The Geometric Canary: Predicting Steerability and Detecting Drift via Representational Stability
| Authors | Prashant C. Raju |
| Year | 2026 |
| HF Upvotes | 2 |
| arXiv | 2604.17698 |
| Download | |
| HF Page | View on Hugging Face |
Abstract
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.
Engineering Breakdown
Plain English
This paper introduces geometric stability as a unified framework for solving two critical problems in language model deployment: predicting whether a model can be steered toward specific behaviors, and detecting when its internal structure degrades. The authors developed Supervised Shesha variants that measure task-aligned geometric stability—essentially tracking how consistent the pairwise distances between data points remain in a model's representation space. Their approach achieves near-perfect prediction accuracy (ρ = 0.89-0.97) for steering capability across 35-69 embedding models and three NLP tasks, and crucially, captures unique variance beyond standard class separability metrics (partial ρ = 0.62-0.76). However, they discovered a critical finding: unsupervised stability measures fail entirely for real-world steering tasks (ρ ≈ 0.10), proving that task alignment is essential rather than optional.
Core Technical Contribution
The core novelty is recognizing that geometric stability—the preservation of distance relationships in learned representations—is the common geometric foundation for both steering robustness and structural degradation detection. Rather than treating steering and robustness as separate problems requiring separate solutions, the authors show they share an underlying geometric property measurable through task-aligned stability metrics. The Supervised Shesha variants operationalize this by computing stability scores that directly correlate with steering success, moving beyond correlation-based or class-separability proxies that have been the standard in interpretability work. This is fundamentally different from prior approaches that either ignore the geometric structure of representations entirely or use unsupervised geometric measures that this paper proves are insufficient for real-world steering tasks.
How It Works
The method takes as input a language model's internal representations (embeddings) from a specific layer and a set of data points aligned to a target task. For each data point, the system computes the pairwise distances to all other points in the representation space, creating a distance matrix. The Supervised Shesha variants then measure geometric stability by tracking whether these pairwise distance relationships remain consistent—essentially asking: if point A was close to point B before the model shifts, is it still close afterward? The task-aligned version crucially supervises this stability measurement using task labels, ensuring the distances being tracked are actually relevant to the behavior you're trying to predict or control. The output is a stability score that predicts both (1) whether a steering intervention will work and (2) whether the model's internal structure has degraded, with scores validated against actual steering outcomes and structural corruption across multiple models and tasks.
Production Impact
For teams deploying language models in safety-critical applications, this provides a concrete pre-deployment test that predicts steering susceptibility without requiring expensive adversarial experiments. You could run Supervised Shesha on a candidate model variant before deployment to get a numerical confidence estimate (ρ = 0.89-0.97 means you catch steering failures 89-97% of the time across test cases). The framework also enables monitoring in production: continuous geometric stability tracking could alert teams when a model's internal structure begins degrading, which might precede obvious behavioral failures. However, there are trade-offs: computing pairwise distances across all embeddings for each data point scales quadratically with dataset size, requiring careful batching in high-throughput systems. The essential requirement for task alignment means you cannot use a single pre-computed stability score—you need to re-measure for each steering target, adding computational overhead but ensuring accuracy on realistic scenarios.
Limitations and When Not to Use This
The paper's most significant limitation is that it only demonstrates effectiveness on 35-69 embedding models and three NLP tasks—the generalization to large language models (7B+ parameters), other modalities (vision, multimodal), or downstream tasks beyond those tested remains unclear. The quadratic scaling of pairwise distance computation becomes prohibitive for very large embedding dimensions or datasets, and the paper doesn't address optimization strategies for production-scale deployments. The requirement for task supervision means this approach cannot automatically detect all forms of degradation; it only detects degradation relevant to the specific tasks you've aligned to, creating blind spots if new failure modes emerge outside those task boundaries. Additionally, the paper identifies that unsupervised stability fails catastrophically but doesn't provide guidance on how much supervised data is needed to achieve reliable task alignment, or what happens in low-data regimes where steering itself is already difficult.
Research Context
This work builds on the broader research into mechanistic interpretability and representation geometry, which has shown that model behavior correlates with structure in learned representations. It extends prior work on geometric properties of embeddings (e.g., anisotropy, class separability) by showing that distance structure specifically predicts behavioral control—a more direct connection than previous proxy metrics. The paper's finding that supervised alignment dramatically outperforms unsupervised metrics challenges the assumption in some interpretability work that geometry alone, without task information, fully explains model behavior. This opens a research direction toward unified frameworks that treat multiple deployment concerns (controllability, robustness, degradation detection) as different facets of a single geometric property, rather than requiring separate probes or monitoring systems for each.
:::tip Subscribe Get weekly breakdowns of papers like this in AI Letters - the newsletter for engineers building production AI systems. :::