Few-shot prompting had a negative effect on HH-intent for smaller models and a significant positive impact on larger models.

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

• Chain-of-thought prompting produced a similar increase in helpful intention of large models as few-shot prompting.finding0.822
  Ablation result from Experiment 3 on chain-of-thought prompting effects.
• System prompting and few-shot produce highly similar activations (cosine 0.83-0.92 at safety-critical layers), indicating a shared prompt-side mechanism distinct from activation steering.claim0.785
  Evidence for two representational pathways based on cross-method activation divergence
• Smallest models have the lowest HH-intent scores, in accordance with their relative weakness at reasoning and adaptation.finding0.781
  Main result from Experiment 3 on HH-intent scaling with model size.
• Calibrated few-shot prompting was a surprisingly weak baseline for truth classification compared to linear probesfinding0.774
  Unexpected finding that behavioral baseline underperforms representational probing approaches
• few-shot promptingmethod0.773
  Providing k labeled examples in the prompt to steer model behavior.
• Fine-tuned LMs displayed higher mean HH-intent scores and increased consistency compared to pre-trained counterparts.finding0.756
  Main result from Experiment 3 on effect of fine-tuning on HH-intent.
• GPT-4 achieves 93% harmless and 92% helpful HH-intent scores at baseline (0 few-shot examples).finding0.742
  Numerical result from Table 3 for GPT-4.
• Benign few-shot exemplars independently prime safety (ASR 2.8% vs 3.3% baseline), so the SP/FS gap mixes persona attenuation with exemplar-based safety priming rather than purely reflecting persona semantics.claim0.730
  Control finding bounding interpretation of FS vs SP difference