Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Programming over Thinking: Efficient and Robust Multi-Constraint Planning

Derrick Goh Xin Deik, Quanyu Long, Zhengyuan Liu, Nancy F. Chen, Wenya Wang

TL;DR

This work introduces SCOPE, a scalable code-based framework for robust multi-constraint planning that decouples query-specific reasoning from generic code execution. It formalizes problem understanding into structured representations and automatically generates reusable solver functions, enabling deterministic and reusable planning across queries. Empirical results on TravelPlanner and Natural Plan benchmarks show state-of-the-art performance with lower inference cost and latency, particularly benefiting weaker models, and demonstrate robustness to rising problem complexity. Overall, SCOPE presents a modular, automated approach that advances reliable planning in real-world tasks by separating reasoning from execution and leveraging executable solvers.

Abstract

Multi-constraint planning involves identifying, evaluating, and refining candidate plans while satisfying multiple, potentially conflicting constraints. Existing large language model (LLM) approaches face fundamental limitations in this domain. Pure reasoning paradigms, which rely on long natural language chains, are prone to inconsistency, error accumulation, and prohibitive cost as constraints compound. Conversely, LLMs combined with coding- or solver-based strategies lack flexibility: they often generate problem-specific code from scratch or depend on fixed solvers, failing to capture generalizable logic across diverse problems. To address these challenges, we introduce the Scalable COde Planning Engine (SCOPE), a framework that disentangles query-specific reasoning from generic code execution. By separating reasoning from execution, SCOPE produces solver functions that are consistent, deterministic, and reusable across queries while requiring only minimal changes to input parameters. SCOPE achieves state-of-the-art performance while lowering cost and latency. For example, with GPT-4o, it reaches 93.1% success on TravelPlanner, a 61.6% gain over the best baseline (CoT) while cutting inference cost by 1.4x and time by ~4.67x. Code is available at https://github.com/DerrickGXD/SCOPE.

Programming over Thinking: Efficient and Robust Multi-Constraint Planning

TL;DR

This work introduces SCOPE, a scalable code-based framework for robust multi-constraint planning that decouples query-specific reasoning from generic code execution. It formalizes problem understanding into structured representations and automatically generates reusable solver functions, enabling deterministic and reusable planning across queries. Empirical results on TravelPlanner and Natural Plan benchmarks show state-of-the-art performance with lower inference cost and latency, particularly benefiting weaker models, and demonstrate robustness to rising problem complexity. Overall, SCOPE presents a modular, automated approach that advances reliable planning in real-world tasks by separating reasoning from execution and leveraging executable solvers.

Abstract

Multi-constraint planning involves identifying, evaluating, and refining candidate plans while satisfying multiple, potentially conflicting constraints. Existing large language model (LLM) approaches face fundamental limitations in this domain. Pure reasoning paradigms, which rely on long natural language chains, are prone to inconsistency, error accumulation, and prohibitive cost as constraints compound. Conversely, LLMs combined with coding- or solver-based strategies lack flexibility: they often generate problem-specific code from scratch or depend on fixed solvers, failing to capture generalizable logic across diverse problems. To address these challenges, we introduce the Scalable COde Planning Engine (SCOPE), a framework that disentangles query-specific reasoning from generic code execution. By separating reasoning from execution, SCOPE produces solver functions that are consistent, deterministic, and reusable across queries while requiring only minimal changes to input parameters. SCOPE achieves state-of-the-art performance while lowering cost and latency. For example, with GPT-4o, it reaches 93.1% success on TravelPlanner, a 61.6% gain over the best baseline (CoT) while cutting inference cost by 1.4x and time by ~4.67x. Code is available at https://github.com/DerrickGXD/SCOPE.
Paper Structure (47 sections, 5 figures, 5 tables, 1 algorithm)

This paper contains 47 sections, 5 figures, 5 tables, 1 algorithm.

Table of Contents

  1. Introduction
  2. Related Work
  3. Methodology
  4. Overview
  5. Query-Specific Problem Reasoning
  6. Problem Formalization.
  7. Problem Optimization.
  8. Generic Solver Generation
  9. Solver Construction.
  10. Solver Refinement.
  11. Inference
  12. Experiments
  13. Datasets
  14. Language Models
  15. Baselines
  16. ...and 32 more sections

Figures (5)

  • Figure 1: The overview workflow of SCOPE. The workflow consists of 2 stages: Query-Specific Problem Reasoning (top-half) and Generic Solver Generation (bottom-half). Different colors of arrows represent different workflow construction: (red, structured representations), (green, structured solution), (gray, solver functions).
  • Figure 2: The output of agents at two different stages of SCOPE and inference.
  • Figure 3: The success rate (y-axis) across different level of complexity (x-axis). Top to bottom (Different benchmarks): TravelPlanner, Trip Planning, Meeting Planning. Left to right (Different models): GPT-5, GPT-o3, GPT-4o, Gemini-2.5-Pro, Gemini-1.5-Pro.
  • Figure 4: The cost/time required per query (y-axis) for different methods across different level of complexity (x-axis). Top to bottom (Different models): GPT-5, GPT-o3, GPT-4o, Gemini-2.5-Pro, Gemini-1.5-Pro. Left to right (Different metrics): Cost and time taken for TravelPlanner, Trip Planning, and Meeting Planing.
  • Figure 5: The inference cost required per query to reach correct solution. Top to bottom (Different metrics): Number of calls, time taken, number of tokens. Left to right (Different datasets): Travel Planner, Trip Planning, Meeting Planning. Infinity means an answer cannot be correctly achieved after many repeated attempts.