Intent-Conditioned Prior–Context Integration via Protected Layerwise Calibration

Integrating parametric knowledge and contextual information requires adapting their influence to the user's intent. Given a statement naming London as France's capital, a task may require the factual answer Paris, the stipulated answer London, or both alternatives. We formulate intent-conditioned knowledge integration through continuous, state-dependent control of source reliance, rather than a single binary source assignment. We introduce EpiSetter, a lightweight two-stage framework with a frozen language model backbone. First, bidirectional coordinate interchange between prompts sharing the same question and context but differing in intent learns layerwise control directions through an empirical residual parameterization. Second, these directions are frozen, and jointly trained scalar calibrators predict bounded, intent-conditioned intervention magnitudes from hidden states modified by preceding interventions. Reference-answer supervision trains the calibration policy, while distribution-preservation regularization, displacement penalties, and protected-validation constraints seek to preserve the backbone's capabilities. Inference requires neither donor prompts nor reference answers. On OLMo3-7B, EpiSetter raises joint correctness across 417 held-out ConFiQA-QA intent pairs from 18.23% to 36.93%, improving both prior- and context-oriented answer accuracy. Its advantage over constant-magnitude control supports input-dependent calibration. Extracted-answer exact match and token F1 improve on five matched MRQA datasets, while capability tests show improved instruction following and largely retained reasoning and QA performance. These results support continuous intent-conditioned source control as a practical foundation for knowledge integration; the direction construction requires further numerical validation, and integration of complementary information from both sources remains to be established.