The Attribution Blind Spot: Detecting When Language Models Rely on Memory Rather Than Retrieved Context

2026-05-26

AI Summary

Existing methods struggle to reliably distinguish whether language model outputs originate from retrieved context or pretrained memory, complicating trustworthy attribution in high-stakes scenarios. This work proposes Computational Reality Monitoring (CRM), which identifies and formally names the “attribution blind spot” problem for the first time. Drawing on reality monitoring theory from cognitive science, CRM discerns information sources by analyzing differences in internal model representations—with versus without retrieval context—rather than relying solely on output-level signals. The approach integrates internal representation comparison, hierarchical difference detection, and module-level noise intervention. Experiments across nine variants of three major model families confirm the existence and regularity of such representational signals, demonstrating their generalization across tasks and datasets, while also revealing failure modes on domain-confusion benchmarks. These findings lay the groundwork for building attributable and traceable generative systems.

Abstract

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.

Problem

• attribution blind spot
• retrieval-augmented generation
• language models
• pretraining memory
• context reliance

Innovation

• attribution blind spot
• retrieval-augmented generation
• Computational Reality Monitoring
• representational divergence
• evidence provenance