Some activation capping settings slightly improve performance on IFEval, MMLU Pro, or GSM8k for both Qwen and Llama

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

• Activation capping reduces harmful response rate by nearly 60% without impacting performance on IFEval, MMLU Pro, GSM8k, and EQ-Benchfinding0.833
  Main quantitative result demonstrating effectiveness of activation capping
• DeepSeek-R1 Llama 8b gains 0.16% accuracy on GSM8k with positive intervention (more reflections) at cost of ~2000 additional tokensfinding0.779
  Only model showing marginal benefit from increased reflection, at substantial token cost
• Optimal activation capping layers for Llama 3.3 70B are layers 56-71 (out of 80) at 25th percentile capfinding0.779
  Specific implementation finding for Llama capping parameters
• Optimal activation capping layers for Qwen 3 32B are layers 46-53 (out of 64) at 25th percentile capfinding0.746
  Specific implementation finding for Qwen capping parameters
• Multi-attempt improvement rate peaks at 83% around -1.0σ below threshold in Llama-3.3-70Bfinding0.742
  Shows slightly weaker steering allows more successful corrections, characterizing optimal ESR conditions
• Layer 24 (indexed at 8) of LLaMA3.1-8B on Hinting satisfies Criteria 1 and 2 under both IIT 3.0 and IIT 4.0 (temporal permutation).finding0.739
  One of the most promising cases; approximately corresponds to the 2/3 layer of LLaMA3.1-8B.
• Scaling Laws for Activation Steering with Llama 2 Models and Refusal Mechanisms (Ali et al., 2025)concept0.738
  Related work finding larger models more resistant to steering, potentially consistent with ESR in 70B
• We hypothesize that Llama-3.1-8B deploys the same base-10 addition circuitry for cyclic reasoning as it uses for general arithmetic, independent of the concept domainhypothesis0.735
  Predictive hypothesis about domain-generality of the identified mechanism