Research map / Large language models / Language, safety and interpretability
A geometric relation of the error introduced by sampling a language model's output distribution to its internal state
Albert F. Modenbach
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| apply the derived holonomy operator instead of a norm-matched random SO(n) matrix (ablation) | q = Hy - y PCA on Qwen2.5-32B-Instruct · Chess branch points | variance in top 2 PCs: 34.8% -> 53.9% | Improved |
| apply the holonomy operator instead of a simple rotation of z_t onto v1 or v2 (ablation) | q PCA on Qwen2.5-32B-Instruct · Chess branch points | variance in top 2 PCs: 31.4% for simple rotation vs 53.9% | Improved |
| use coupling of Hy rather than q = Hy - y (curvature) (ablation) | holonomy coupling analysis · Chess branch points | semantic coupling to world vectors | Hurt |
| separate active and bulk probes by selecting the maximum coupling probe instead of averaging over all probes (ablation) | coupling analysis · Chess branch points | separability of active vs bulk probe couplings | Improved |
| consider delta q across greedy/branch boundary, coupling by chess piece type | Qwen2.5-32B-Instruct · Chess branch points | Spearman rho of coupling vs piece value: rho = 0.90, p = 0.037 | Improved |
| replace Qwen with held-out Mistral model | holonomy analysis · Chess branch points | piece-importance coupling and PCA structure: rho = -0.90, p = 0.037; flow direction reversed vs Qwen | Mixed |
| use poorer probes (Mistral) in place of Qwen probes | holonomy analysis · Chess branch points | PCA axis magnitude | Hurt |
| compare delta y (state difference) with delta q (curvature) coupling to world vectors | Qwen2.5-32B-Instruct · Chess branch points | coupling to probes | Improved |
Closest papers by meaning
- Token Embeddings Violate the Manifold Hypothesis (NeurIPS 2025 · similarity 0.87)
- Controlling Equational Reasoning in Large Language Models with Prompt Interventions (AAAI 2025 · similarity 0.86)
- Evaluating the World Model Implicit in a Generative Model (NeurIPS 2024 · similarity 0.86)
- Tracing the Representation Geometry of Language Models from Pretraining to Post-training (NeurIPS 2025 · similarity 0.86)
- Invariant Features in Language Models: Geometric Characterization and Model Attribution (NeurIPS 2026 · similarity 0.86)
- Lines of Thought in Large Language Models (ICLR 2025 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
