Typicality weight α = 0.65 ± 0.07 (p < 10^-14) on correctness-matched HelpSteer pairs using GLM-4.5 as reference

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

• Typicality weight α = 0.57 ± 0.07 (p < 10^-14) on correctness-matched HelpSteer pairs using Llama-3.1-405B as referencefinding0.911
  Empirical evidence for positive typicality bias in human preference data independent of true task utility
• Typicality bias with weight α>0 mathematically sharpens the aligned policy via a power transform of the reference distributionclaim0.811
  Theoretical result showing that any positive typicality bias weight γ-sharpens the reference distribution, amplifying modes
• After controlling for token count, Flesch-Kincaid readability, type-token ratio, and sentence length, typicality coefficient α remains positive and significant (0.260-0.326, p<10^-6)finding0.787
  Shows typicality bias is not fully explained by surface-form confounds
• Under matched-strength calibration (alpha*=1.0), P12 remains safe under SP (3.8% ASR) and dangerous under matched AS (81.8%), while P04 does not reverse (41.4% SP vs 29.5% AS).finding0.785
  Confirms the prosocial paradox is not due to mismatched intervention strength
• 47.69% of 130 injection-manipulated alpha trends have near-linear fits (R2 >= 0.95); 96.15% have roughly linear fits (R2 >= 0.75)finding0.779
  Demonstrates alignment with Linear Representation Hypothesis: target trait steers approximately linearly with alpha
• Strength comparison pair (3,7) with |Δα|=4 outperforms pair (3,5) with |Δα|=2, indicating graded sensitivity to perturbation magnitudefinding0.760
  Shows that introspective accuracy scales with injection strength difference, not binary detection
• Binary detection accuracy (up to 97.3% at L0 α=5) is entirely explained by global logit shifts (r=0.999 correlation with control)finding0.756
  Core negative result: the binary detection paradigm cannot distinguish genuine introspection from uniform output bias
• Binary detection adjusted accuracy reaches 97.3% at layer 0 with α=5 before baseline control is appliedfinding0.751
  The misleadingly high result that prior paradigm would report as evidence of introspection