IC-105Value anchoring produces a sinusoidal scoring pattern around the circular value structure, with scores decreasing as circular distance from the anchor increases

Naama Rozen, Liat Bezalel, Gal Elidan, Amir Globerson, Ella Daniel

SourceDo LLMs have Consistent Values?

To explain why value anchoring yields human-like correlations, the paper arranged the 19 anchor values in their circular order and, for each anchor condition, normalized scores by setting the anchored value to zero. The mean normalized scoring pattern across all 19 anchor conditions followed a sinusoidal function: the anchored value received the highest score, neighboring values on the circle received similarly high scores, and values farther around the circle received progressively lower scores. This pattern held across all models except Gemma-2-9B, confirming that the model's scoring is governed by circular proximity in the value structure.

Evidence
correlational
Caveat
The sinusoidal pattern does not hold for Gemma-2-9B, which is the exception among the six models tested.
Model
GPT-4 / ChatGPT4 / GPT-4 Code Interpreter / GPT-4 Technical Report GPT-4-0314, Gemini 1.0 Pro, Llama 3.1 8B, 70B, Gemma 2 9B, 27B
Concepts
Circular representation
Datasets
PVQ-RR [eval]
Methods
PVQ-RR [eval]
Related findings
IC-103, IC-104
Extraction
automatic-extraction