Geometric and Behavioral Stratification in Transformer Residual Streams

• URL: http://arxiv.org/abs/2608.12447v3
• Date: Wed, 19 Aug 2026 21:18:06 GMT
• Title: Geometric and Behavioral Stratification in Transformer Residual Streams
• Abstract summary: Trained transformer models develop privileged bases: coordinate axes whose statistics differ from the rest of the residual stream.<n>We investigate the prediction direction, the unembedding direction of the token a model currently predicts, and find that it functions as a content-defined privileged anchor.
• Score: 0.0
• License: http://creativecommons.org/licenses/by/4.0/
• Abstract: Trained transformer models develop privileged bases: coordinate axes whose statistics differ from the rest of the residual stream. But what kind of direction does such a basis select? We investigate the prediction direction, the unembedding direction of the token a model currently predicts, and find that it functions as a content-defined privileged anchor. Measured with respect to this anchor, residual-stream variation is geometrically and behaviorally stratified by proximity to the prediction. The stratification holds in all eighteen models tested (dense and mixture-of-experts, 7B-120B, base and instruction-tuned). A narrow, scale-invariant prediction interface concentrates readout-relevant structure, while the vast prediction-distal complement expands with model scale. Because the prediction direction sits nearly orthogonal to the principal variance axes, variance-based analyses recover this organization only partly, and the shortfall grows with prompt heterogeneity. Anchoring reveals a steep geometric gradient: prediction-proximal regions are highly structured and cluster related prompts, while the complement is flatter and anti-discriminates among prompt groups. The interface is a narrow slice but functionally decisive. Disrupting the variance directions closest to the prediction causes immediate divergence and frequent task-frame shifts; disrupting the next level down delays divergence and preserves framing. The complement is weakly readout-aligned per direction yet causally and temporally load-bearing, and behavior is driven by direction rather than magnitude. These results establish the prediction direction as a privileged anchor distinct from previously described coordinate axes, and give a geometric account of how high-dimensional computation coexists with linear readout.

Related papers

• Structured Extrema Errors in Classical Surrogates for Viscous Burgers: A Physics-Consistent Interpretation [0.0]
  We study the local errors of machine-learning surrogate models, which approximate the time evolution of the viscous Burgers equation.<n>Across all four models, the one-step residual, defined here as the true value minus the predicted value at each grid point, forms clear curved branches near predicted maxima and minima.<n>A more detailed analysis of KRR shows that these errors are much more strongly related to the second spatial derivative, which measures local curvature, than to the first spatial derivative.
  arXiv  Detail & Related papers  (2026-09-07T20:17:58Z)
• Orthogonal JEPA: Factorized Predictive States for Latent World Models [95.20158869896393]
  We introduce method, a latent world-modeling framework based on predictive factorization.<n>We show that method can allocate redundant capacity to dominant signals while providing weak or conflicting gradients to less dominant predictive structure.<n>Experiments on controlled vision, single-cell transcriptomics, longitudinal health records, continuous control, and molecular dynamics evaluate representation quality, forecasting, planning, and long-horizon stability.
  arXiv  Detail & Related papers  (2026-08-20T13:59:57Z)
• Predicting When Random Low-Dimensional Reparameterizations Train Neural Networks [3.980892205755342]
  We present an orientation-resolved quadratic master formula that predicts the random-slice residual from both the curvature spectrum and the reference-to-solution displacement profile.<n>We then introduce Random Mapping Networks (RaMaN), which instantiate the predicted latent dimension using structured Hadamard or seed-regenerated Gaussian maps.<n>RaMaN closely tracks measured transition locations and outperforms orientation-agnostic approximations when displacement direction matters.
  arXiv  Detail & Related papers  (2026-08-12T21:09:54Z)
• High-dimensional ridgeless least squares interpolation under spiked covariance structures [3.656751299156262]
  This paper investigates the behavior of the out-of-sample prediction risk of the high-dimensional ridgeless least-squares estimator when the feature dimension $p$ and the sample size $n$ grow proportionally.<n>We show that the signal energy distributed along latent spike directions determines whether overfitting leads to benign, tempered, or catastrophic overfitting.
  arXiv  Detail & Related papers  (2026-08-07T14:40:44Z)
• Binary Flow Matching: Prediction-Loss Space Alignment for Robust Learning [23.616336786063552]
  Flow matching has emerged as a powerful framework for generative modeling.<n>We identify a latent structural mismatch that arises when it is coupled with velocity-based objectives.<n>We prove that re-aligning the objective to the signal space eliminates the singular weighting.
  arXiv  Detail & Related papers  (2026-02-11T02:02:30Z)
• Revisiting Zeroth-Order Optimization: Minimum-Variance Two-Point Estimators and Directionally Aligned Perturbations [57.179679246370114]
  We identify the distribution of random perturbations that minimizes the estimator's variance as the perturbation stepsize tends to zero.<n>Our findings reveal that such desired perturbations can align directionally with the true gradient, instead of maintaining a fixed length.
  arXiv  Detail & Related papers  (2025-10-22T19:06:39Z)
• Causal vs. Anticausal merging of predictors [57.26526031579287]
  We study the differences arising from merging predictors in the causal and anticausal directions using the same data.<n>We use Causal Maximum Entropy (CMAXENT) as inductive bias to merge the predictors.
  arXiv  Detail & Related papers  (2025-01-14T20:38:15Z)
• Refined Risk Bounds for Unbounded Losses via Transductive Priors [67.12679195076387]
  We revisit the sequential variants of linear regression with the squared loss, classification problems with hinge loss, and logistic regression.<n>Our key tools are based on the exponential weights algorithm with carefully chosen transductive priors.
  arXiv  Detail & Related papers  (2024-10-29T00:01:04Z)
• Learning When the Concept Shifts: Confounding, Invariance, and Dimension Reduction [5.38274042816001]
  In observational data, the distribution shift is often driven by unobserved confounding factors. This motivates us to study the domain adaptation problem with observational data. We show a model that uses the learned lower-dimensional subspace can incur nearly ideal gap between target and source risk.
  arXiv  Detail & Related papers  (2024-06-22T17:43:08Z)
• Structured Radial Basis Function Network: Modelling Diversity for Multiple Hypotheses Prediction [51.82628081279621]
  Multi-modal regression is important in forecasting nonstationary processes or with a complex mixture of distributions. A Structured Radial Basis Function Network is presented as an ensemble of multiple hypotheses predictors for regression problems. It is proved that this structured model can efficiently interpolate this tessellation and approximate the multiple hypotheses target distribution.
  arXiv  Detail & Related papers  (2023-09-02T01:27:53Z)
• Neuro-Symbolic Entropy Regularization [78.16196949641079]
  In structured prediction, the goal is to jointly predict many output variables that together encode a structured object. One approach -- entropy regularization -- posits that decision boundaries should lie in low-probability regions. We propose a loss, neuro-symbolic entropy regularization, that encourages the model to confidently predict a valid object.
  arXiv  Detail & Related papers  (2022-01-25T06:23:10Z)
• Which Invariance Should We Transfer? A Causal Minimax Learning Approach [18.71316951734806]
  We present a comprehensive minimax analysis from a causal perspective. We propose an efficient algorithm to search for the subset with minimal worst-case risk. The effectiveness and efficiency of our methods are demonstrated on synthetic data and the diagnosis of Alzheimer's disease.
  arXiv  Detail & Related papers  (2021-07-05T09:07:29Z)
• Amplifying the Anterior-Posterior Difference via Data Enhancement -- A More Robust Deep Monocular Orientation Estimation Solution [7.540176446791261]
  Existing deep-learning based monocular orientation estimation algorithms face the problem of confusion between the anterior and posterior parts of the objects. We propose a pretraining method which focuses on predicting the left/right semicircle in which the orientation of the object is located. Experiment results show that the proposed semicircle prediction enhances the accuracy of orientation estimation, and mitigates the problem stated above.
  arXiv  Detail & Related papers  (2020-12-21T15:36:13Z)

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.