Why Reasoning Fails to Plan: A Planning-Centric Analysis of Long-Horizon Decision Making in LLM Agents

Zehong Wang, Fang Wu, Hongru Wang, Xiangru Tang, Bolian Li, Zhenfei Yin, Yijun Ma, Yiyang Li, Weixiang Sun, Xiusi Chen, Yanfang Ye

Abstract

Large language model (LLM)-based agents exhibit strong step-by-step reasoning capabilities over short horizons, yet often fail to sustain coherent behavior over long planning horizons. We argue that this failure reflects a fundamental mismatch: step-wise reasoning induces a form of step-wise greedy policy that is adequate for short horizons but fails in long-horizon planning, where early actions must account for delayed consequences. From this planning-centric perspective, we study LLM-based agents in deterministic, fully structured environments with explicit state transitions and evaluation signals. Our analysis reveals a core failure mode of reasoning-based policies: locally optimal choices induced by step-wise scoring lead to early myopic commitments that are systematically amplified over time and difficult to recover from. We introduce FLARE (Future-aware Lookahead with Reward Estimation) as a minimal instantiation of future-aware planning to enforce explicit lookahead, value propagation, and limited commitment in a single model, allowing downstream outcomes to influence early decisions. Across multiple benchmarks, agent frameworks, and LLM backbones, FLARE consistently improves task performance and planning-level behavior, frequently allowing LLaMA-8B with FLARE to outperform GPT-4o with standard step-by-step reasoning. These results establish a clear distinction between reasoning and planning.

Introduction

Large Language Models (LLMs) have demonstrated remarkable proficiency in "Chain-of-Thought" reasoning, effectively breaking down complex problems into immediate next steps. However, a persistent paradox remains: these models often struggle with long-horizon tasks requiring a sequence of coherent decisions. This paper investigates this failure, proposing that the root cause lies in the fundamental difference between step-wise reasoning and long-term planning.

The Myopic Trap: Reasoning vs. Planning

Standard LLM reasoning typically operates as a "greedy" local policy. When an agent reasons about its next action, it selects the most plausible step based on immediate local information. In complex environments, such actions may appear optimal in the short term but lead to irreversible dead ends later. This "myopic trap" occurs because standard agents lack a mechanism to allow future consequences to inform present decisions.

The authors provide theoretical and empirical evidence that step-wise greedy reasoning is inherently sub-optimal for long-horizon tasks. Even techniques like Beam Search, which increase the breadth of exploration, often fail because they still rely on local scoring metrics that do not account for delayed rewards.

FLARE: Future-aware Lookahead with Reward Estimation

To address these limitations, the researchers introduced FLARE, a framework designed to incorporate essential planning principles into LLM agents:

  1. Explicit Lookahead: Instead of selecting only the immediate next step, FLARE simulates potential future trajectories (rollouts) using principles from Monte Carlo Tree Search (MCTS).
  2. Backward Value Propagation: The outcomes of these simulated futures are propagated back to the current decision point. If a path leads to failure, the value of the initial action is penalized.
  3. Limited Commitment (Receding Horizon): The agent executes only the first step of the best-simulated plan and then replans from the new state, maintaining flexibility in the face of imperfect simulations.

To maintain computational efficiency, FLARE employs Action Pruning (limiting candidate actions to high-quality suggestions from the LLM) and Trajectory Memory (caching simulations to reduce token consumption).

Experimental Evaluation

The framework was tested on Knowledge Graph Question Answering (KGQA) datasets (CWQ, WebQSP, GrailQA) and tool-use environments (ALFWorld).

Performance across planning horizons and trap avoidance rates for CWQ and WebQSP datasets.

As shown in Figure 5, FLARE (Lookahead) consistently outperforms both Single Step reasoning and Beam Search across varying planning horizons. Notably, the "Trap select @step1" analysis reveals that FLARE is significantly less prone to picking locally attractive but globally harmful actions compared to greedy baselines.

Key Results and Insights

  • Performance Gains: FLARE significantly improved success rates. For instance, on the CWQ dataset, performance rose from 59.8% (Single Step) to 71.8%.
  • Error Recovery: A critical finding is FLARE's ability to recover from early mistakes. While standard reasoning agents rarely recover (approx. 5% rate), FLARE increases this to nearly 30% by identifying errors during lookahead simulations.
Analysis of prefix correctness and success rates after the first error for CWQ and WebQSP.

Figure 4 illustrates that FLARE maintains a "still-correct prefix" for longer and demonstrates a much higher success rate even after an initial error has occurred, highlighting its robustness compared to myopic reasoning.

  • Model Scale vs. Architecture: Remarkably, a LLaMA-3.1-8B model using FLARE can outperform GPT-4o using standard reasoning. This suggests that for long-horizon tasks, the planning architecture can be more impactful than the raw scale of the model.

Conclusion

This research clarifies the distinction between reasoning and planning in LLM agents. By implementing minimal planning requirements—lookahead, value propagation, and limited commitment—FLARE enables agents to overcome the myopic traps inherent in step-wise reasoning. Future work will explore extending these principles to stochastic environments and amortizing the cost of planning through model fine-tuning.

Summary updated

Abstract

Large language model (LLM)-based agents exhibit strong step-by-step reasoning capabilities over short horizons, yet often fail to sustain coherent behavior over long planning horizons. We argue that this failure reflects a fundamental mismatch: step-wise reasoning induces a form of step-wise greedy policy that is adequate for short horizons but fails in long-horizon planning, where early actions must account for delayed consequences. From this planning-centric perspective, we study LLM-based agents in deterministic, fully structured environments with explicit state transitions and evaluation signals. Our analysis reveals a core failure mode of reasoning-based policies: locally optimal choices induced by step-wise scoring lead to early myopic commitments that are systematically amplified over time and difficult to recover from. We introduce FLARE (Future-aware Lookahead with Reward Estimation) as a minimal instantiation of future-aware planning to enforce explicit lookahead, value propagation, and limited commitment in a single model, allowing downstream outcomes to influence early decisions. Across multiple benchmarks, agent frameworks, and LLM backbones, FLARE consistently improves task performance and planning-level behavior, frequently allowing LLaMA-8B with FLARE to outperform GPT-4o with standard step-by-step reasoning. These results establish a clear distinction between reasoning and planning.
BibTeX
@misc{wang2026why,
  author = {Zehong Wang and Fang Wu and Hongru Wang and Xiangru Tang and Bolian Li and Zhenfei Yin and Yijun Ma and Yiyang Li and Weixiang Sun and Xiusi Chen and Yanfang Ye},
  title = {{Why Reasoning Fails to Plan: A Planning-Centric Analysis of Long-Horizon Decision Making in LLM Agents}},
  year = {2026},
  eprint = {2601.22311},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2601.22311},
}
Figure