Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

ALAS: A Stateful Multi-LLM Agent Framework for Disruption-Aware Planning

Edward Y. Chang, Longling Geng

TL;DR

ALAS introduces a stateful, modular framework that decomposes planning into role-specific agents with persistent memory to enable disruption-aware planning. The three-layer architecture (workflow template, agent factory, and runtime monitor) supports localized compensation via LRCP, reducing the need for costly global replanning. Across URS, event coordination, and job-shop scheduling benchmarks, ALAS delivers strong static and dynamic performance, with superior feasibility, reduced makespan, and robust disruption handling compared to monolithic LLMs and classical solvers. The approach demonstrates practical impact for large-scale, time-sensitive planning in logistics, coordination, and manufacturing, enabling transparent audit trails, rollback, and cross-domain generalization.

Abstract

Large language models (LLMs) excel at rapid generation of text and multimodal content, yet they falter on transaction-style planning that demands ACID-like guarantees and real-time disruption recovery. We present Adaptive LLM Agent System (ALAS), a framework that tackles four fundamental LLM deficits: (i) absence of self-verification, (ii) context erosion, (iii) next-token myopia, and (iv) lack of persistent state. ALAS decomposes each plan into role-specialized agents, equips them with automatic state tracking, and coordinates them through a lightweight protocol. When disruptions arise, agents apply history-aware local compensation, avoiding costly global replanning and containing cascade effects. On real-world, large-scale job-shop scheduling benchmarks, ALAS sets new best results for static sequential planning and excels in dynamic reactive scenarios with unexpected disruptions. These gains show that principled modularization plus targeted compensation can unlock scalable and resilient planning with LLMs.

ALAS: A Stateful Multi-LLM Agent Framework for Disruption-Aware Planning

TL;DR

ALAS introduces a stateful, modular framework that decomposes planning into role-specific agents with persistent memory to enable disruption-aware planning. The three-layer architecture (workflow template, agent factory, and runtime monitor) supports localized compensation via LRCP, reducing the need for costly global replanning. Across URS, event coordination, and job-shop scheduling benchmarks, ALAS delivers strong static and dynamic performance, with superior feasibility, reduced makespan, and robust disruption handling compared to monolithic LLMs and classical solvers. The approach demonstrates practical impact for large-scale, time-sensitive planning in logistics, coordination, and manufacturing, enabling transparent audit trails, rollback, and cross-domain generalization.

Abstract

Large language models (LLMs) excel at rapid generation of text and multimodal content, yet they falter on transaction-style planning that demands ACID-like guarantees and real-time disruption recovery. We present Adaptive LLM Agent System (ALAS), a framework that tackles four fundamental LLM deficits: (i) absence of self-verification, (ii) context erosion, (iii) next-token myopia, and (iv) lack of persistent state. ALAS decomposes each plan into role-specialized agents, equips them with automatic state tracking, and coordinates them through a lightweight protocol. When disruptions arise, agents apply history-aware local compensation, avoiding costly global replanning and containing cascade effects. On real-world, large-scale job-shop scheduling benchmarks, ALAS sets new best results for static sequential planning and excels in dynamic reactive scenarios with unexpected disruptions. These gains show that principled modularization plus targeted compensation can unlock scalable and resilient planning with LLMs.
Paper Structure (92 sections, 2 theorems, 13 equations, 10 figures, 20 tables, 2 algorithms)

This paper contains 92 sections, 2 theorems, 13 equations, 10 figures, 20 tables, 2 algorithms.

Table of Contents

  1. Introduction
  2. Related Work
  3. Structural Limitations of LLMs
  4. Multi-Agent Planning Systems
  5. LLM-Based Planning and Benchmarks
  6. LLM-Based Planning.
  7. Planning Benchmarks.
  8. The ALAS Three-Layer Architecture for Adaptive and Reactive Planning
  9. Workflow Blueprinting Layer: Template Construction
  10. Agent Factory Layer: Agent Sourcing and Compensation Logic
  11. Runtime Layer: Execution and Reactive Adaptation
  12. Disruption Handling.
  13. Master Coordinator and Memory Logging.
  14. Experimental Evaluation
  15. Metrics:
  16. ...and 77 more sections

Key Result

Lemma 1

For a system with: The worst-case time complexity is:

Figures (10)

  • Figure 1: Network $G=(V,E)$ with urban travel times $\tau_{ij}=10$ minutes and airport routes distance specified on the figure. Static scenarios can be solved by MILP or Column Generation. Dynamic scenarios (e.g., an accident, a cancellation, new passenger requests) must be addressed by $\mathsf{Alas}$.
  • Figure 2: Comparison of ride-sharing solutions generated by $\mathsf{Alas}$ and baseline LLMs. (Left) Mean total travel distance (km) with standard deviation error bars over 10 independent runs for each method, illustrating $\mathsf{Alas}$'s improved efficiency. (Right) Optimal schedule generated by $\mathsf{Alas}$ for the URS task, utilizing two vehicles ($k_1, k_2$) to serve four passengers ($r_1$-$r_4$).
  • Figure 3: Mean Gap to Upper Bound comparison across two benchmark datasets
  • Figure 4: LRCP Phase #1 Local Compensation (makespan = 22): (a) Static baseline schedule; (b) $M_1$ failure between $t = 5$–$8$; (c) $M_1$ notifies $M_2$ to delay $J3(2)$; (d) $M_2$ informs $M_0$ to push $J3(3)$ back.
  • Figure 5: LRSP Phase #2 Queue Reordering (makespan = 22): (a) Safe moves: moving last operations down, first operations forward with potential penalty; (b) Resolving remaining operations.
  • ...and 5 more figures

Theorems & Definitions (3)

  • Lemma 1: Generalized LCSR Complexity
  • proof
  • Corollary 1: Special Cases