The Geometry of Forgetting: Temporal Knowledge Drift as an Independent Axis in LLM Representations

Researchers demonstrate that large language models encode temporal knowledge drift—whether facts have become outdated since training—as a geometrically orthogonal direction in their internal representations, separate from correctness and uncertainty signals. This structural property explains why existing detection methods fail and why LLMs confidently produce outdated information, with implications for AI reliability and deployment.

This research identifies a fundamental architectural problem in how large language models store and retrieve knowledge. The study reveals that temporal drift operates as an independent geometric axis within model representations, meaning standard approaches measuring confidence or semantic consistency cannot detect outdated information by design. The researchers validated this across multiple instruction-tuned models using five distinct geometric tests, all confirming the orthogonality claim.

The findings connect to broader concerns about AI system reliability. As LLMs become integrated into critical applications—from healthcare to financial advising—the inability to distinguish stale facts from accurate ones poses significant risks. The mechanistic insight that MLP circuits produce identical dynamics for outdated recall and confabulation explains why users cannot rely on model confidence as a signal for trustworthiness.

From an industry perspective, this work has immediate implications for developers building LLM-based applications. Organizations cannot depend on existing uncertainty quantification methods to flag temporal issues. The paper's proposed linear probe achieves strong performance (AUROC 0.83-0.95) but requires explicit training on drift labels, suggesting specialized monitoring infrastructure may be necessary for production systems dealing with time-sensitive information.

The cross-cutoff experiments demonstrating that probes respond to model training dates rather than input properties establish that temporal drift is genuinely encoded in model weights. This opens pathways for targeted interventions—either through training procedures, retrieval-augmented generation, or direct geometric manipulation of representations. Future work likely focuses on whether drift can be detected without explicit labels or mitigated during inference.

• →Temporal knowledge drift is geometrically orthogonal to correctness and uncertainty in LLM representations, making standard detection methods structurally blind to outdated information.
• →Linear probes trained on drift labels achieve 83-95% AUROC while entropy and semantic uncertainty methods perform at chance level (49-57%), confirming the independence of these signal dimensions.
• →MLP retrieval circuits produce identical activation patterns for stale recall and confabulation, explaining why output confidence cannot distinguish outdated from fabricated responses.
• →The geometric orthogonality holds consistently across six instruction-tuned models and survives multiple validation tests including weight analysis and null-space projection.
• →Production systems using LLMs for time-sensitive tasks cannot rely on existing confidence signals and may require specialized monitoring mechanisms to detect outdated knowledge.