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

2026-04-30 · Source: stat.ML updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Emerging Technologies & Innovation · Depth: Expert, extended

Summary

A new diagnostic framework, "Geometric Canary," introduces supervised and unsupervised variants of Shesha, a geometric stability metric, to address two critical language model deployment challenges: predicting steerability and detecting representational drift. Supervised Shesha, which measures task-aligned geometric stability, accurately predicts linear steerability with near-perfect accuracy (Spearman's $\rho=0.89$–$0.97$) across 35-69 embedding models and three NLP tasks, capturing unique variance beyond class separability (partial $\rho=0.62$–$0.76$). Conversely, unsupervised Shesha, which measures intrinsic representational consistency, fails for steering prediction on real-world tasks ($\rho\approx 0.10$) but excels at drift detection. It measures nearly $2\times$ greater geometric change than CKA during post-training alignment (up to $5.23\times$ in Llama), provides earlier warning in 73% of models, and maintains a $6\times$ lower false alarm rate than Procrustes. This dissociation highlights that task alignment is crucial for controllability prediction, while intrinsic consistency is vital for post-deployment monitoring.

Key takeaway

For research scientists developing or deploying large language models, understanding representational geometry is crucial. You should integrate supervised Shesha into your pre-deployment evaluation pipeline to predict a model's linear steerability, especially for applications requiring fine-grained behavioral control. Post-deployment, continuously monitor unsupervised Shesha to detect subtle representational drift, as it offers earlier and more reliable warnings than traditional metrics like CKA or Procrustes, preventing alarm fatigue and ensuring model integrity.

Key insights

Geometric stability, measured by task-aligned and task-agnostic Shesha variants, predicts LLM steerability and detects representational drift.

Principles

• Task alignment is essential for predicting model controllability.
• Intrinsic geometric consistency is vital for detecting structural degradation.
• Supervised contrastive training enhances steerability and geometric rigidity.

Method

Shesha quantifies representational self-consistency by correlating RDMs from complementary views. Supervised Shesha uses label-derived RDMs for task alignment; unsupervised Shesha splits embedding dimensions for intrinsic consistency.

In practice

• Use supervised Shesha pre-deployment for steerability assessment.
• Employ unsupervised Shesha post-deployment for drift monitoring.
• Prioritize supervised contrastive models for steerable applications.

Topics

• Geometric Canary
• Representational Stability
• LLM Steerability
• Representational Drift Detection
• Supervised Shesha

Code references

• prashantcraju/geometric-canary

Best for: Research Scientist, AI Scientist, Machine Learning Engineer, MLOps Engineer

Related on AIssential

• A Geometric View of SRC: Learning Representations for Stable Residual Inference — SRC reliability hinges on representation geometry, which can be optimized without direct SRC training.
• Relational Rank Geometry in Transformers: Detecting and Steering Hidden-State Relation Frames — Transformers encode multi-argument relations as detectable and steerable rank-indexed hidden-state geometries.
• You Have Been Doing Geometry All Along — Data science workflows are fundamentally geometric, shaping how models perceive and learn structure.
• Sensor-Conditioned Representation Learning via Scene-Relevant Observation Quotients — Sensor-conditioned representations require evaluation beyond predictive utility, focusing on preserving sensing-justified scene distinctions in their latent geometry.
• When Probing Accuracy Saturates, Fragility Resolves: A Complementary Metric for LLM Pre-Training Analysis — Fragility, a new metric, reveals LLM pre-training dynamics invisible to standard probing accuracy.
• Structure Retention in Embedding Spaces as a Predictor of Benchmark Performance — Embedding space geometry, specifically local structure retention and linear separability, strongly predicts model performance.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by stat.ML updates on arXiv.org.