Two Ways to Stabilize LLM Agents on Complex Tasks: Hierarchical Planning and CAP-CoT

Two recent papers examine a similar weakness in LLM-based agents from different angles. “From Coarse to Fine: Self-Adaptive Hierarchical Planning for LLM Agents” focuses on planning in dynamic, multi-step tasks and argues that many current agents rely on plans with a fixed level of detail, which can be too coarse for hard tasks and too detailed for simple ones. “CAP-CoT: Cycle Adversarial Prompt for Improving Chain of Thoughts in LLM Reasoning” looks at reasoning instead, starting from the observation that Chain-of-Thought prompting can become unstable on long, multi-step problems and may produce inconsistent answers even when the task does not change. Together, the papers ask how LLM agents can become more stable when both planning and reasoning stretch across many steps. [S9][S11] [S9] [S11]

What the papers are about

The first paper, “From Coarse to Fine: Self-Adaptive Hierarchical Planning for LLM Agents,” was released on arXiv in April 2026. Its stated target is LLM-based agents operating in dynamic and multi-step environments, where planning is needed to guide long-term action. The paper’s core problem statement is that existing planning methods often work at a fixed granularity level, which creates a mismatch between the complexity of the task and the detail of the plan. [S9]

The second paper, “CAP-CoT: Cycle Adversarial Prompt for Improving Chain of Thoughts in LLM Reasoning,” also appeared on arXiv in April 2026. It focuses on reasoning rather than action planning. Its starting point is that Chain-of-Thought prompting is useful for step-by-step solutions, but can be unstable across runs on long, multi-step problems, leading to inconsistent answers for the same task. [S11]

Sources: [S9], [S11]

Core idea: adaptive planning granularity and iterative correction

The central idea in the hierarchical planning paper is straightforward: not every task should be planned at the same resolution. According to the abstract, fixed-granularity planning has a basic limitation. If the plan is always very detailed, simple tasks may be over-specified. If the plan is always high-level, complex tasks may not get enough structure. The proposed direction is a self-adaptive hierarchical planning approach that moves from coarse to fine, adjusting the level of detail to the task rather than assuming one planning scale fits all cases. My interpretation is that this is an attempt to make planning more proportionate: broad structure first, then finer decomposition where complexity actually requires it. [S9]

CAP-CoT addresses a different failure mode. Instead of asking only how to generate a forward reasoning chain once, it emphasizes iterative and contrastive correction. The paper proposes a cycle-based adversarial prompt approach for improving Chain-of-Thought reasoning. From the abstract alone, the key point is that prior work has focused more on strengthening a single forward pass, while this paper shifts attention toward repeated correction using a cycle. In plain terms, the idea is not to trust the first reasoning trace as final, but to pressure-test and revise it through an adversarial, repeated prompting process. My interpretation is that this aims to reduce answer instability by making reasoning self-corrective rather than purely one-shot. [S11]

Sources: [S9], [S11]

How they differ from existing approaches

The hierarchical planning paper differs from existing planning methods by explicitly criticizing fixed granularity. The source states that current approaches often operate at one fixed level of detail and therefore fail to match the needs of tasks with different complexity. Its contribution, at least at the abstract level, is to replace that fixed planning scale with a self-adaptive hierarchical one. This is a structural change in how plans are formed: the issue is not only what plan is chosen, but at what level of abstraction planning happens. [S9]

CAP-CoT differs from standard CoT use in a different way. The paper argues that much prior work tries to improve the forward reasoning chain within a single pass. CAP-CoT instead emphasizes iterative and contrastive correction through a cycle adversarial prompt. So the shift is from single-pass reasoning to repeated examination and revision. In other words, where conventional CoT often assumes the main job is to produce a better initial chain, CAP-CoT treats instability itself as a target and introduces a mechanism meant to revisit the chain rather than simply extend it. [S11]

Sources: [S9], [S11]

Where these ideas may matter in practice

Based on the problem settings described in the abstracts, the hierarchical planning approach is most relevant to agents that must act over multiple steps in dynamic environments. When a task unfolds over time and conditions may change, a plan that is too rigidly detailed or too vague can both become problematic. A self-adaptive hierarchy could therefore be useful in settings where an agent needs both long-term structure and local flexibility. This is an interpretation of the paper’s framing, not a claim about proven deployment outcomes. [S9]

CAP-CoT appears most relevant to tasks that require long reasoning chains, especially when the same prompt may yield inconsistent outputs across runs. In such cases, a cycle-based correction process could matter wherever reliability of multi-step reasoning is more important than getting a first-pass answer quickly. Again, the source supports the motivation around long, multi-step reasoning instability; any concrete application domain beyond that would need evidence from the full paper. [S11]

Taken together, the two papers suggest a useful division of labor for complex LLM systems: one line of work tries to stabilize how an agent organizes actions over time, while the other tries to stabilize how it constructs and revises reasoning traces. That does not mean they solve the same problem, but they address adjacent sources of brittleness in multi-step tasks. [S9][S11]

Sources: [S9], [S11]

Limitations and open questions

Both papers begin from real weaknesses, but neither should be read as a complete solution. In the case of hierarchical planning, the abstract clearly identifies the limitation of fixed granularity and proposes self-adaptive hierarchy as a remedy. Still, the source summary alone does not establish that adaptive planning removes all planning failures in dynamic environments. A remaining question is how reliably an agent can decide when to stay coarse and when to refine into more detail, especially as tasks change over time. That concern follows from the paper’s own framing around granularity mismatch. [S9]

For CAP-CoT, the paper directly addresses instability in Chain-of-Thought reasoning and proposes iterative, adversarial correction. But repeated correction also raises practical questions that are not resolved by the abstract alone: when should the cycle stop, how often does revision genuinely improve reasoning rather than merely alter it, and how stable is the correction process itself across long problems? These are not criticisms of the paper’s premise; they are open issues implied by the fact that the method is designed for unstable multi-step reasoning in the first place. [S11]

There is also a broader shared limitation. One paper focuses on planning structure, the other on reasoning correction. In real agents, planning and reasoning interact. A better plan does not guarantee correct reasoning at each step, and a better reasoning cycle does not guarantee that the overall plan is set at the right level of abstraction. So the two approaches look complementary, but the available source material does not support saying that they fully resolve instability in complex LLM agents. [S9][S11]

Sources: [S9], [S11]

One-paragraph takeaway

These two papers approach the same broad problem—unstable performance on complex, multi-step tasks—from different layers of the stack. “From Coarse to Fine” argues that planning fails when agents are forced to use one fixed granularity for every task, and proposes self-adaptive hierarchical planning to better match plan detail to task complexity. “CAP-CoT” argues that Chain-of-Thought reasoning can be inconsistent on long problems, and proposes a cycle-based adversarial prompting method that emphasizes iterative correction rather than a single forward pass. Read together, they suggest that making LLM agents more dependable may require both better task decomposition and better mechanisms for revising reasoning, while still leaving open questions about when and how those mechanisms should be applied. [S9][S11]

Sources: [S9], [S11]

---

One-line takeaway: One paper targets unstable planning by adapting plan granularity, while the other targets unstable reasoning by adding cycle-based correction to Chain-of-Thought. [S9][S11] [S9] [S11]

Short summary: Two recent papers tackle instability in LLM agents from different directions: one adjusts planning granularity, and the other revises reasoning through cycle-based correction. Together they show why complex multi-step tasks need more than a single fixed plan or a single-pass Chain-of-Thought.

Sources and references: - [S9] cs.AI updates on arXiv.org - From Coarse to Fine: Self-Adaptive Hierarchical Planning for LLM Agents - URL: https://arxiv.org/abs/2604.23194 - [S11] cs.AI updates on arXiv.org - CAP-CoT: Cycle Adversarial Prompt for Improving Chain of Thoughts in LLM Reasoning - URL: https://arxiv.org/abs/2604.23270

Internal link ideas: - How planning granularity affects LLM agent behavior in multi-step tasks - Why Chain-of-Thought becomes unstable on long reasoning problems - A practical guide to evaluating reliability in LLM agent workflows

LLM agents #hierarchical planning #chain-of-thought #reasoning stability #paper brief

---

Note AI-assisted content
This post was drafted with AI (gpt-5.4) using source-grounded inputs.
Please review the citations and original links below.