IC-1632Tulu-13B uses a direct binding mechanism rather than binding ids for multiple-choice question tasks

Jiahai Feng, Jacob Steinhardt

SourceHow do Language Models Bind Entities in Context?

For a reversed MCQ format on SST-2, the paper shows that Tulu-13B's activations are not factorizable in the binding id sense. Instead of an abstract binding id mediating between label and option, the label and option text are directly bound in the same token activations. Copying only the suffix 'ative' and the following whitespace from a source context into the target context suffices to almost completely change the model's belief, indicating the binding information is stored directly at those positions. This contrasts with the binding id mechanism where such a substitution would not change the binding.

Evidence
interventional
Caveat
The reversed MCQ format is described as 'too out of distribution for llama-30b to solve,' so the finding is specific to Tulu-13B and this particular prompt format. Only a single task (SST-2) was tested for this alternative mechanism.
Model
Tulu
Datasets
SST-2 [eval]
Methods
Causal mediation analysis / Vig et al. 2020 (causal mediation analysis) [primary]
Related findings
IC-1630, IC-1631
Extraction
automatic-extraction