Summary
Transformer AI models can recognize how expert their conversation partner is early inside their processing but only start using this understanding much later to influence their responses. The authors studied dialogues between personas with different expertise levels and found that information about partner expertise is detectable in early layers of the model but does not affect output until after the model’s midpoint. They showed that changing this expertise information early in the model hardly changes the final answer, while changes later have a big effect. This means the model stores inferred information well before it actually uses it to guide its behavior.
transformerresidual streamlinear decodabilitypartner expertisedialogue systemscounterfactual patchinglayerwise representationinferencecausal influence
Authors
Mika Okamoto, Gabriele Sarti
Abstract
A transformer can make an attribute linearly decodable in its residual stream at a depth where that attribute does not yet influence the output. This gap between where information is readable and where it is used has been shown for attributes stated directly in the input. We ask whether it also holds for an attribute the model must infer gradually over a conversation, namely how expert its dialogue partner is. Using ExpertCollab, a corpus of multi-turn research-planning dialogues between model-played personas at four expertise levels, we find that partner expertise is most decodable in the early layers and falls to near chance before the midpoint of the network. Counterfactual patching shows that injecting the expertise difference at the layer of peak decodability barely changes a fixed late-layer readout, whereas the same difference injected past the midpoint propagates almost completely, a separation of more than an order of magnitude. A content-matched random control and a probe-free diagnostic place the transition at the same early layer, and a statically specified control attribute stays decodable throughout. An inferred relational attribute is therefore represented well before it becomes causally active, which bounds where any attempt to read out or steer partner-conditioned behavior must intervene. We use one model on a synthetic corpus as an initial demonstration.