Abstract:Retrieval-augmented LLMs are deployed for tasks where evidence quality determines action safety, yet evaluation protocols assume that single-turn robustness predicts robustness when evidence accumulates across turns. We show this assumption is fundamentally incorrect. Models exhibit a monitoring-control gap: they readily acknowledge contradictory evidence, yet this awareness fails to constrain their final recommendations - detecting epistemic conflict does not imply resolving it safely. Through a multi-turn document accumulation protocol across four model families (1.5B-32B parameters) and over 50,000 turn-level evaluations, we demonstrate that single-turn diagnostics systematically overestimate RAG safety, that contradiction acknowledgement is uncorrelated with safe resolution, a pattern corroborated by targeted human validation, and that no universal prompt fix exists. Converging mechanism evidence - hidden-state probing, attention analysis, and response-strategy taxonomy - points to action selection as the most plausible locus of the deficit: danger-relevant information is internally represented and receives enhanced attention during unsafe generation, yet fails to constrain output behavior. The gap between what models recognize and what they do must be measured and closed before retrieval-augmented systems can be trusted in high-stakes settings.
Abstract:Post-training is routinely evaluated through aggregate benchmark scores that treat multi-hop reasoning as a single capability -- as if a model that answers more questions correctly must be better at assembling facts. We show that this assumption can be misleading: recipes with statistically indistinguishable atomic knowledge produce composition behaviour separated by over 40 percentage points, a phenomenon we call composition collapse: the systematic failure to assemble stably-known facts into chains, invisible to aggregate metrics. We introduce a double-gate protocol that changes the estimand from an aggregate compositionality gap to residual composition failure conditioned on stable atomic access, decomposing post-training gains into three independent channels: atomic stability, residual composition, and critical depth. On a benchmark of temporal factual chains spanning depths 2--11 across four post-training recipes, this decomposition reveals that post-training objectives shift composition capability in directions that aggregate metrics mask, and suggests that claims about multi-hop reasoning improvement should be accompanied by atomic-gate-controlled composition metrics. Diagnostic probes further show that a substantial share of measured composition failure reflects generation-time computation constraints rather than permanent inability to compose.
Abstract:Retrieval-augmented generation (RAG) increasingly underpins high-stakes applications, yet remains vulnerable to Confundo-style poisoning where adversarially optimized documents manipulate generated outputs. Existing defenses assume that detecting poisoned evidence prevents harm. We show this assumption is incorrect: models exhibit a monitoring-control gap -- they can detect contradictions in retrieved evidence yet still act on poisoned claims. We introduce the Cordon Principle -- no agent capable of final synthesis may access untrusted natural-language evidence -- and realize it through CORDON-MAS, a compartmentalized framework that enforces this principle architecturally by separating evidence extraction, cross-source audit, and answer synthesis into agents with asymmetric memory privileges. Across five BEIR datasets, CORDON-MAS reduces attack success rate by 92.4\% relative to undefended RAG. This reframes RAG poisoning from a detection problem to an information-flow control problem.