Finding Belief Geometries with Sparse Autoencoders

• URL: http://arxiv.org/abs/2604.02685v1
• Date: Fri, 03 Apr 2026 03:29:48 GMT
• Title: Finding Belief Geometries with Sparse Autoencoders
• Abstract summary: We introduce a pipeline for discovering candidate simplex-structured subspaces in transformer representations.<n>We validate the pipeline on a transformer trained on a hidden Markov model with known belief-state geometry.<n>Preliminary evidence that genuine belief-like geometry exists in Gemma-2-9B's representation space.
• Score: 0.0
• License: http://creativecommons.org/licenses/by-sa/4.0/
• Abstract: Understanding the geometric structure of internal representations is a central goal of mechanistic interpretability. Prior work has shown that transformers trained on sequences generated by hidden Markov models encode probabilistic belief states as simplex-shaped geometries in their residual stream, with vertices corresponding to latent generative states. Whether large language models trained on naturalistic text develop analogous geometric representations remains an open question. We introduce a pipeline for discovering candidate simplex-structured subspaces in transformer representations, combining sparse autoencoders (SAEs), $k$-subspace clustering of SAE features, and simplex fitting using AANet. We validate the pipeline on a transformer trained on a multipartite hidden Markov model with known belief-state geometry. Applied to Gemma-2-9B, we identify 13 priority clusters exhibiting candidate simplex geometry ($K \geq 3$). A key challenge is distinguishing genuine belief-state encoding from tiling artifacts: latents can span a simplex-shaped subspace without the mixture coordinates carrying predictive signal beyond any individual feature. We therefore adopt barycentric prediction as our primary discriminating test. Among the 13 priority clusters, 3 exhibit a highly significant advantage on near-vertex samples (Wilcoxon $p < 10^{-14}$) and 4 on simplex-interior samples. Together 5 distinct real clusters pass at least one split, while no null cluster passes either. One cluster, 768_596, additionally achieves the highest causal steering score in the dataset. This is the only case where passive prediction and active intervention converge. We present these findings as preliminary evidence that genuine belief-like geometry exists in Gemma-2-9B's representation space, and identify the structured evaluation that would be required to confirm this interpretation.

Related papers

• ActMap: Single-Pass Uncertainty Quantification from Generation-Time Activation Maps [1.7403133838762448]
  ActMap is a representation that compresses the generation-time hidden-state trajectory into a fixed $12times32times128$ of temporal-statistic channels.<n>A lightweight classifier, instantiated as a compact Vision Transformer, reads an estimated correctness probability from each map in a fraction of a millisecond.
  arXiv  Detail & Related papers  (2026-09-10T13:04:24Z)
• Hyper-Fold: Exploring the Expressive Limit of Sequence-Geometry Learning for Proteins via Hypergraph Modeling [55.36837305276515]
  We show that the complete bilinear operator over content-geometry outer products--the sufficient statistic of all second-order interactions--is the expressive ceiling.<n>We then introduce Hyper-Fold, a rank-K separable convolutional backbone approaching this ceiling at message-passing cost.
  arXiv  Detail & Related papers  (2026-08-29T11:43:25Z)
• Tropical Algebraic Geometry for Neuronal Representations: An Arakelov-Green Measure Based Descriptor for Graph Learning [14.563432689352084]
  We propose a training-free geometric prior based on tropical algebraic geometry.<n>We convert spatial trees into cyclic metric graphs suitable for embedding into the Tropical Jacobian.<n>We show that the discrete Arakelov-Green measure, computed in closed form via the graph Laplacian's generalized inverse, decomposes exactly into the intrinsic path metric minus the unquantized polarization distance on this cover.
  arXiv  Detail & Related papers  (2026-08-05T05:34:50Z)
• Geometry of Ordinal Representations in Language Models [0.0]
  We show that geometric computation is architecture-dependent, with twisters preserving ordinal order, unlike numeric twisters.<n>We also find that Qwen3-4B shows substantially stronger twisting than Gemma models for indentation, and its twisters preserve ordinal order, unlike its numeric twisters.<n> Activation patching confirms that the identified manifold subspaces concentrate task-relevant information, with manifold-direction ablation causing dramatically larger probe accuracy drops than random-direction controls.
  arXiv  Detail & Related papers  (2026-07-05T08:13:26Z)
• Understanding Geometric Representations in Self-Supervised Vision Transformers via Subspace Intervention [7.887469434417159]
  We introduce a controlled subspace intervention framework to investigate how self-supervised Vision Transformers (ViTs) encode dense geometric information.<n>By connecting internal encoding mechanics with downstream performance, these findings provide a basis for effective feature selection and lightweight decoder design.
  arXiv  Detail & Related papers  (2026-07-02T10:18:02Z)
• MeshWeaver: Sparse-Voxel-Guided Surface Weaving for Autoregressive Mesh Generation [53.06620176910962]
  MeshWeaver is an autoregressive framework that treats mesh generation as a surface weaving process.<n>MeshWeaver achieves a state-of-the-art compression ratio of 18%, can generate meshes with up to 16K faces, and significantly improves geometric fidelity over prior approaches.
  arXiv  Detail & Related papers  (2026-06-03T10:15:43Z)
• From Extrinsic to Intrinsic: Geodesic-Guided Representation Learning for 3D Geometric Data [56.41479248637621]
  We introduce a novel 3D representation learning paradigm, namely textbfPRISM, for textbfPre-training.<n>PRISM learns embeddings by textbfRecovering the textbfIntrinsic textbfSurface geodesic textbfMetric.<n> Experiments demonstrate that our approach shows satisfactory accuracy, robustness, and high efficiency in geodesic distance prediction.
  arXiv  Detail & Related papers  (2026-06-01T13:54:26Z)
• Geometry-Aware Decoding with Wasserstein-Regularized Truncation and Mass Penalties for Large Language Models [9.059725329168435]
  Top-W is a geometry-aware truncation rule that uses Wasserstein distance-defined over token-embedding geometry.<n>We show that Top-W consistently outperforms prior state-of-the-art decoding approaches achieving up to 33.7% improvement.
  arXiv  Detail & Related papers  (2026-02-10T22:36:48Z)
• Automated Circuit Interpretation via Probe Prompting [0.0]
  We present probe prompting, an automated pipeline that transforms attribution graphs into compact, interpretable subgraphs.<n>Across five prompts, probe-prompted subgraphs preserve high explanatory coverage while compressing complexity.<n>Compared to geometric clustering baselines, concept-aligned groups exhibit higher behavioral coherence.
  arXiv  Detail & Related papers  (2025-11-10T11:53:36Z)
• Identifying General Mechanism Shifts in Linear Causal Representations [58.6238439611389]
  We consider the linear causal representation learning setting where we observe a linear mixing of $d$ unknown latent factors. Recent work has shown that it is possible to recover the latent factors as well as the underlying structural causal model over them. We provide a surprising identifiability result that it is indeed possible, under some very mild standard assumptions, to identify the set of shifted nodes.
  arXiv  Detail & Related papers  (2024-10-31T15:56:50Z)
• SpaceMesh: A Continuous Representation for Learning Manifold Surface Meshes [61.110517195874074]
  We present a scheme to directly generate manifold, polygonal meshes of complex connectivity as the output of a neural network.<n>Our key innovation is to define a continuous latent connectivity space at each mesh, which implies the discrete mesh.<n>In applications, this approach not only yields high-quality outputs from generative models, but also enables directly learning challenging geometry processing tasks such as mesh repair.
  arXiv  Detail & Related papers  (2024-09-30T17:59:03Z)
• Disentangled Representation Learning with the Gromov-Monge Gap [65.73194652234848]
  Learning disentangled representations from unlabelled data is a fundamental challenge in machine learning.<n>We introduce a novel approach to disentangled representation learning based on quadratic optimal transport.<n>We demonstrate the effectiveness of our approach for quantifying disentanglement across four standard benchmarks.
  arXiv  Detail & Related papers  (2024-07-10T16:51:32Z)
• G3Reg: Pyramid Graph-based Global Registration using Gaussian Ellipsoid Model [21.189016878269104]
  This study introduces a novel framework, G3Reg, for fast and robust global registration of LiDAR point clouds. In contrast to conventional complex keypoints and descriptors, we extract fundamental geometric primitives. We present a distrust-and-verify scheme based on a Pyramid Graph for Global Registration.
  arXiv  Detail & Related papers  (2023-08-22T17:23:00Z)
• SE(3)-Equivariant Attention Networks for Shape Reconstruction in Function Space [50.14426188851305]
  We propose the first SE(3)-equivariant coordinate-based network for learning occupancy fields from point clouds. In contrast to previous shape reconstruction methods that align the input to a regular grid, we operate directly on the irregular, unoriented point cloud. We show that our method outperforms previous SO(3)-equivariant methods, as well as non-equivariant methods trained on SO(3)-augmented datasets.
  arXiv  Detail & Related papers  (2022-04-05T17:59:15Z)

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.