IC-562Applying the layer-16 self-recognition vector to input tokens (not output) of Llama3-8B-Instruct alters the model's perception of authorship, making it believe or disbelieve it wrote arbitrary texts in both individual and paired paradigms.
The authors test whether the self-recognition vector can influence the model's perception rather than just its output. By adding or subtracting the layer-16 vector from every token of the text being evaluated (leaving the output token untouched), they can shift the model's judgment in both directions. In the individual paradigm, zeroing out the vector's projection from both texts causes the model to decrease its confidence in the correct answer in 55.6% of SAD trials (p=0.0153) and 58.1% of QA trials (p=0.0008). In the paired paradigm, the same manipulation shifts the model's choice in 52.9% (SAD, p=0.1188) and 55.9% (QA, p=0.0021) of trials.
Evidence
interventional
Key metric
Zeroing out vector from text tokens: individual paradigm SAD 55.6% decreasing (p=0.0153), QA 58.1% decreasing (p=0.0008); paired paradigm SAD 52.9% decreasing (p=0.1188), QA 55.9% decreasing (p=0.0021).
Caveat
The paired SAD result (p=0.1188) is not statistically significant at the 0.05 level. The authors note that the vector was formed over a few hundred short texts and is unlikely to capture all signals about writing characteristics, so the effect is partial rather than complete.