Reliability of internal-state-based measures of factual encoding

Determine the extent to which internal-state-based methods for measuring factual encoding in large language models reliably capture whether a fact is truly stored in the model’s parameters, as opposed to merely correlating with behavioral reproduction under training-like contexts.

Background

The paper distinguishes between encoding (parametric storage) and recall (access) and notes that many prior approaches to measuring encoding rely on access to internal model states. Because frontier models often lack accessible internals, the authors adopt a behavioral, pre-training-like reproduction criterion to operationalize encoding.

They highlight that, despite widespread use, the validity of internal-state-based approaches as faithful indicators of true parametric encoding has not been firmly established. This motivates a need to rigorously assess whether such methods genuinely reflect stored facts rather than artifacts or proxies.

References

Existing approaches to measuring encoding often rely on access to internal states, a requirement that does not align with our focus on evaluating frontier LLMs. Moreover, the extent to which these methods reliably capture whether a model truly encodes a fact remains an open question \citep{HaseBKG23, Ma2024bird, Huang2024Demys, WeiYWMZ0024, ChenC00025, Haller2025brittle}.

— Empty Shelves or Lost Keys? Recall Is the Bottleneck for Parametric Factuality  (2602.14080 - Calderon et al., 15 Feb 2026) in Section 2.1, Operationalizing Encoding and Knowledge (Encoding)

Open controls remain: plan-content specificity vs. position/length/discourse; whether the orthogonal R1 direction is a distinct concept or a rotated encoding; strict stripping specificity; a keep_recent budget sweep; probe causality beyond step index at scale ($n{\geq}100$); and replication beyond our models and domains.

— Plans Don't Persist: Why Context Management Is Load Bearing for LLM Agents  (2606.22953 - Mehta et al., 22 Jun 2026) in Section 6, Discussion and Limitations, subsection “What we do not establish”