ARTFEED — Contemporary Art Intelligence

HoloAegis: Zero-Shot LLM Guardrails via Topological Safety Manifolds

ai-technology · 2026-08-11

A recent preprint on arXiv (2608.08485) presents HoloAegis, a framework designed for LLM safety with minimal parameters, functioning on static semantic representations without the need for fine-tuning. This method separates representation from reasoning by projecting text onto a unit sphere using an encoder that remains un-fine-tuned, relying solely on geometric decision-making. Safety assessments are defined through a Gibbs-Boltzmann Free Energy calculation based on a pre-established System Topology Anchor Bank, utilizing just two constant hyperparameters: anchor count K and temperature tau. Additionally, it employs Dual Time-Scale Exponential Moving Averages to identify gradual multi-turn semantic shifts. This approach challenges the traditional reliance on fine-tuning or generative judges, providing a zero-shot solution that preserves pre-trained representations and lowers inference expenses.

Key facts

  • HoloAegis is a minimally parametric topological inference framework for LLM safety guardrails.
  • It uses frozen semantic representations and requires no gradient-based training.
  • Only free parameters are anchor count K and temperature tau, fixed after construction.
  • Safety evaluation is based on Gibbs-Boltzmann Free Energy computation over a System Topology Anchor Bank.
  • Introduces Dual Time-Scale Exponential Moving Averages for detecting progressive multi-turn semantic drift.
  • The paper is available on arXiv with ID 2608.08485.
  • The approach decouples representation from reasoning, mapping text to a unit sphere.
  • It aims to address the tension between fine-tuning distortion and high inference costs of generative judges.

Entities

Institutions

  • arXiv

Sources