The Geometric Canary: Predicting Steerability and Detecting Drift via Representational Stability

Hugging Face Daily Papers Papers

Summary

This paper introduces geometric stability measures—based on pairwise distance consistency in representations—to predict language model steerability and detect structural drift. Supervised variants achieve near-perfect correlation (ρ=0.89-0.97) with linear steerability across 35-69 embedding models, while unsupervised variants outperform CKA and Procrustes for post-deployment drift detection.

Reliable deployment of language models requires two capabilities that appear distinct but share a common geometric foundation: predicting whether a model will accept targeted behavioral control, and detecting when its internal structure degrades. We show that geometric stability, the consistency of a representation's pairwise distance structure, addresses both. Supervised Shesha variants that measure task-aligned geometric stability predict linear steerability with near-perfect accuracy (ρ= 0.89-0.97) across 35-69 embedding models and three NLP tasks, capturing unique variance beyond class separability (partial ρ= 0.62-0.76). A critical dissociation emerges: unsupervised stability fails entirely for steering on real-world tasks (ρapprox 0.10), revealing that task alignment is essential for controllability prediction. However, unsupervised stability excels at drift detection, measuring nearly 2times greater geometric change than CKA during post-training alignment (up to 5.23times in Llama) while providing earlier warning in 73\% of models and maintaining a 6times lower false alarm rate than Procrustes. Together, supervised and unsupervised stability form complementary diagnostics for the LLM deployment lifecycle: one for pre-deployment controllability assessment, the other for post-deployment monitoring.
Original Article
View Cached Full Text

Cached at: 04/21/26, 07:21 AM

Paper page - The Geometric Canary: Predicting Steerability and Detecting Drift via Representational Stability

Source: https://huggingface.co/papers/2604.17698

Abstract

Geometric stability measures predict language model controllability and detect structural degradation, with supervised variants excelling at steering prediction and unsupervised variants at drift detection.

Reliable deployment of language models requires two capabilities that appear distinct but share a common geometric foundation: predicting whether a model will accept targeted behavioral control, and detecting when its internal structure degrades. We show thatgeometric stability, the consistency of a representation’s pairwise distance structure, addresses both.Supervised Shesha variantsthat measuretask-aligned geometric stabilitypredictlinear steerabilitywith near-perfect accuracy (ρ= 0.89-0.97) across 35-69 embedding models and three NLP tasks, capturing unique variance beyond class separability (partial ρ= 0.62-0.76). A critical dissociation emerges:unsupervised stabilityfails entirely for steering on real-world tasks (ρapprox 0.10), revealing that task alignment is essential for controllability prediction. However,unsupervised stabilityexcels atdrift detection, measuring nearly 2times greater geometric change thanCKAduring post-training alignment (up to 5.23times in Llama) while providing earlier warning in 73\% of models and maintaining a 6times lower false alarm rate thanProcrustes. Together, supervised andunsupervised stabilityform complementary diagnostics for theLLM deployment lifecycle: one for pre-deployment controllability assessment, the other for post-deployment monitoring.

View arXiv pageView PDFGitHub0Add to collection

Get this paper in your agent:

hf papers read 2604\.17698

Don’t have the latest CLI?curl \-LsSf https://hf\.co/cli/install\.sh \| bash

Models citing this paper1

#### pcr2120/shesha-geometry Other• Updatedabout 5 hours ago

Datasets citing this paper0

No dataset linking this paper

Cite arxiv.org/abs/2604.17698 in a dataset README.md to link it from this page.

Spaces citing this paper0

No Space linking this paper

Cite arxiv.org/abs/2604.17698 in a Space README.md to link it from this page.

Collections including this paper0

No Collection including this paper

Add this paper to acollectionto link it from this page.

Similar Articles

Perfect Detection, Failed Control: The Geometry of Knowing vs. Steering in Language Models

arXiv cs.CL

This paper investigates the geometric relationship between directions in language model activations that detect a behavior versus those that control it, finding that for hallucination detection they are nearly orthogonal (cosine ~0.12), while for output format they align perfectly, challenging a common assumption in mechanistic interpretability.

Controlling Tool Use with Heading-Specific Activation Steering

arXiv cs.AI

This paper investigates whether tool-use decisions in large language models have stable internal representations that can be extracted and manipulated via activation steering, demonstrating that heading-specific steering vectors can suppress unnecessary tool use across five open-source models and three domains. The geometric analysis reveals that tool-invocation steps exhibit diffuse, bimodal alignment rather than the clean linear structure expected for parametrically grounded concepts.