IC-1372Successor heads in Pythia-1.4b exhibit a greater-than bias: the OV circuit assigns systematically higher logits to tokens with greater ordinal values than the input, impairing decrementation

Rhys Gould, Euan Ong, George Ogden, Arthur Conmy

SourceSuccessor Heads: Recurring, Interpretable Attention Heads In The Wild

The effective OV circuit of the successor head shows that logits on or above the main diagonal (output token ≤ input token) are lower than logits below the diagonal (output token > input token), for cardinal words, numbers, and other ordinal tasks. This means the head is biased toward predicting values greater than the input. As a consequence, vector arithmetic using mod-10 features to shift toward a lower ordinal value fails: for example, wu·Wov(MLP0(We('35')) − kf5 + kf3) places high logits on '43' rather than '33', a +10 effect occurring for 2/3 of entries below the diagonal in the 20–29 numbers table. The mod-10 features themselves do not exhibit this bias, indicating the head uses additional numeric information beyond the mod-10 features.

Evidence
correlational
Key metric
the +10 effect occurs for 2/3 of entries below the diagonal in the 20-29 numbers table; vector arithmetic works on 53% (months) and 89% (digits 20-29) only when target residue > source residue
Caveat
The authors note that the absence of a strong greater-than bias in the mod-10 features suggests the feature-level description is incomplete and that successor heads must use other numeric information to produce the bias.
Model
Pythia 1.4B
Related findings
IC-1369, IC-1370, IC-1371
Extraction
automatic-extraction