When Gemma retrieves a bound entity, it does not rely on one pointer. It mixes a positional pointer (which slot in the list) with a lexical pointer (which words) and a reflexive one (the answer's own identity).

The setup is an interchange intervention. The model reads a list of bound groups, for example "{name} performed {genre} music on the {instrument}", repeated for 20 groups, and is asked for one entity. A counterfactual run is built so its positional slot (iₚ), lexical match (iₗ) and reflexive match (iᵣ) all point to different groups. Patching the last-token residual at layer 18 from that run, then reading which group the model now answers, reveals which pointer the patch carried.

Each panel below is one value of t_entity (which of the three slots is queried). Bars stack the four outcomes across positional index iₚ = 1…20. The positional band (orange) is the signature: tall at the edges, low in the middle. Where the positional pointer fades, the others fill in — and which one fills in flips from reflexive to lexical as the queried slot moves from first to last. That is the mixing.

Every step here — reading the paper, rebuilding the task and the §3.2 counterfactual from scratch on causalab, running the interchange sweep, and verifying the numbers against the raw data — was produced by Silico from a single prompt.