The Attribution Blind Spot: Detecting When Language Models Rely on Memory Rather Than Retrieved Context
For high-stakes deployment of retrieval-augmented generation, this work provides a method to detect when models rely on memory rather than retrieved evidence, addressing a critical failure mode that output-level monitors miss.
The paper identifies the 'attribution blind spot' where language models can produce context-consistent output from parametric memory rather than retrieved context, and introduces Computational Reality Monitoring (CRM) to detect this via internal representational divergence. Across nine model variants, CRM shows that internal signals can diagnose reliance on memory versus retrieved context, a signal invisible at the output level.
Retrieval-augmented generation promises to ground language model outputs in external evidence, yet the field has no reliable way to verify whether retrieved context actually governs generation -- a prerequisite for any high-stakes deployment. The standard assumption, that context-consistent output implies context-governed output, breaks when the retrieved document overlaps with the model's pretraining data: the model can produce faithful-looking text entirely from parametric memory, and both pathways yield indistinguishable output. We name this failure the attribution blind spot and introduce Computational Reality Monitoring (CRM) to address it. CRM operationalizes a principle adapted from cognitive science's reality monitoring framework: comparing internal representations with and without context reveals membership-conditioned representational divergence that output-level monitors systematically miss. CRM does not certify which source an individual generation used; it detects whether pretraining exposure leaves a measurable internal trajectory signature, establishing a necessary substrate for source attribution. Across nine model variants spanning three families, this divergence concentrates in architecture-specific layer patterns, receives converging support from block-level noise intervention, and generalizes across tasks and datasets while collapsing on domain-confounded benchmarks. The attribution blind spot is measurable and partially addressable: internal representations carry a diagnostic signal invisible at the output level, establishing a foundation for systems whose internal awareness of evidence provenance governs their external behavior.