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

• URL: http://arxiv.org/abs/2605.09195v1
• Date: Sat, 09 May 2026 22:27:31 GMT
• Title: The Geometry of Forgetting: Temporal Knowledge Drift as an Independent Axis in LLM Representations
• Abstract summary: Large language models confidently produce outdated answers, and no existing method can detect them.<n>We show this is not an engineering failure but a structural one: temporal drift is encoded as a direction in the residual stream geometrically to both correctness and uncertainty.
• Score: 50.43168858368539
• License: http://creativecommons.org/licenses/by/4.0/
• Abstract: Large language models confidently produce outdated answers, and no existing method can detect them. We show this is not an engineering failure but a structural one: temporal drift, whether a stored fact has changed since training, is encoded as a direction in the residual stream geometrically orthogonal to both correctness and uncertainty. Any method operating on correctness or uncertainty signals is therefore blind to drift by construction. We verify this across six instruction-tuned models. A linear probe trained directly on drift labels achieves AUROC $0.83$--$0.95$; methods based on token entropy, semantic entropy, CCS, and SAPLMA all remain near chance ($0.49$--$0.57$). Five tests confirm the geometric orthogonality: weight cosines ($|\cos| \leq 0.14$), score correlations ($|r| \leq 0.20$), bidirectional null-space projection ($|Δ| \leq 0.008$), iterative null-space projection with $k{=}10$, and difference-of-means dissociation. Mechanistically, the MLP retrieval circuit produces identical dynamics for stale recall and confabulation ($r > 0.81$, six models), explaining why output confidence cannot separate them. A cross-cutoff experiment holds inputs constant and varies only the model: the probe fires on the model whose training predates the fact's transition and stays silent otherwise ($P(A{>}B) = 0.975$--$0.998$, twelve model pairs), confirming it reads model-internal knowledge state rather than input properties. Our code and datasets will be publicly released.

Related papers

• The exact price of local realism in CHSH experiments: a measurement-dependence-detection trade-off surface, a moiré phase-locking mechanism that saturates it, and an unmeasured fringe in the fourfold coincidence sum [0.0]
  I solve the unique detection profile $D(m)=sqrtm,h(m)$ under which a deterministic sign model reproduces the singlet exactly.<n>Surviving local accounts trade off measurement dependence against a $cos 4(a-b)$ modulation of the fourfold coincidence sum.<n>I exhibit a local mechanism -- moire phase locking -- with deterministic phase evolution and all randomness quenched in frozen offsets and flight times.
  arXiv  Detail & Related papers  (2026-08-19T13:14:52Z)
• Round-Trip Consistency: Bidirectional Diffusion Models Can Predict Their Own Rollout Errors [51.56484100374058]
  Autoregressive models accumulate error over long rollouts, yet at deployment there is no ground truth to measure it against.<n>We train a single conditional latent diffusion model that steps a dynamical system forward or backward in time via a direction flag.<n>We show that this bidirectionality supplies a measurement-free test-time error signal.
  arXiv  Detail & Related papers  (2026-08-01T13:49:46Z)
• Efficiently Learning Drifting Halfspaces with Massart Noise [50.4331323695175]
  We study the problem of learning a drifting concept in the presence of Massart noise.<n>In this framework, an online learner has access to a history of independent samples.<n>The goal is to output, in each round, a hypothesis with small prediction error.
  arXiv  Detail & Related papers  (2026-06-09T17:35:18Z)
• When and How Long? The Readout-Mediator Angle in Temporal Reasoning [1.0195618602298682]
  linear probe can decode a representation almost perfectly and yet be completely irrelevant to how the model uses it.<n> ablating a four-dimensional subspace found by Distributed Alignment Search (DAS) at the same layer collapses performance entirely.<n>We reverse-engineering the circuit reveals why: attention heads month-grained context through learned QK offsets at $pm30$ and $pm61$ days.
  arXiv  Detail & Related papers  (2026-05-27T21:38:17Z)
• Wavelet Variance Equipartition as a Threshold for World-Model Quality and Quantum Kernel TN-Simulability [0.0]
  We identify the wavelet scaling exponent $$ as a critical diagnostic.<n>We establish $= 1/2$ as a sharp transition boundary for the classical simulability of amplitude-encoded quantum kernels.<n>We prove this variance scales strictly as $Var[X] = (d-2)$.
  arXiv  Detail & Related papers  (2026-05-12T05:41:12Z)
• Information Routing in Atomistic Foundation Models: How Task Alignment and Equivariance Shape Linear Disentanglement [0.0]
  We introduce Compositional Probe Decomposition (CPD), which linearly projects out composition signal and measures how much geometric information remains accessible to a Ridge probe.<n>We find a emphlinear accessibility gradient: models differ by $6.6times$ in geometric information accessible after composition removal.<n>We also show that nonlinear probes produce misleading results on residualized representations, recovering $R2 = 0.68$--$0.95$ on a purely compositional target, and recommend linear probes for this setting.
  arXiv  Detail & Related papers  (2026-03-03T16:52:06Z)
• INC: An Indirect Neural Corrector for Auto-Regressive Hybrid PDE Solvers [61.84396402100827]
  We propose the Indirect Neural Corrector ($mathrmINC$), which integrates learned corrections into the governing equations.<n>$mathrmINC$ reduces the error amplification on the order of $t-1 + L$, where $t$ is the timestep and $L$ the Lipschitz constant.<n>We test $mathrmINC$ in extensive benchmarks, covering numerous differentiable solvers, neural backbones, and test cases ranging from a 1D chaotic system to 3D turbulence.
  arXiv  Detail & Related papers  (2025-11-16T20:14:28Z)
• IT$^3$: Idempotent Test-Time Training [95.78053599609044]
  Deep learning models often struggle when deployed in real-world settings due to distribution shifts between training and test data.<n>We present Idempotent Test-Time Training (IT$3$), a novel approach that enables on-the-fly adaptation to distribution shifts using only the current test instance.<n>Our results suggest that idempotence provides a universal principle for test-time adaptation that generalizes across domains and architectures.
  arXiv  Detail & Related papers  (2024-10-05T15:39:51Z)
• Improved Algorithm for Adversarial Linear Mixture MDPs with Bandit Feedback and Unknown Transition [71.33787410075577]
  We study reinforcement learning with linear function approximation, unknown transition, and adversarial losses. We propose a new algorithm that attains an $widetildeO(dsqrtHS3K + sqrtHSAK)$ regret with high probability.
  arXiv  Detail & Related papers  (2024-03-07T15:03:50Z)
• Beyond Invariance: Test-Time Label-Shift Adaptation for Distributions with "Spurious" Correlations [44.99833362998488]
  Changes in the data distribution at test time can have deleterious effects on the performance of predictive models. We propose a test-time label shift correction that adapts to changes in the joint distribution $p(y, z)$ using EM applied to unlabeled samples.
  arXiv  Detail & Related papers  (2022-11-28T18:52:33Z)
• Mean Estimation in High-Dimensional Binary Markov Gaussian Mixture Models [12.746888269949407]
  We consider a high-dimensional mean estimation problem over a binary hidden Markov model. We establish a nearly minimax optimal (up to logarithmic factors) estimation error rate, as a function of $|theta_*|,delta,d,n$.
  arXiv  Detail & Related papers  (2022-06-06T09:34:04Z)
• Online Robust Regression via SGD on the l1 loss [19.087335681007477]
  We consider the robust linear regression problem in the online setting where we have access to the data in a streaming manner. We show in this work that the descent on the $ell_O( 1 / (1 - eta)2 n )$ loss converges to the true parameter vector at a $tildeO( 1 / (1 - eta)2 n )$ rate which is independent of the values of the contaminated measurements.
  arXiv  Detail & Related papers  (2020-07-01T11:38:21Z)
• How isotropic kernels perform on simple invariants [0.5729426778193397]
  We investigate how the training curve of isotropic kernel methods depends on the symmetry of the task to be learned. We show that for large bandwidth, $beta = fracd-1+xi3d-3+xi$, where $xiin (0,2)$ is the exponent characterizing the stripe of the kernel at the origin.
  arXiv  Detail & Related papers  (2020-06-17T09:59:18Z)
• Quantum Algorithms for Simulating the Lattice Schwinger Model [63.18141027763459]
  We give scalable, explicit digital quantum algorithms to simulate the lattice Schwinger model in both NISQ and fault-tolerant settings. In lattice units, we find a Schwinger model on $N/2$ physical sites with coupling constant $x-1/2$ and electric field cutoff $x-1/2Lambda$. We estimate observables which we cost in both the NISQ and fault-tolerant settings by assuming a simple target observable---the mean pair density.
  arXiv  Detail & Related papers  (2020-02-25T19:18:36Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.