Generalization Beyond Binary Factual Verdicts
Determine how the dose-based coupling and counterweight account of self-repair translates to tasks other than the single-token True/False factual-verdict task studied in the paper, especially to generative tasks.
References
We studied one task at one answer position, a True/False verdict read at a single token. How it may translate to other tasks, and especially generative ones, remains an open question.
Whether the same accounting explains self-repair in settings other than IOI, the sensitivity of circuit discovery to ablation type, or partial recovery after refusal or unlearning interventions is left to future work.
The routing gains are established mainly for binary attribute claims; broader generation and faithful end-to-end comparisons with external decoding methods remain open.