Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Entropic-Time Inference: Self-Organizing Large Language Model Decoding Beyond Attention

Andrew Kiruluta

TL;DR

This work introduces a self\-organizing inference architecture that jointly couples scheduling, attention sparsification, and sampling temperature under a unified entropy control objective, and transforms inference into a resource\-intelligent thermodynamic process that allocates computation where uncertainty reduction is maximized.

Abstract

Modern large language model (LLM) inference engines optimize throughput and latency under fixed decoding rules, treating generation as a linear progression in token time. We propose a fundamentally different paradigm: entropic\-time inference, where decoding is governed by the flow of uncertainty rather than token index. We introduce a self\-organizing inference architecture that jointly couples scheduling, attention sparsification, and sampling temperature under a unified entropy control objective. Our method extends vLLM with entropy-aware scheduling, entropic pruning of paged attention blocks, and adaptive temperature control that stabilizes generation near a target entropy regime. This transforms inference into a resource\-intelligent thermodynamic process that allocates computation where uncertainty reduction is maximized. We present a concrete systems design, pseudocode, and integration plan, demonstrating how entropy can serve as a first\-class control signal for scalable LLM inference.

Entropic-Time Inference: Self-Organizing Large Language Model Decoding Beyond Attention

TL;DR

This work introduces a self\-organizing inference architecture that jointly couples scheduling, attention sparsification, and sampling temperature under a unified entropy control objective, and transforms inference into a resource\-intelligent thermodynamic process that allocates computation where uncertainty reduction is maximized.

Abstract

Modern large language model (LLM) inference engines optimize throughput and latency under fixed decoding rules, treating generation as a linear progression in token time. We propose a fundamentally different paradigm: entropic\-time inference, where decoding is governed by the flow of uncertainty rather than token index. We introduce a self\-organizing inference architecture that jointly couples scheduling, attention sparsification, and sampling temperature under a unified entropy control objective. Our method extends vLLM with entropy-aware scheduling, entropic pruning of paged attention blocks, and adaptive temperature control that stabilizes generation near a target entropy regime. This transforms inference into a resource\-intelligent thermodynamic process that allocates computation where uncertainty reduction is maximized. We present a concrete systems design, pseudocode, and integration plan, demonstrating how entropy can serve as a first\-class control signal for scalable LLM inference.
Paper Structure (112 sections, 1 theorem, 50 equations, 1 figure, 1 table)

This paper contains 112 sections, 1 theorem, 50 equations, 1 figure, 1 table.

Table of Contents

  1. Introduction
  2. Operational time and entropy flow.
  3. Inference as an optimization problem.
  4. Toward self-organizing inference.
  5. Novelty and scope.
  6. Practical considerations.
  7. Entropic-Time Principle
  8. Entropy dynamics during decoding.
  9. Definition of entropic time.
  10. Inference efficiency as an information–resource tradeoff.
  11. Thermodynamic interpretation.
  12. Entropy Estimation and Computational Cost
  13. Cost model.
  14. Top-$k$ entropy approximation.
  15. Tail-corrected estimator.
  16. ...and 97 more sections

Key Result

Theorem 1

Under Assumptions ass:lipschitz_entropy--ass:local_slope, choose $\eta$ such that $0<\eta\mu<1$. Then for all trajectories remaining in the neighborhood $\mathcal{N}$, Equivalently, the log-temperature converges exponentially to an $\epsilon_H/\mu$-radius neighborhood of $x^*$.

Figures (1)

  • Figure 1: Entropic-time inference architecture as a coupled control system. Entropy drives macro scheduling, meso memory interaction, and micro sampling control, forming a closed feedback loop.

Theorems & Definitions (1)

  • Theorem 1: Local ISS robustness to bounded entropy error