ACE: A Minimum-Energy Criterion for Entropy-Controlled Language Generation Modeling Semantic Stability and Structural Grounding in Large Language Models

Read full text Summarise with VIRAT
Authors: ERNESTO ROSATI BERISTAIN
Cited by: 0

Summary

No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.

Summarise with VIRAT Free to try with your Science Catalogs account.

Abstract

This work introduces the Axiomatic Criterion Engine (ACE), a framework that models language generation in Large Language Models (LLMs) as a minimum-energy optimization process under semantic constraints. Unlike standard probabilistic decoding, ACE incorporates structural penalties that account for: contextual disorder conceptual misalignment knowledge inconsistency and structural grounding (origin cost) The extended energy formulation: E'(z | x, A, K) = λ_p (-log P(z|x)) + λ_h H_c(x) + λ_a D(z,A) + λ_k C(z,K) + λ_o O(z) introduces a novel term O(z), representing the degree to which a response is anchored to a shared semantic premise. This formulation provides a mechanistic explanation of hallucination, showing that instability emerges from: contextual disorder (H_c) lack of structural grounding (O) Empirical observations across multiple LLMs (ChatGPT, Gemini, Grok) demonstrate: consistent reduction in token usage (~25–35%) improved semantic coherence convergence toward stable attractor states From an agnostic and instrumental perspective, ACE reveals a general principle: semantic generation requires structural grounding to achieve stability and efficiency This work connects probabilistic language modeling with energy-based systems and semantic entropy research, offering a unified and practical framework for improving reliability in LLMs without modifying model weights.

Fields of Study & Subjects

Topic Modeling Natural Language Processing Techniques Adversarial Robustness in Machine Learning

Publication details

Journal / source
Zenodo (CERN European Organization for Nuclear Research)
Publication date
22 March 2026
Language
EN
Document type
Preprint