acceptodds
Under review as a conference paper at ICLR 2027

ReHoPER: Receding-Horizon Planning for Enhanced Reasoning

Abstract

We propose ReHoPER, an inference-only, zero-shot method that improves large language models’ reasoning by generating and answering intermediate questions along multiple paths before the final answer. It iteratively plans a horizon of candidate intermediate questions, selects one to answer, and replans from the updated history. ReHoPER is task-agnostic, using the same generic instructions across datasets and models without labeled data or task-specific prompt design. Across multiple datasets, including a new controlled benchmark for compositional reasoning, ReHoPER outperforms strong baselines, with the largest gains in the most compositional settings. We will release our implementation and the iLLC generator to support future work.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Machine Learning, Large Language Models, Compositional Reasoning

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_WevWOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes RoHoper, a general prompting strategy that can help in reasoning. Instead of following a fixed plan, the model generates several intermediate questions, answers one of them, and then updates its plan based on the new information. Then it uses multiple reasoning paths and majority voting to produce the final answer. The proposed method shows improvement on 3 datasets across many LLMs.

Critical strengths
  1. The proposed method is simple and easy to use. It does not require task-specific training or examples. The authors also test it on multiple LLMs.

  2. Detailed analysis. The paper provides many useful analyses, including how the number of reasoning paths and steps affects performance.

  3. Covered different prompting strategies. The paper covers several common prompting strategies.

Critical weaknesses
  1. Limited novelty. I have some concerns about the novelty of the proposed method. Replanning is already a common idea in LLM agents: the agent first makes a plan, executes one step, and then updates its plan based on new information. ReHoPER mainly applies this idea to generating and answering intermediate questions. However, I think generating new intermediate questions is also a form of planning, so the main idea does not seem very inspiring. I suggest comparing against agentic planning methods, such as RAFA [1], or other methods that dynamically update their plans. I understand this paper also takes the way about how to aggregate these paths as a new idea, but I believe it is still worth exploring if deleting the question regeneration and making it an easy replanning will make any difference.

  2. Limited diversity of reasoning benchmarks. The paper evaluates on MuSR, MoreHopQA, and iLLC (a new dataset from the author). While testing many models is a strength, I think the number and diversity of datasets are still too limited. In particular, iLLC mainly tests string operations, which may not be that useful. Actually, for the other prompting papers, they usually include many other tasks, such as mathematical and symbolic reasoning. Moreover, the improvements on MoreHopQA are not always positive across models and prompting strategies (Table 1). Besides, based on the LLM's ability, I think evaluating on more challenging QA benchmarks on some advanced LLMs, such as BrowseComp-Plus [2], would make the paper more impactful, especially if the method can show improvements on these more complex tasks.

[1] RAFA: Reason for Future, Act for Now. International Conference on Machine Learning (ICML 2024). https://proceedings.mlr.press/v235/liu24ab.html

[2] BrowseComp-Plus: A More Fair and Transparent Evaluation Benchmark of Deep-Research Agent. https://arxiv.org/abs/2508.06600

Questions for authors
  1. Does replanning really improve the reasoning process, or does it mainly increase answer diversity? Compared with Self-Consistency, I wonder whether the improvements come from better reasoning plans or simply from generating more diverse answers. I think this is an interesting question worth further discussion and analysis, especially to see if this could be extended in self-improvement.

  2. How well does ReHoPER work without Self-Consistency? In practice, we often do not use multiple reasoning paths or majority voting because of the inference cost. What happens if ReHoPER uses only one reasoning path and returns the final answer directly, without majority voting?

4 comments
AuthorsOpenReview ↗

Part 1 / 2

We thank the reviewer for their positive assessment of the clarity, simplicity, and soundness of our approach, and for recognizing its training-free nature, evaluation across multiple LLMs, detailed analysis of reasoning paths and steps, and comparisons with various prompting strategies. We address the concerns below and will incorporate the feedback into the paper.


W1. We agree that replanning and majority voting are established ideas. Our contribution is not either component in isolation. Rather, ReHoPER studies their integration into a task-agnostic, inference-only reasoning procedure. It plans intermediate questions, answers one before replanning from the accumulated question-answer history, and aggregates goal-question predictions across both intermediate steps and independent paths. Our paper already compares ReHoPER against several established approaches, including Chain-of-Thought, Plan-and-Solve, Self-Consistency, Self-Refine, Self-Discover, and Least-to-Most prompting.

While RAFA is related to our work in its planning mechanism, it addresses a different problem setting. RAFA relies on external state transitions and reward feedback to learn environmental dynamics and optimize actions, whereas ReHoPER solves a fixed question using only the input context and model-generated intermediate answers, without external feedback. RAFA also uses task-specific prompts and configurations, so a direct comparison would require adapting its planning and feedback mechanisms to our setting. We will better clarify these differences in the paper.

Based on your feedback, we also conducted additional experiments comparing ReHoPER against Tree of Thoughts (ToT) across six models on MuSR-MM and MoreHopQA. The results are shown below:

ModelMuSR MM ToT (%)MuSR MM ReHoPER+IO (%)MoreHopQA ToT (%)MoreHopQA ReHoPER+IO (%)
Gemma 358.061.254.064.7
gpt-oss68.076.030.078.7
Llama 3.1 8B55.662.448.757.3
Mistral Small66.473.260.071.3
Phi-471.274.859.377.3
Qwen369.672.066.068.0
Average64.869.953.069.6

ReHoPER outperforms ToT in all 12 comparisons, with average improvements of 5.1 pp on MuSR-MM and 16.6 pp on MoreHopQA, or 10.8 pp overall. These results provide additional evidence of ReHoPER's effectiveness compared to an established search-based reasoning approach.

Finally, Section 5.2 (Table 2) already evaluates the suggested simplifications. -RP generates a fixed list of questions without replanning, while -LA replans by generating only the next question based on the updated history. Both use the same prediction aggregation procedure as ReHoPER. The full approach outperforms the stronger ablation by 1.9/2.7/2.5 pp under IO/CoT/PS, showing that combining look-ahead with replanning improves performance beyond aggregation alone.


W2. We use the same zero-shot planning instructions across three benchmark families that test different forms of multi-step reasoning. MuSR tests narrative deduction, object tracking, and allocation under constraints. MoreHopQA combines evidence across passages and adds arithmetic, symbolic, or commonsense steps (Schnitzler et al., 2024). iLLC extends the last-letter task of Wei et al. (2022) by varying the target character position and the number of words, giving us a controlled test of how models compose a sequence of operations. Focused evaluations of this size have precedents in Reasoning via Planning (Hao et al., 2023), Everything of Thoughts (Ding et al., 2024), and Tree of Thoughts (Yao et al., 2023), which each evaluate three main tasks.

On MoreHopQA, ReHoPER improves average accuracy over Self-Consistency (SC) under all three prompting styles. The gains are 0.8, 2.6, and 2.3 percentage points for input-output, chain-of-thought, and Plan-and-Solve, respectively (Table 1).

To extend the QA evaluation, we followed your suggestion and ran a preliminary experiment on BrowseComp-Plus (Chen et al., 2026). We first took all examples, but some inputs exceeded the models’ context windows even when we supplied only the relevant documents. We then evaluated the 250 instances (a test set comparable in size to those in the submission) with the shortest combined question and document context to avoid this issue.

ModelSC+IOReHoPER+IOΔ (pp)
gpt-oss76.0%82.8%+6.8
Llama 3.1 8B91.6%92.8%+1.2
Qwen396.0%97.6%+1.6

With the relevant documents supplied, Self-Consistency already performs well on this subset, exceeding 90% accuracy on two of the three models. ReHoPER improves accuracy on all three. For the final paper, we will evaluate a harder subset of BrowseComp-Plus or a more challenging QA dataset to test transfer in a setting with more room for improvement.

AuthorsOpenReview ↗

Part 2 / 2

Q1. We thank the reviewer for this interesting question. We believe the improvements come from both better reasoning and answer aggregation, rather than simply increased answer diversity. In ReHoPER, each prediction is generated using only the original input and the intermediate question-answer history of its own path. The model does not have access to previous predictions from the same path or predictions from other paths. Therefore, later predictions cannot benefit from observing earlier predictions.

As shown in Figure 2 (top row) in the paper, the mean correctness of individual predictions, or Current-Step Mean (CSM), generally increases as more intermediate questions are answered, even before applying any majority voting. This suggests that the evolving reasoning process improves the quality of individual predictions, independently of aggregation.

Furthermore, Table 2 directly examines the contribution of planning. When we remove replanning or look-ahead while keeping the prediction and aggregation procedures unchanged, performance generally decreases. Full ReHoPER outperforms the stronger ablation by 1.9, 2.7, and 2.5 percentage points under IO, CoT, and PS, respectively. These findings support the benefit of adaptive replanning beyond simply generating more diverse answers. We will clarify this distinction and better highlight these analyses in the paper.


Q2. ReHoPER can also operate with a single reasoning path, returning its final prediction directly without majority voting. In fact, our analysis in Figure 2 and Appendix B already provides evidence for this setting through Current-Step Mean (CSM), which measures the accuracy of individual paths without aggregating their predictions. Since the paths are independent, this corresponds to the expected accuracy of a single-path configuration.

Across the 63 model-dataset-prefix configurations analyzed, the final CSM achieves 55.5% accuracy, compared with 64.6% for the full cumulative-voting approach. While aggregation provides an additional performance boost, Figure 2 also shows that individual predictions generally improve as more intermediate questions are answered, even without voting. This suggests that ReHoPER's reasoning process can improve individual predictions independently of Self-Consistency. In addition, using only one path also eliminates the inference cost of the other five paths, providing a lower-cost alternative. We will clarify this distinction and better highlight these results in the revised paper.


Once again, we appreciate your detailed feedback and hope our revisions and clarifications address your concerns. Thank you for contributing to the improvement of our work.

Thank you for the detailed response and the additional experiments.

  1. On novelty: At each step, ReHoPER generates the full list of remaining intermediate questions, answers only the first one, and then regenerates the list based on the accumulated question-answer history. This is essentially receding-horizon planning: plan the full trajectory, execute only the first step, observe the result, and replan. RAFA follows the same idea. The main difference is that RAFA relies on external state and reward feedback, while ReHoPER relies on the model’s own evaluation. I agree this is a difference in the problem setting, but I do not think it is a fundamental difference in the method.

  2. On BrowseComp-Plus: Since Self-Consistency already performs very well on this subset, it would be more convincing to also try newer and stronger models on harder instances, where there is more room for improvement.

My other concerns have been addressed, and I appreciate the clarifications. Overall, I think this is a good paper with a simple and practical method and thorough analysis. I will keep my positive score.

We thank the reviewer for the follow-up and for their positive assessment of our work.


1. We agree that ReHoPER and RAFA share the receding-horizon planning principle, which we do not claim as novel. Our contribution is applying this principle to task-agnostic, inference-only reasoning through intermediate question answering and prediction aggregation, without external state transitions or reward feedback. Following your suggestion, we also compared ReHoPER against RAFA adapted to our setting across six models:

ModelMuSR MM RAFA (%)MuSR MM ReHoPER+IO (%)MoreHopQA RAFA (%)MoreHopQA ReHoPER+IO (%)
Gemma 360.461.251.364.7
gpt-oss64.876.075.378.7
Llama 3.1 8B55.662.451.357.3
Mistral Small60.473.262.771.3
Phi-463.674.870.777.3
Qwen364.472.062.768.0
Average61.569.962.369.6

ReHoPER outperforms RAFA in all 12 comparisons, with average improvements of 8.4 pp on MuSR MM and 7.2 pp on MoreHopQA (7.8 pp overall). Together with our ToT comparisons and planning ablations, these results further support the effectiveness of our approach. We will include these results and better clarify the differences from RAFA in the paper.


2. We agree that evaluating stronger models on harder BrowseComp-Plus instances would provide a more convincing comparison. Our preliminary experiments focused on instances with shorter contexts to avoid exceeding the models' context windows. We will extend this evaluation to harder instances with stronger models in the revised work.


We appreciate your constructive feedback and are glad that our previous response addressed your other concerns.

Reviewer_7VKmOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
3
Soundness
3
Read review
Paper summary

The paper proposes a prompt framework that performs well on several benchmarks. The method works by breaking the task down into sub questions which are answered, punctuated with periodic replanning steps.

Critical strengths

The paper makes connections to classical control theory and appears to use principles from the literature to improve model results.

Critical weaknesses

While the control theory connection is nice to have, the overall method does not seem to be rooted in a well-developed theory of why this method should work better than other decomposed prompting methods. More importantly, I worry that this method will not stand the test of time. More up-to-date language models than the ones tested (even ones from >6 months ago) already have quite strong planning/long task abilities, and I wouldn't be surprised that any gains from this paper's method are already superseded by improved post-training methods. I think you can already see some of this, for instance enabling thinking on Qwen3 appears to have almost wiped out the gains from the method (+8.7 percentage points to +1.1 for IO). I also worry that, given the abundance of existing work on prompting frameworks for decomposable tasks, this paper might not have much to add in terms of scientific understanding of language models or practical applications.

Questions for authors

The paper is clearly written, so I do not have much in the way of questions. My critique is more about the pursuit of prompting frameworks as a scientific endeavor at this point in time.

1 comment

We thank Reviewer 7VKm for noting the clarity of our presentation and highlighting how ReHoPER draws on classical control theory. We address the concerns below and will incorporate them into the paper.


When and Why Replanning Helps

While we do not claim that the theoretical guarantees of classical control theory directly apply to LLM reasoning, ReHoPER is motivated by the benefits of feedback-driven planning. The key intuition is that the most useful next question may depend on the answers to previous questions. Unlike fixed decomposition methods, ReHoPER uses the updated reasoning history to revise its plan after each intermediate answer. This combines look-ahead planning with the flexibility to adapt as new information becomes available.

Our experiments support this motivation. In Table 2, ReHoPER outperforms the stronger planning ablation by 1.9-2.7 pp across prompting styles. Appendix F confirms that replanning frequently changes the reasoning trajectory, while Appendix E shows that the method generates more unique intermediate questions than planning without look-ahead, despite using fewer questions overall. Although these results do not establish a formal theoretical guarantee, they provide empirical support for the proposed mechanism. We will clarify this motivation in the revised paper.


Relevance of ReHoPER as Language Models Improve

While stronger native reasoning capabilities can reduce the gains from inference-time methods, our results suggest that ReHoPER remains complementary to these advances rather than being superseded by them.

First, enabling thinking on Qwen3 substantially improves the SC$_{61}$ (Self-Consistency) baseline from 61.5% to 79.1% (+17.6 pp), reducing the remaining room for improvement from 38.5 to 20.9 pp. Thus, smaller additional gains are expected. Importantly, ReHoPER still improves accuracy by +1.1 pp on average across all eight subsets, even with thinking enabled, suggesting that explicit planning can complement native reasoning capabilities.

More importantly, motivated by the reviewer's feedback, we conducted additional experiments with two recent reasoning-capable models: Gemma-4-31B (April 2026) and Qwen3.8-Flash-Next 125B (August 2026).

ModelDatasetSC$_{61}$ (%)ReHoPER (%)Δ (pp)
Gemma-4-31BMuSR MM70.072.8+2.8
MoreHopQA67.378.0+10.7
Qwen3.8-Flash-Next 125BMuSR MM74.077.6+3.6
MoreHopQA76.780.0+3.3
Average72.077.1+5.1

ReHoPER improves performance across all four settings, with an average gain of +5.1 pp, including a substantial +10.7 pp on MoreHopQA with Gemma 4. These results provide further evidence that ReHoPER remains effective with newer models, even as their underlying reasoning capabilities improve.

We agree that as models become stronger and existing benchmarks approach saturation, evaluating more challenging compositional tasks becomes increasingly important. Our iLLC benchmark supports this direction by allowing systematic increases in reasoning difficulty. Overall, these findings suggest that ReHoPER is not rendered obsolete by advances in post-training, but provides a complementary, training-free mechanism for improving reasoning, which can be applied directly to newer models without task-specific adaptation.


Scientific Contributions and Practical Relevance of ReHoPER

Although many prompting frameworks have been proposed for decomposable tasks, most of them rely on task-specific demonstrations, handcrafted reasoning structures, or manual adaptation. In contrast, ReHoPER uses the same generic instructions across different tasks and models without requiring such modifications. We compare against general-purpose approaches, including Self-Consistency and Self-Discover, and show improvements on average.

More importantly, our work investigates how different planning strategies affect LLM reasoning. Our ablations (Table 2) demonstrate the benefit of combining look-ahead planning with iterative replanning, while our trajectory analyses (Appendices E and F) show how replanning changes reasoning paths and reduces repeated intermediate questions. Our aggregation analysis (Figure 2) further shows that predictions often improve as reasoning progresses, even before voting. Finally, our compute-accuracy analysis (Table 4) demonstrates that allocating computation to structured reasoning can be substantially more effective than simply increasing the number of independent samples.

We believe these findings go beyond proposing another prompting framework and provide useful insights into how to improve reasoning through inference-time planning, which remains relevant as LLMs continue to advance.


Once again, we appreciate your feedback and hope our revisions and clarifications address your concerns. Thank you for contributing to the improvement of our work.

Reviewer_Cz66OpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
3
Soundness
1
Read review
Paper summary

This paper proposes ReHoPER, an inference-only LLM reasoning workflow on complex question answering. Given the task input, ReHoPeR first issues generate an initial prediction based on the context and the question. It then runs n stochastic reasoning paths, each of them is organized as an iterative plan-select-answer-update loop. At each step, the LLM plans a horizon of candidate intermediate questions, answers only the first one not yet asked, and appends the question-answer pair to the path history before replanning. After each intermediate question is answered, the LLM also produces a new prediction for the original question conditioned on the updated history. Finally, the initial prediction and all per-step prediction of the original question aggregated across all reasoning paths are considered together to produce the final answer with majority vote.

Critical strengths
  1. The proposed ReHoPER workflow is simple and intuitive.

  2. This paper evaluates how increasing the number of reasoning steps and the number of reasoning paths affects final performance.

  3. The evaluation results also provide useful insight into how to allocate inference-time computation between deeper reasoning and more diverse paths.

Critical weaknesses
  1. Although the workflow is simple and easy to understand, it is more computationally complex than existing solutions since it needs replicated predictions. This may require an extra number of LLM calls. For example, answering a complex question with m steps takes O(m·n+1) predictions.

  2. The contribution of this paper is somewhat incremental. Some techniques introduced in this paper have already been investigated and proposed in previous research. For example, planning multiple reasoning paths over a short horizon, executing one step, and replanning from the updated state has been adopted in LLMPC. Dynamic replanning from history has also been widely investigated in reflection-related papers. Aggregating predictions across different reasoning paths is the key idea of Self-Consistency.

  3. This paper lacks a detailed comparison with widely used LLM reasoning workflows, such as Tree-of-Thoughts [1], ReAct [2], and Reflexion [3], ReWOO [4], and ADaPT[5].

[1] Yao et al., Tree of Thoughts: Deliberate Problem Solving with Large Language Models (Neurips 2023)

[2] Yao et al., ReAct: Synergizing Reasoning and Acting in Language Models (ICLR 2023)

[3] Shinn et al., Reflexion: language agents with verbal reinforcement learning (Neurips 2023)

[4] Xu et al., ReWOO: Decoupling Reasoning from Observations for Efficient Augmented Language Models

[5] Prasad et al., ADAPT: As-Needed Decomposition and Planning with Language Models (NAACL 2024 findings)

Questions for authors
  1. What will happen if we scale up ReAct or other reasoning workflows with extra reasoning steps/paths?
2 comments
AuthorsOpenReview ↗

Part 1 / 2

We thank the reviewer for their positive assessment of the clarity and simplicity of our approach, and for recognizing the value of our analysis of reasoning steps, paths, and their implications for allocating inference-time computation.


W1. Using additional inference-time computation to improve reasoning is well established in the literature, including several methods cited by the reviewer (e.g., Tree-of-Thoughts and Reflexion). We believe the key consideration is therefore not simply the number of LLM calls, but how effectively the additional computation is used.

To address this, Section 5.6 (Table 4) explicitly compares ReHoPER against Self-Consistency with 181 samples (SC181) to account for its additional inference cost. The results show that increasing SC samples provides limited improvements, whereas ReHoPER achieves substantially higher accuracy, often with lower runtime. For example, on iLLC L2-4, ReHoPER improves Qwen3 accuracy from 70.5% to 83.6% while reducing runtime from 1:44 to 1:20. With Phi-4, accuracy increases from 10.8% to 59.3%, with runtime decreasing from 2:01 to 1:18. These results demonstrate that ReHoPER uses additional inference-time computation more effectively than simply increasing the number of independent predictions. We will clarify this trade-off and better emphasize our compute-matched comparisons in the final version.


W2. We appreciate the reviewer's observation that ReHoPER builds on ideas explored in prior research. However, building on existing methods and investigating what their combination makes possible is a natural part of research. We believe ReHoPER makes a distinct contribution beyond these individual methods, as each cited method covers a limited part of ReHoPER’s procedure.

  • LLMPC

    • LLMPC carries a single trajectory forward per run, while ReHoPER advances multiple independent trajectories for the same question.

    • It omits the cumulative history of answered subquestions that ReHoPER gives both the planner and the answer predictor.

    • LLMPC’s trip- and meeting-planning experiments use task-specific prompts, worked examples, and constraint checkers. By comparison, ReHoPER uses the same zero-shot reasoning instructions across MuSR, MoreHopQA, and iLLC and builds each path’s history from the test instance itself.

  • Reflexion

    • Reflexion tailors its prompts and feedback to each task type and leaves transfer to new tasks with the same instructions untested, while ReHoPER demonstrates this reuse across three benchmarks.

    • Reflexion’s sequential retries omit ReHoPER’s independent reasoning paths and its vote over predictions from every step across those paths.

    • Its reflection loop also omits the explicit look-ahead that ReHoPER uses to plan upcoming subquestions and replan after each answer within the same attempt.

    • Table 3 tests a related approach, Self-Refine, which has similar limitations as Reflexion. Averaged across four models and six subsets, ReHoPER+PS achieves 62.9% accuracy compared with 43.0% for Self-Refine, a gain of 19.9 pp. The gap grows to 33.5 points on the harder iLLC L2 subsets.

  • Self Consistency (SC)

    • SC leaves reasoning to the underlying prompt and votes over one final answer per sampled trace. The procedure adds no explicit planning, look-ahead, replanning, or updated history. ReHoPER, by comparison, uses its growing history to decide what to ask next and to produce a fresh prediction at each step.

    • We already compare against SC as a main baseline and show that ReHoPER makes better use of inference compute. Even with SC’s larger sampling budget, ReHoPER achieves 14.9 pp higher accuracy on average across the four models and four iLLC subsets in Table 4.

We will better clarify these differences in the revised paper and explain how combining these mechanisms enables ReHoPER to achieve the observed improvements.

AuthorsOpenReview ↗

Part 2 / 2

W3. Thank you for the suggestion. We would like to clarify that our paper already compares ReHoPER against several established prior approaches, including Chain-of-Thought, Plan-and-Solve, Self-Consistency, Self-Refine, Self-Discover, and Least-to-Most prompting.

Based on your feedback, we also conducted additional experiments comparing ReHoPER against Tree of Thoughts (ToT) across six models on MuSR-MM and MoreHopQA. The results are shown below:

ModelMuSR MM ToT (%)MuSR MM ReHoPER+IO (%)MoreHopQA ToT (%)MoreHopQA ReHoPER+IO (%)
Gemma 358.061.254.064.7
gpt-oss68.076.030.078.7
Llama 3.1 8B55.662.448.757.3
Mistral Small66.473.260.071.3
Phi-471.274.859.377.3
Qwen369.672.066.068.0
Average64.869.953.069.6

ReHoPER outperforms ToT in all 12 comparisons, with average improvements of 5.1 pp on MuSR-MM and 16.6 pp on MoreHopQA, or 10.8 pp overall. These results provide additional evidence of ReHoPER's effectiveness compared to an established search-based reasoning approach.

Regarding the other mentioned methods, ReAct, Reflexion, ReWOO, and ADaPT use different reasoning and planning mechanisms, some of which rely on tool use or external feedback rather than the zero-shot, tool-free question-answering setting we study. We will expand our related-work discussion to better clarify these differences and include the new ToT results in the paper.


Q1. We agree that increasing the number of steps or paths could also improve the performance of other methods. However, additional computation does not necessarily lead to proportional or even consistent improvements. For example, Self-Refine (Madaan et al., 2023, Figure 4) shows diminishing returns from additional iterations. On Code Optimization, the score improves from 22.0 to 27.0 after one iteration, but only reaches 28.8 after three. The authors also report cases where additional refinements actually reduce output quality (Appendix K, Table 15).

We already investigate this question for SC in Section 5.4 (Figure 3) and Section 5.6 (Table 4). As discussed in W1, increasing SC from 61 to 181 samples provides only marginal improvements, while ReHoPER achieves substantially higher accuracy, often with lower runtime. For example, on iLLC L2-4 with Phi-4, SC181 achieves 10.8% accuracy in 2:01, compared to 59.3% in 1:18 for ReHoPER. This suggests that our improvements come from how we structure the reasoning process, rather than simply increasing the inference-time computation.

A systematic scaling analysis of ReAct and other methods, each with different reasoning mechanisms and computational requirements, would be a separate research study. Moreover, as discussed in W3, several of these methods involve tool interactions or external feedback, which makes the comparison between their reasoning steps and paths and our zero-shot, tool-free setting less straightforward. We will clarify these distinctions and the scope of our analysis in the revised paper.


Once again, we appreciate your feedback and we hope that our revisions and clarifications address your concerns. Thank you for contributing to the improvement of our work.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.