3 layer optimally predicts human brain activity.finding0.798
  Connects this study's results to Schrimpf et al. 2021 and Caucheteux et al. 2022/2023 findings on brain-LLM alignment.
• Meta-LLaMA-3.1-8B-Instruct shows optimal anchoring at layer 9 (S ≈ −1.90, median peak layer ℓ* = 10 [IQR 0.384])finding0.789
  E3 result establishing the Goldilocks zone at mid-layers for LLaMA architecture
• Llama-3.3-70B shows multi-attempt rate of 7.4% vs. ≤1.2% for all other models testedfinding0.779
  Supporting finding showing ESR is driven by both higher multi-attempt rates and comparable improvement rates
• Some activation capping settings slightly improve performance on IFEval, MMLU Pro, or GSM8k for both Qwen and Llamafinding0.779
  Unexpected positive finding suggesting capping may sometimes help capabilities
• 25th percentile of Assistant Axis projection distribution gives the most Pareto-optimal safety-capability tradeoff for activation capping, and approximately matches mean Assistant response activationfinding0.778
  Calibration finding for choosing the activation cap threshold