A verbalizable workspace holds concepts but not role bindings: probes decode roles near ceiling, yet the workspace reads below a rank-matched random subspace
Synopsis
A verbalizable workspace proposed as a substrate for runtime alignment monitoring was tested for role-filler binding, the assignment that separates propositions sharing a concept set, such as the model plans to deceive the user versus the user plans to deceive the model. Across Qwen3.6-27B, Gemma-3-12B and Gemma-3-27B-IT, two lens families and two role-direction estimators, the residual stream decodes role near ceiling while the workspace slice scores below a rank-matched random subspace, and pushing along a fitted role direction changes what the model reports about roles without changing which entity it binds as agent. Ablating the workspace damages binding specifically only on Gemma-3-27B-IT at the entity token, replicating across three lens variants.
Interpretation
The verbalizable workspace carries a readout about roles rather than the representation the model uses to compute role assignment, and this dissociation appears consistently across three frontier language models. The workspace had been proposed as a candidate substrate for runtime alignment monitoring, but whether its concepts are bound into semantic roles was explicitly left open. A linear probe decodes role from the residual stream near ceiling (0.989 on Qwen3.6-27B, 0.975 on Gemma-3-12B, 0.978 on Gemma-3-27B-IT) and nearly as well from the orthogonal remainder the lens discards, while J-space decodes at 0.587, 0.573 and 0.574, at or below the matched-rank random subspace at 0.643, 0.589 and 0.590.
Pushing activations along a role direction fitted inside the workspace changes what the model reports about role relations but not which entity it actually binds. If the workspace carried binding, a role push should shrink the log-odds gap between the two role sentences of a family from opposite sides and exceed strength-matched controls. On Gemma-3-27B-IT at the entity token, the pooled effect grows more than tenfold as the LRE coefficient rises from 64 to 256, but the non-participant control grows with it, the real-push-to-control ratio stays between 0.83 and 1.01, and the pooled mean stays inside the null band throughout; all three models are classified as indistinguishable from controls.
Ablating the workspace produces binding-specific damage only on Gemma-3-27B-IT at the entity token, showing the design detects binding when it is present. Two models show no binding-specific damage at that site, so an internal positive control is needed to separate a genuine absence from an intervention too weak to reach binding. The binding-specific deficit on Gemma-3-27B-IT at the entity token is 0.284 (95% CI 0.275 to 0.293), concept accuracy holds at 0.920, the two edits' residual-norm perturbations stay within 7% of each other, and three lens variants give 0.284, 0.334 and 0.297.
Perspective
The test covers three dense language models (Qwen3.6-27B, Gemma-3-12B and Gemma-3-27B-IT) on English agent-patient minimal-pair stimuli, with interventions and ablation applied at a single read layer per model at roughly 75% relative depth, across the doctor/nurse, teacher/student and driver/passenger concept pairs and four syntactic constructions.
The identity-swap edit-strength check fell below the 0.05 floor on all three models (counterpart-probability shifts of 0.0012, 0.0048 and 0.0001), so the steering null is consistent both with the fitted direction being irrelevant to binding and with the edit being too weak to reach it; the positive ablation on Gemma-3-27B-IT is confounded between instruction tuning and model scale, so which factor the workspace-load-bearing signal belongs to is unresolved.
