HoloAegis: Frozen Representation, Topological Inference: Minimally Parametric Safety Manifolds for Zero-Shot LLM Guardrails

2026-08-09

AI Summary

This work addresses the dilemma faced by current large language models, where safety alignment either distorts semantic representations through fine-tuning or incurs high inference costs. The authors propose a training-free geometric safety mechanism that freezes the pretrained encoder and maps text embeddings onto the unit hypersphere. Leveraging a precomputed library of topological anchor points, the method performs zero-shot safety judgments via Gibbs–Boltzmann free energy and a dual-timescale exponential moving average, effectively decoupling representation learning from inference. Requiring only a few fixed hyperparameters, the approach significantly enhances robustness against high-frequency perturbations and achieves state-of-the-art performance across eight benchmarks—e.g., AuthenHallu AUC = 1.0000 and HarmBench AUC = 0.9802—while offering sub-millisecond latency, zero cold-start overhead, and strong cross-lingual transferability, as demonstrated by CHIFRAUD AUC = 0.9758 on Chinese data.

Abstract

Current LLM safety guardrails face a fundamental tension: fine-tuning distorts pre-trained representations while generative judges incur prohibitive inference costs. We challenge the prevailing paradigm by asking: can safety be achieved through pure geometric reasoning over frozen semantic representations? We present HoloAegis, a minimally parametric topological inference framework that decouples representation from reasoning. We term our approach minimally parametric because the only free parameters are the anchor count K and the temperature tau, both fixed after construction and requiring no gradient-based training. An un-fine-tuned encoder maps text to a unit sphere, after which all decisions are purely geometric. We formalize safety evaluation as a Gibbs-Boltzmann Free Energy computation over a pre-computed System Topology Anchor Bank, and we introduce Dual Time-Scale Exponential Moving Averages to detect progressive multi-turn semantic drift. Our key theoretical insight is a Topological Boundary Stability Conjecture: we provide theoretical motivation and strong empirical evidence that sparse anchor centroids stabilize the decision boundary against high-frequency lexical perturbations far better than full vector space methods. Evaluated across 8 benchmarks, HoloAegis achieves state-of-the-art accuracy (1.0000 AUC on AuthenHallu, 0.9802 on HarmBench) with sub-millisecond latency, zero cold-start data, and cross-lingual transfer (0.9758 AUC on Chinese CHIFRAUD).

Problem

• LLM safety
• representation distortion
• inference cost
• guardrails
• semantic drift

Innovation

• topological inference
• frozen representations
• minimally parametric
• safety manifolds
• geometric reasoning