What Matters Is Which Inputs You Protect: Behavioural Constraints in Continual Adaptation of Language Models

Continual adaptation allows a language model to learn new tasks, but often reduces performance on tasks learned earlier. A common way to limit this forgetting is to penalise changes in the model's output distribution relative to a reference model. Existing methods differ in the divergence they use, the teacher they compare against, and the inputs on which the penalty is evaluated, which makes it difficult to tell which choice is responsible for improved retention. We isolate one of them: the support of the behavioural constraint, or the inputs on which the regularisation term is evaluated. Keeping the teacher, coefficient, temperature, divergence, and batch size fixed, we compare constraints evaluated on the current task, on inputs from previous tasks and general data, or on both. Across three task orders, methods whose constraint includes inputs outside the current task outperform an active-only constraint in 8 of 9 order-matched comparisons (sign test $p=0.0195$), with an average improvement in backward transfer over sequential fine-tuning 1.81 times larger. We find no evidence that excluding current-task inputs is beneficial: the strongest and most order-stable configuration constrains both sets. Protected coverage also makes results much less sensitive to task order: the worst order costs sequential fine-tuning 0.159 in backward transfer, against 0.011--0.050 for the constrained methods. We then ask whether behavioural change can predict forgetting. Change measured after individual updates has essentially no out-of-sample predictive power ($R^2\leq 0.001$), whereas change accumulated over a complete task is moderately associated with subsequent forgetting ($\rho=0.379$). Finally, teacher-forced token-level measurements substantially understate sequence-level changes during generation, by factors of 47--98 on long-output tasks. Behavioural change therefore appears more useful for deciding where to regularise than as an update-level predictor of future forgetting.