Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Deliberate Planning in Language Models with Symbolic Representation

Siheng Xiong, Zhangding Liu, Jieyu Zhou, Yusen Su

TL;DR

SymPlanner tackles the chronic planning failures of large language models by grounding planning in a deterministic symbolic world model. The framework couples a policy model, a symbolic simulator, and a discriminator, augmented by Iterative Correction and Contrastive Ranking to repair invalid actions and compare candidate plans. On PlanBench, SymPlanner consistently outperforms natural language based baselines, especially on long-horizon tasks, demonstrating improved validity, coherence, and execution fidelity. This symbol grounded, deliberative planning approach bridges language reasoning with formal planning, enabling robust, verifiable multi-step action sequences in structured domains.

Abstract

Planning remains a core challenge for large language models (LLMs), particularly in domains that require coherent multi-step action sequences grounded in external constraints. We introduce SymPlanner, a novel framework that equips LLMs with structured planning capabilities by interfacing them with a symbolic environment that serves as an explicit world model. Rather than relying purely on natural language reasoning, SymPlanner grounds the planning process in a symbolic state space, where a policy model proposes actions and a symbolic environment deterministically executes and verifies their effects. To enhance exploration and improve robustness, we introduce Iterative Correction (IC), which refines previously proposed actions by leveraging feedback from the symbolic environment to eliminate invalid decisions and guide the model toward valid alternatives. Additionally, Contrastive Ranking (CR) enables fine-grained comparison of candidate plans by evaluating them jointly. Conceptually, SymPlanner operationalizes two cognitive faculties: (i) error monitoring and repair via externalized feedback (IC) and (ii) preference formation among alternatives via pairwise comparison (CR), advancing cognitively plausible, symbol-grounded planning aligned with the rich structure in intelligent systems. We evaluate SymPlanner on PlanBench, demonstrating that it produces more coherent, diverse, and verifiable plans than pure natural language baselines.

Deliberate Planning in Language Models with Symbolic Representation

TL;DR

SymPlanner tackles the chronic planning failures of large language models by grounding planning in a deterministic symbolic world model. The framework couples a policy model, a symbolic simulator, and a discriminator, augmented by Iterative Correction and Contrastive Ranking to repair invalid actions and compare candidate plans. On PlanBench, SymPlanner consistently outperforms natural language based baselines, especially on long-horizon tasks, demonstrating improved validity, coherence, and execution fidelity. This symbol grounded, deliberative planning approach bridges language reasoning with formal planning, enabling robust, verifiable multi-step action sequences in structured domains.

Abstract

Planning remains a core challenge for large language models (LLMs), particularly in domains that require coherent multi-step action sequences grounded in external constraints. We introduce SymPlanner, a novel framework that equips LLMs with structured planning capabilities by interfacing them with a symbolic environment that serves as an explicit world model. Rather than relying purely on natural language reasoning, SymPlanner grounds the planning process in a symbolic state space, where a policy model proposes actions and a symbolic environment deterministically executes and verifies their effects. To enhance exploration and improve robustness, we introduce Iterative Correction (IC), which refines previously proposed actions by leveraging feedback from the symbolic environment to eliminate invalid decisions and guide the model toward valid alternatives. Additionally, Contrastive Ranking (CR) enables fine-grained comparison of candidate plans by evaluating them jointly. Conceptually, SymPlanner operationalizes two cognitive faculties: (i) error monitoring and repair via externalized feedback (IC) and (ii) preference formation among alternatives via pairwise comparison (CR), advancing cognitively plausible, symbol-grounded planning aligned with the rich structure in intelligent systems. We evaluate SymPlanner on PlanBench, demonstrating that it produces more coherent, diverse, and verifiable plans than pure natural language baselines.
Paper Structure (29 sections, 7 equations, 1 figure, 2 tables, 1 algorithm)

This paper contains 29 sections, 7 equations, 1 figure, 2 tables, 1 algorithm.

Table of Contents

  1. Introduction
  2. Methodology
  3. Problem definition
  4. System design
  5. Action proposal (generation).
  6. Symbolic simulation (validation & transition).
  7. Budgets and widths.
  8. Planning process.
  9. Why not just a classical planner?
  10. Iterative Correction (IC)
  11. Mechanism.
  12. Contrastive Ranking (CR)
  13. Mechanism.
  14. Experiments
  15. Experimental Setup
  16. ...and 14 more sections

Figures (1)

  • Figure 1: Overview of SymPlanner. The language agent consists of a generator and a discriminator, which interact with a symbolic environment. The generator proposes candidate actions, which are refined via iterative correction using feedback from the environment. The discriminator then applies contrastive ranking over simulated outcomes to select the most coherent successor state. On the right, an example blocksworld task illustrates how invalid actions are pruned (red crosses) and valid trajectories are retained until the goal configuration is achieved (green check).