PlanGEN: A Multi-Agent Framework for Generating Planning and Reasoning Trajectories for Complex Problem Solving

TL;DR

PlanGEN introduces a model-agnostic, multi-agent framework for natural planning that couples constraint extraction, verification, and adaptive algorithm selection to improve inference-time reasoning. By integrating constraint-guided verification with three established inference-time strategies (Best of N, Tree-of-Thought, REBASE) and a dynamic mixture framework, PlanGEN achieves state-of-the-art results across NATURAL PLAN, OlympiadBench, DocFinQA, and GPQA, while adapting to instance complexity. Key findings show that the verification agent robustly correlates plan quality with outcomes and that the selection agent substantially boosts performance on complex tasks. The work demonstrates broad applicability and robustness across multiple LLMs, suggesting strong potential for scalable, interpretable planning in diverse domains.

Abstract

Recent agent frameworks and inference-time algorithms often struggle with complex planning problems due to limitations in verifying generated plans or reasoning and varying complexity of instances within a single task. Many existing methods for these tasks either perform task-level verification without considering constraints or apply inference-time algorithms without adapting to instance-level complexity. To address these limitations, we propose PlanGEN, a model-agnostic and easily scalable agent framework with three key components: constraint, verification, and selection agents. Specifically, our approach proposes constraint-guided iterative verification to enhance performance of inference-time algorithms--Best of N, Tree-of-Thought, and REBASE. In PlanGEN framework, the selection agent optimizes algorithm choice based on instance complexity, ensuring better adaptability to complex planning problems. Experimental results demonstrate significant improvements over the strongest baseline across multiple benchmarks, achieving state-of-the-art results on NATURAL PLAN ($\sim$8%$\uparrow$), OlympiadBench ($\sim$4%$\uparrow$), DocFinQA ($\sim$7%$\uparrow$), and GPQA ($\sim$1%$\uparrow$). Our key finding highlights that constraint-guided iterative verification improves inference-time algorithms, and adaptive selection further boosts performance on complex planning and reasoning problems.

Figures (12)

• Figure 1: Schematic representation of PlanGEN (Mixture of Algorithms). An initial plan and constraints guide iterative plan refinement. The verification agent provides reward scores for plan quality, and the selection agent chooses inference algorithms until the highest-reward plan is found and used for downstream reasoning (if needed). UCB: Upper Confidence Bound, BoN: Best of $\mathcal{N}$, ToT: Tree-of-Thought, RS: REBASE.
• Figure 2: Schematic representation of PlanGEN (Best of $\mathcal{N}$) (BoN).
• Figure 3: Schematic representation of PlanGEN (ToT). Highest-reward steps are highlighted in green.
• Figure 4: Schematic representation of PlanGEN (REBASE). Green shading indicates step reward (darker $=$ higher). Darker steps prioritized for exploration.
• Figure 5: Performance comparison of the proposed multi-agent frameworks against baselines across four benchmarks. All experiments are conducted using Gemini-1.5-Pro. Algo: Algorithms, MATH: Mathematics, PHY: Physics.