Rewriting or Reweighting? A Geometric Account in Language Models

• URL: http://arxiv.org/abs/2608.01835v1
• Date: Mon, 03 Aug 2026 07:46:28 GMT
• Title: Rewriting or Reweighting? A Geometric Account in Language Models
• Abstract summary: Post-training can substantially alter language-model behavior.<n>We study two mechanistically distinct failures, repetition as a decoding-attractor pathology and sycophancy as a preference-related alignment failure.<n>We introduce behavioral manifold analysis, which isolates behavior-specific geometry by selecting sparse behavior-associated coordinates.
• Score: 44.66056836026845
• License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/
• Abstract: Post-training can substantially alter language-model behavior, yet aggregate behavior rates do not reveal whether training removes an existing mechanism, creates a new one, or changes how an inherited mechanism is used. We study this question through two mechanistically distinct failures, repetition as a decoding-attractor pathology and sycophancy as a preference-related alignment failure. We introduce behavioral manifold analysis, which isolates behavior-specific geometry by selecting sparse behavior-associated coordinates and lifting them into low-dimensional local charts. We construct these charts in two complementary spaces. ACT captures runtime activation states, while NOC quantifies how strongly the model routes functional information flow through the shared behavior-associated subspace. Across multiple model families, the resulting charts are highly compressed and partially alignable across architectures. Contribution-space charts expose a more architecture-robust shared core, whereas activation-space charts retain stronger family-specific structure. Tracking these charts through controlled post-training reveals a consistent asymmetry. Supervised fine-tuning substantially alters the inherited behavioral geometry, whereas reward optimization changes behavior while largely preserving the underlying chart. This geometric perspective provides a unified framework for understanding the mechanistic distinction between the two objectives. SFT tends to rewrite behavioral geometry, whereas reward optimization primarily reweights it. Code is available at https://github.com/ronglingze/Manifold-Analysis

Related papers

• The information geometry of large language models is shared, learned, and controllable [0.0]
  We show that behaviour determines geometry up to output-preserving symmetries.<n>Across transformer, state-space and recurrent models, output geometries agree more strongly than activation geometries.<n> Controlled language assignments show that geometry follows the language law across architectures.
  arXiv  Detail & Related papers  (2026-09-10T04:09:49Z)
• MENTIS: What Belief Changes Under Alignment? Measuring Multi-Scale Latent Torsion in Language Models [16.071341117554038]
  We introduce MENTIS, a framework for measuring alignment-induced internal reorganization.<n>Across four 7-8B model pairs on LITMUS, our study reveals that alignment-induced change is selective rather than uniform.<n>These results suggest preference alignment leaves structured, depth-localized geometric signatures in internal computation beyond what behavior-level evaluation alone can reveal.
  arXiv  Detail & Related papers  (2026-05-31T07:05:51Z)
• Failure Modes for Deep Learning-Based Online Mapping: How to Measure and Address Them [10.503307305047487]
  We propose a framework to identify and measure the underlying failure modes by disentangling two effects: Memorization of input features and overfitting to known map geometries.<n>We introduce Fréchet distance-based reconstruction statistics that capture per-element shape fidelity without threshold tuning.<n>We formulate an MST-based sparsification strategy that reduces redundancy and improves balancing and performance while shrinking training size.
  arXiv  Detail & Related papers  (2026-03-20T11:09:11Z)
• GraphShaper: Geometry-aware Alignment for Improving Transfer Learning in Text-Attributed Graphs [16.624063216788638]
  We introduce textbfGraphShaper, a geometry-aware framework that enhances graph encoding through multi-geometric specialization.<n>Our approach employs expert networks tailored to different geometric spaces, dynamically computing fusion weights to adaptively integrate geometric properties.<n>It achieves 9.47% accuracy improvements on citation networks and 7.63% on social networks in zero-shot settings.
  arXiv  Detail & Related papers  (2025-10-14T02:48:50Z)
• Graph Your Own Prompt [44.358377952850994]
  Graph Consistency Regularization (GCR) is a framework that injects relational graph structures, derived from model predictions, into the learning process.<n>GCR promotes cleaner feature structure, stronger intra-class cohesion, and improved generalization, offering a new perspective on learning from prediction structure.
  arXiv  Detail & Related papers  (2025-09-27T15:45:07Z)
• A Top-down Graph-based Tool for Modeling Classical Semantic Maps: A Crosslinguistic Case Study of Supplementary Adverbs [50.982315553104975]
  Semantic map models (SMMs) construct a network-like conceptual space from cross-linguistic instances or forms.<n>Most SMMs are manually built by human experts using bottom-up procedures.<n>We propose a novel graph-based algorithm that automatically generates conceptual spaces and SMMs in a top-down manner.
  arXiv  Detail & Related papers  (2024-12-02T12:06:41Z)
• G-MSM: Unsupervised Multi-Shape Matching with Graph-based Affinity Priors [52.646396621449]
  G-MSM is a novel unsupervised learning approach for non-rigid shape correspondence. We construct an affinity graph on a given set of training shapes in a self-supervised manner. We demonstrate state-of-the-art performance on several recent shape correspondence benchmarks.
  arXiv  Detail & Related papers  (2022-12-06T12:09:24Z)
• Decoupled Multi-task Learning with Cyclical Self-Regulation for Face Parsing [71.19528222206088]
  We propose a novel Decoupled Multi-task Learning with Cyclical Self-Regulation for face parsing. Specifically, DML-CSR designs a multi-task model which comprises face parsing, binary edge, and category edge detection. Our method achieves the new state-of-the-art performance on the Helen, CelebA-HQ, and LapaMask datasets.
  arXiv  Detail & Related papers  (2022-03-28T02:12:30Z)
• Mitigating Generation Shifts for Generalized Zero-Shot Learning [52.98182124310114]
  Generalized Zero-Shot Learning (GZSL) is the task of leveraging semantic information (e.g., attributes) to recognize the seen and unseen samples, where unseen classes are not observable during training. We propose a novel Generation Shifts Mitigating Flow framework for learning unseen data synthesis efficiently and effectively. Experimental results demonstrate that GSMFlow achieves state-of-the-art recognition performance in both conventional and generalized zero-shot settings.
  arXiv  Detail & Related papers  (2021-07-07T11:43:59Z)
• Spatial-spectral Hyperspectral Image Classification via Multiple Random Anchor Graphs Ensemble Learning [88.60285937702304]
  This paper proposes a novel spatial-spectral HSI classification method via multiple random anchor graphs ensemble learning (RAGE) Firstly, the local binary pattern is adopted to extract the more descriptive features on each selected band, which preserves local structures and subtle changes of a region. Secondly, the adaptive neighbors assignment is introduced in the construction of anchor graph, to reduce the computational complexity.
  arXiv  Detail & Related papers  (2021-03-25T09:31:41Z)

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.