The Geometry of Forgetting: Temporal Knowledge Drift as an Independent Axis in LLM Representations
A new research paper published on arXiv reveals a structural limitation in Large Language Models (LLMs) regarding outdated information. The study, titled 'The Geometry of Forgetting,' demonstrates that temporal knowledge drift—when stored facts change after training—is encoded in the model's residual stream along a geometric axis orthogonal to both correctness and uncertainty. Consequently, existing detection methods relying on confidence or entropy signals are inherently blind to this drift. The authors verified this phenomenon across six instruction-tuned models, showing that a linear probe trained on drift labels achieved high accuracy (AUROC 0.83–0.95), while traditional methods performed near chance levels. Mechanistic analysis indicates that MLP retrieval circuits produce identical dynamics for stale recall and confabulation, explaining why output confidence fails to distinguish them. Cross-cutoff experiments confirmed that the probe detects internal model knowledge states rather than input properties. This findings suggest that addressing outdated answers requires new architectural approaches rather than simple engineering fixes. The researchers plan to publicly release their code and datasets to facilitate further investigation into this structural issue within AI systems.
Wire timeline
The Geometry of Forgetting: Temporal Knowledge Drift as an Independent Axis in LLM Representations
A new research paper published on arXiv reveals a structural limitation in Large Language Models (LLMs) regarding outdated information. The study, titled 'The Geometry of Forgetting,' demonstrates that temporal knowledge drift—when stored facts change after training—is encoded in the model's residual stream along a geometric axis orthogonal to both correctness and uncertainty. Consequently, existing detection methods relying on confidence or entropy signals are inherently blind to this drift. The authors verified this phenomenon across six instruction-tuned models, showing that a linear probe trained on drift labels achieved high accuracy (AUROC 0.83–0.95), while traditional methods performed near chance levels. Mechanistic analysis indicates that MLP retrieval circuits produce identical dynamics for stale recall and confabulation, explaining why output confidence fails to distinguish them. Cross-cutoff experiments confirmed that the probe detects internal model knowledge states rather than input properties. This findings suggest that addressing outdated answers requires new architectural approaches rather than simple engineering fixes. The researchers plan to publicly release their code and datasets to facilitate further investigation into this structural issue within AI systems.
cs.AI updates on arXiv.org