Feedback-Aware MCTS for Goal-Oriented Information Seeking

Thank you for thoroughly reviewing our paper and sharing your feedback.

We appreciate your positive assessment of the paper, particularly your recognition of the intuitive design of our MCTS modifications and the strength of our experimental validation across diverse tasks.

---

Addressing W1: Complexity and Runtime

Our paper includes a brief note on efficiency in Appendix G. We also address the computational complexity in detail in our response to Q4 by Reviewer 2p6X. In our setting, the main bottleneck is LLM call invocations, which dominate overall runtime and cost. Tree traversal and clustering/embedding operations are negligible by comparison (on average, cluster assignment takes about 0.007 seconds per data point).

To provide an estimate, we recorded the average wall-clock time (in seconds) per full MCTS-based question selection step (decision turn) for the following datasets in the closed set scenario with Llama 3.3 70B:

As the results in Table 1 of the paper show, MISQ-HF achieves substantial efficiency gains over UoT due to its selective expansion and feedback reuse. Due to compute and time constraints, we cannot report runtime for every setting, but will aim to include these numbers with standard deviation in the Appendix. We believe this further supports our primary focus on LLM call counts as the most relevant cost metric for our framework.

---

Addressing W2: Prior Work

We acknowledge the importance of comprehensively situating our work within the landscape of LLM-based MCTS research. The referred papers: Language Agent Tree Search (LATS) by [1]Zhou et al. and ExACT by [2]Yu et al. focus on open-ended, interactive agent tasks involving rich, unstructured environmental actions, where the LLM reasons about complex state transitions and performs future action simulations through language prompts. Their settings differ fundamentally from ours; MISQ-HF centers specifically on structured information-seeking dialogues with a predefined or dynamically constructed possibility set. It leverages entropy-based information gain as a quantitative reward to guide selective question generation within bounded domains - an approach that is both computationally efficient and principled.

Implementing LATS or ExACT directly in our setting would require redesigning our interaction model as open-ended agent-environment loops, adding complexity and computational overhead. Importantly, please note that our baseline MISQ (without HF) already captures a vanilla LLM+MCTS approach tailored to question selection, serving as a relevant comparison. Nonetheless, we will more clearly mention these connections and expand our related work section to discuss the relationships and boundaries between these methods and ours.

---

Response to Q1: Questions that maximize information gain

Our empirical analysis across LLM calls for question generation in the MD (DX, MedDG) and TS (FloDial) domains shows that the average ratio of affirmative-consistent possibilities to total possibilities is approximately 52.82% ±3.49%. This near-50% balance indicates that, when appropriately prompted, LLMs are reasonably effective at producing questions that closely approximate optimal information gain, as entropy peaks at balanced partitions. We will add this detail in the revised manuscript.

---

Response to Q2: Random rollout in MCTS Simulation

In traditional MCTS, random simulations (where actions from a given node are chosen at random during the rollout phase) are a foundational practice used to provide unbiased, tractable estimates of long-term utility and facilitate theoretical convergence [1,2,3]. This approach is not only computationally efficient but also widely adopted in influential systems like AlphaGo, where random or fast simulated rollouts underpin robust planning performance.

Recent LLM + MCTS frameworks like GDP‑Zero[4], RoT[5], and PG-TD[6] replace random rollouts with LLM-driven simulation to better capture the complexities of natural language tasks. GDP‑Zero leverages a single LLM to simulate both user and system responses during search, acting jointly as policy, value function, and world model for dialogue planning. RoT (Reflection on Trees) enhances tree search by using a strong LLM to reflect on past search trajectories and generate guidelines that improve future search accuracy in semantically rich domains. PG-TD employs LLM-driven code simulations, prompting the model to generate likely next code snippets at each step.

The objective and quantifiable nature of our reward (information gain measured by entropy reduction) allows us to rely on random rollouts during simulation. Since R_IG can be calculated directly from the current partition of the possibility set, this approach yields an objective and meaningful signal for the search tree without incurring repeated or costly LLM calls during simulation. It also ensures that each simulation step precisely reflects progress toward reducing uncertainty, which is crucial for efficient information-seeking. This method is consistent with the theoretical roots of MCTS, and we will add that future work could incorporate semantically-aware or language-based rollout policies if richer feedback or open-domain reasoning becomes essential for the task

---

[1]: Kocsis, L. & Szepesvári, C. “Bandit based Monte-Carlo Planning.” ECML 2006.

[2]: James, S., Konidaris, G., & Rosman, B. (2017). An Analysis of Monte Carlo Tree Search. Proceedings of the AAAI Conference on Artificial Intelligence, 31(1).

[3] Vodopivec, T., Samothrakis, S., & Ster, B. (2017). On monte carlo tree search and reinforcement learning. Journal of Artificial Intelligence Research, 60, 881-936.

[4] Yu, X., Chen, M., and Yu, Z. 2023. Prompt-Based Monte-Carlo Tree Search for Goal-oriented Dialogue Policy Planning. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing (EMNLP ’23). Association for Computational Linguistics, 7101–7125.

[5] Hui, W., and Tu, K. 2024. RoT: Enhancing Large Language Models with Reflection on Search Trees. arXiv:2404.05449 [cs].

[6] Zhang, S., Chen, Z., Shen, Y., Ding, M., Tenenbaum, J. B., and Gan, C. 2023. Planning with Large Language Models for Code Generation. In International Conference on Learning Representations (ICLR ’23).

We thank the reviewer for taking the time to acknowledge our rebuttal. We appreciate that their concerns and questions regarding runtime, MCTS-based baselines, and the ability of LLMs to generate information-seeking questions were addressed by our response.

We would like to highlight that our work focuses on feedback-guided planning for information-seeking, where we express efficiency via reduced LLM calls, which constitute the dominant cost in practice, considering both latency and resource demands of LLM inference. We have demonstrated wall-clock runtime estimates for completeness and aim to include these in the appendix. However, we note that our main metric for efficiency centers on reducing the number of expensive model invocations during planning (measured by QGC in Table 1). It offers a model-agnostic, reproducible signal of efficiency, consistent across different hardware and execution environments.

Importantly, as the tree expands over time with more data, both the number of calls and overall runtime decreases because the system increasingly reuses prior high-reward paths and only needs to traverse and update parts of the tree. This effect is driven by both the standard MCTS selection step (based on UCT) and our hierarchical feedback mechanism, contributing meaningfully to long-term efficiency gains.

We hope this clarifies the rationale behind our choice of evaluation and reinforces the practical value of our approach in resource-constrained settings. We thank the reviewer again for recognizing our contributions.

Thank you for thoroughly reviewing our paper and sharing your feedback. We appreciate your positive assessment of the paper and are encouraged by the recognition of our method’s novel contributions, clarity and technical soundness.

---

Addressing W1:

To strengthen the positioning of our approach with respect to existing literature, we can refine Section 2, Related Work, to more clearly articulate the limitations of prior work and how MISQ-HF fills these gaps. While we discuss UoT in the second paragraph of section 2, we will elaborate on its performance bottlenecks and how our proposed approach overcomes them. We would like to emphasize these conceptual distinctions:

• UoT performs exhaustive tree expansion up to a fixed depth, incurring high computational cost;
• MISQ-HF avoids this overhead by employing selective expansion through MCTS based on an exploration-exploitation tradeoff.
• UoT does not incorporate any form of feedback; MISQ-HF enables feedback-driven adaptation during inference time by using semantic similarity to generalize and improve tree search for new cases.

To make the distinctions more accessible, we will add a compact visual comparing the exhaustive tree expansion with MISQ-HF’s selective, feedback-driven search. We believe this will enhance the foundation of our approach.

---

Addressing W2:

Thank you for highlighting the need for an intuitive comparison. The distinction between MISQ-HF and UoT does not just lie in efficiently reaching the target, but also in how questions are selected during the search/planning phase. While the semantic form of the questions may not always be very different, the key distinction lies in when (depth-aware) and why (cluster-specific bonus) a question is prioritized, based on the combination of current information gain and historical feedback.

UoT performs an exhaustive expansion, relying solely on immediate information‐gain at each node. In contrast, MISQ-HF selectively focuses on the most promising subtrees, skipping low‐reward or semantically redundant branches.

Beyond efficiency, MISQ-HF uses feedback‐driven prioritization based on the depth‐decayed bonus derived from past successes in similar cases. Where UoT ‘forgets’ previous interactions and treats every decision anew, MISQ-HF remembers which early‐stage questions historically delivered the greatest narrowing of the solution space. For example, if identifying upper abdomen pain as an early question consistently leads to rapid and accurate diagnoses of gastric ulcers, MISQ-HF will accelerate the selection of such broad, high-impact questions in future, similar cases. In contrast, UoT does not retain or reuse information from previous dialogues, so it tends to ask less strategic or even repetitive questions. This learning and adaptation allow MISQ-HF to generate higher-quality, more generalizable question sequences tailored by experiential feedback.

In addition to discussing the above details, we will include a compact table of qualitative examples in Section 5, contrasting MISQ-HF vs UoT trajectories on the same input.

---

Response to Q1: Human judgements

We agree that incorporating human judgments would be valuable in assessing semantic appropriateness and human preference of questions. However, due to resource and time constraints, this was out of scope for the current version. That said, the core design idea behind the system is to reflect generalization across cases - which aligns with how humans tend to learn from repeatedly seeing similar examples. We will note this point in the Limitations section, highlighting the potential for future work on a case study comparing human and model-generated questions.

---

Response to Q2: Effort cost of questions

Thank you for raising this point. In our current setup, we do not explicitly model the effort cost of answering each question. However, our medical diagnosis prompt is carefully designed to elicit symptom-based questions, typically answerable via yes/no or brief open-ended responses, thus naturally avoiding high-effort or invasive queries. In contrast, our troubleshooting domain does include more situation-dependent questions, where the cost to respond could vary more.

While incorporating response cost remains out of scope for this work, we fully agree it’s a critical consideration for real-world deployment. One promising direction we are interested in exploring is to have the LLM assign a cost score to each question it generates, which could be integrated into the reward design or selection criteria. That said, real-world datasets often lack ground truth for high-effort questions, making such modeling challenging in practice. We will include this as a future extension in the paper.

---

Response to Q3: Figure showing closed-vs-open set

We will add some details to Figure 1 to highlight the difference between closed and open set scenarios. Specifically, around steps 3 and 8 in the diagram, we will visually separate the initialization of the possibility set: for the closed set, the system will start from a fixed, pre-defined set of outcomes, while for the open set, the system will begin with a dynamically generated or expanded set based on user input. We will also add clear legends and labels to explain this distinction to aid readers’ understanding.

---

Response to Q4: Complexity analysis

Our paper includes a brief complexity analysis in Appendix G: QGC Efficiency. As suggested, we will move it towards the end of Section 3. For clarity, we will focus on the derivations for theoretical grounding and add the following details to Appendix, adhering to page limits.

UoT uses exhaustive expansion to maximize information gain. At each question node $v$, it generates $m$ candidate questions and evaluates both branches, creating $b = 2m$ child nodes.

Total number of question nodes up to depth $d_s$:

$N_{\text{nodes}} = \sum_{k=0}^{d_s-1} b^k = \frac{b^{d_s} - 1}{b - 1}$

Each node requires one LLM call. So:

$C_{\text{UoT}} = O(b^{d_s}) = O((2m)^{d_s})$

This exponential growth reflects a scaling challenge.

MISQ-HF employs selective expansion via MCTS, balancing exploration and exploitation with UCT (Eq. 7; 13). Each iteration has:

1. Selection: Tree traversal (no LLM calls)
2. Expansion: ≤1 new node (≤1 LLM call)
3. Simulation: Random rollout to depth $d_s$ (≤$d_s$ LLM calls)
4. Backpropagation: Value updates ($O(d_s)$, no LLM calls)

Per iteration: ≤ $d_s + 1$ LLM calls

Total per decision: $C_{\text{MISQ}} \leq K(d_s + 1) = O(K \cdot d_s)$

Hierarchical Feedback: The bonus reward term $ B_k(v) = \beta \cdot R_{\text{total}}(v) \cdot \gamma^{d_v} $ adds efficiency without changing asymptotic complexity.

Cluster Assignment Complexity: Semantic clustering via embeddings requires $O(n_c)$ comparisons per sample, where $n_c$ is the number of clusters. This is amortized across turns and negligible compared to LLM cost.

Hence, MISQ-HF offers a key advantage: linear scaling $O(K \cdot d_s)$ vs exponential growth $O((2m)^{d_q})$ in exhaustive tree search.

---

Addressing Broader Impact:

We appreciate your encouragement to highlight the broader impacts of optimal question selection, especially around user expectations and social dynamics. We acknowledge that maximizing information gain does not always align with conversational norms, particularly in sensitive settings like healthcare or customer support. In practice, users may expect certain standard question sequences for reassurance or trust-building. They may also provide unprompted information or interpret omitted questions as inattentiveness. Skipping over familiar or obvious questions, even when optimal, can lead to confusion or make the dialog feel unnatural. Similarly, some questions are naturally perceived as coupled and are best asked together for reasons beyond efficiency.

In the MedDG dataset, we allow for open-ended responses where the user (simulator) often shares additional details. We use the classification prompt (Appendix D.5) to infer the response and update the possibility set accordingly, though this does not fully address all scenarios of spontaneous disclosure.

To reflect these considerations, we will expand our discussion in Appendix H: Broader Impacts, to emphasize the importance of balancing algorithmic optimality with user-centered design in the domain of information seeking. We would also propose that future work models spontaneous disclosures and adaptive conversational flows more comprehensively, allowing systems to integrate ancillary information and adjust strategies mid-dialogue.

Dear Reviewer,

We sincerely appreciate the time and effort you have taken to review our paper. We hope that our rebuttal effectively addresses your concerns and strengthens your confidence in our work. If there are any remaining questions or clarifications we can provide to support your evaluation, we would be happy to engage in a discussion.

Thank you again for recognizing our contributions and for a careful assessment. We believe that your feedback has been valuable in enhancing our paper's quality and broader impacts.

Best regards,

The authors

Thank you for thoroughly reviewing our paper and sharing your feedback.

We appreciate your positive recognition of the strengths of our work, including the clear problem formulation, architecture design, implementation, and extensive empirical validation across domains. We hope the clarifications below will adequately address your concerns.

---

W1: Baselines

We benchmarked our approach against three main baselines: Direct Prompting (DP), Uncertainty of Thoughts (UoT), and MISQ without the hierarchical feedback mechanism.

UoT (Hu et al., 2025) is the current state-of-the-art in this domain. We would like to highlight that the UoT paper includes rigorous evaluations against prior baselines such as Chain-of-Thought (CoT), Tree-of-Thought (ToT), and Reflexion on the same datasets we use. Their results show that UoT consistently outperforms these methods, which justifies our use of UoT as the strongest comparison. We cite these works accordingly in the Related Work and Baselines sections and did not duplicate experiments that were already well-covered by the state-of-the-art baseline.

In the Experiments section, our goal is to highlight the contribution of our hierarchical feedback mechanism by comparing against DP (a minimal-effort baseline), UoT (the strongest available method), and MISQ without feedback as an ablation. This allowed us to isolate the impact of our contribution while also keeping computational demands manageable.

---

W2 and Q1: Defining UCT

Thank you for requesting more information on this important component of our paper. We agree that further explanation of the UCT formulation in Line 176 would improve clarity. UCT (Upper Confidence bounds applied to Trees), originally introduced by Kocsis and Szepesvári (2006), is used as a principled strategy for node selection in Monte Carlo Tree Search. In our implementation, each node $v$ maintains a visit count $N_v$ and an accumulated reward $R_{\mathrm{total}}(v)$, initialized to zero when the node is first created during the Expansion step. The UCT score used in the Selection step is given by

$

UCT(v) = \frac{R_{\mathrm{total}}(v)}{N_v} + C \sqrt{\frac{\ln N_p}{N_v}}

$

where $N_p$ is the visit count of the parent node and $C$ is a tunable exploration coefficient. This formulation balances exploitation (favoring actions with high empirical reward via $\tfrac{R_{\mathrm{total}}(v)}{N_v}$) and exploration (favoring less‑visited actions via $\sqrt{\tfrac{\ln N_p}{N_v}}$). The logarithmic scaling ensures diminishing exploration pressure as nodes become well‑sampled, while still encouraging discovery in underexplored parts of the tree. This approach builds directly on foundational work in bandit‑based planning [1,2]. We will cite these foundational papers and revise Section 3.4 to make the motivation, variable definitions, and intuition behind UCT more explicit.

Proposed Addition to Section 3.4 point 1. Selection

""" This selection strategy follows the UCT algorithm [1], which applies UCB1 (Upper Confidence Bounds) from multi-armed bandit theory to guide Monte Carlo Tree Search. We initialize $R_{\mathrm{total}}(v)=0$ and $N_v=0$ when a node is first added to the tree.​​​ The first term estimates the average reward of the node, promoting high-reward questions (maximizing information gain). The second term $$ C \sqrt{\frac{\ln N_p}{N_v}} $$, derived from UCB1, encourages exploration of under-visited nodes and grows as $N_v$ remains small relative to the parent’s visit count $N_p$. This balance enables UCT to trade off exploitation and exploration effectively. UCT has been proven consistent, with the probability of selecting suboptimal actions at the root decreasing polynomially with more samples [1]. Empirically, it has shown strong performance in large-scale domains such as Go and planning under uncertainty [2].

"""

References:

• [1] Kocsis, L. and Szepesvári, C. “Bandit based Monte-Carlo Planning.” ECML 2006.
• [2] Chaslot, G., Bakkes, S., Szita, I., and Spronck, P. “Monte-Carlo Tree Search: A New Framework for Game AI.” Proceedings of the AAAI Conference on Artificial Intelligence and Interactive Digital Entertainment (AIIDE), vol. 4, pp. 216–217, 2008.

---

Q2: Uncertainty reduction

Thank you for raising this important clarification question. Uncertainty evaluation is a fundamental part of our approach and directly addresses the core challenge of information-seeking under incomplete information.

Uncertainty stems from the fact that users often begin with underspecified problem descriptions, missing out various relevant details. This creates an uncertain decision space where the system must systematically gather missing information through targeted questioning to converge on the correct solution by narrowing the possibility space $\Omega$.

• Each time the system generates a candidate question, it is rewarded based on how well it splits the current possibility space into two balanced subsets. This is measured using the entropy $IG_v$ $= -p_A \log p_A - p_N \log p_N$, where $p_A$ and $p_N$ are the proportions of possible outcomes consistent with an affirmative or negative response. According to information theory, entropy is maximized when $p_A = p_N = 0.5$.

• The logic and evaluation closely follow prior work (Hu et al., 2025), where entropy and information gain drive the reward signals that are used for selecting questions. By always aiming for questions that most sharply reduce uncertainty, our method is designed to be principled and efficient in uncertainty resolution.

• The closer a question gets to dividing the remaining options in half, the more it reduces entropy, which is inherently incentivized in our reward calculation for question selection. In Response to Q1 by Reviewer 54vp, we highlight that this ratio of splitting the possibility space is empirically found to be near 52.82% on average across samples. This supports the uncertainty quantification and the premise that LLMs can reliably generate questions that closely approximate the optimal information gain.

Thank you for thoroughly reviewing our paper and sharing your feedback.

We appreciate your positive recognition of our method’s core strengths, including the practical efficiency of feedback-aware MCTS along with the system’s memory and reuse abilities. Below, we address the two main weaknesses you identified.

---

W1. Lack of Penalty Signals

Our feedback mechanism is designed so that positive rewards are propagated only along successful questioning trajectories, that is, when the system solves a case, those specific question nodes receive a bonus, making them more likely to be chosen for similar future problems. As a result, effective questions in the beginning are reinforced and preferred in analogous contexts, improving efficiency and accuracy over time.

However, in this work, we do not actively penalize failed questioning paths when a trajectory does not succeed. They are neither incentivized with a bonus reward nor punished, effectively remaining neutral in the search process. We would like to highlight that the absence of feedback does not lead to repeating bad ideas. Importantly, because MCTS uses visit counts and average reward estimates to guide selection, low-reward nodes naturally become less likely to be explored again. In practice, when our system encounters a new case that does not correspond to any known cluster or lacks prior successful trajectories, it defaults to standard MCTS behavior without using bonus term. This means such cases proceed without any inherited bias from past failures, eliminating the risk of getting stuck on repeatedly unsuccessful paths.

Assigning a simple penalty signal presents certain challenges.

• Penalizing every failure could introduce noise in the learning process, as a question that is unhelpful for one specific scenario might be the right choice for another that is only subtly different.
• Propagating negative feedback uniformly could suppress useful exploration or inadvertently destabilize long-term learning.

Crafting an effective penalty mechanism in this setting is non-trivial and requires careful consideration to avoid discouraging paths that may actually be valuable in neighboring problems. Approaches such as confidence-weighted penalties, where only repeatedly unsuccessful and low-confidence trajectories are penalized, represent promising directions.

We have discussed this challenge in our Limitations section (line 353) and will expand our rationale and potential solutions in the revised version. We sincerely acknowledge that this represents a valuable avenue for future work.

---

W2. Generalization in Open-Ended Domains

The primary reason for the overall drop in hit rate in open-set domains is the combinatorial complexity and ambiguity of unconstrained target spaces, which make the task inherently more difficult, affecting not just our approach but all other planning methods [Hu et al., 2025]. In such settings, the system must not only generate effective questions but also dynamically construct the set of possible outcomes on the fly. This compounds the uncertainty and leads to lower success rates because we expect the response to exactly match the ground truth.

As shown in Table 5 of the Appendix, all tested baselines exhibit similar or lower performance in the open-set conditions, with MISQ-HF still outperforming them. This suggests that the performance drop is not specific to our method, but reflects the underlying challenge of generalization in these domains.

We agree that improving robustness in open-ended settings is a valuable direction. Extensions of this work could integrate retrieval-augmented methods or external knowledge bases to help dynamically construct and refine candidate outcome sets.

We appreciate your feedback and will clarify that while our method is primarily designed for and excels in closed set scenarios, addressing open-ended answer spaces remains an important open challenge we hope to advance in future work.

Dear reviewer,

Thank you for bringing this ICML 2025 paper to our attention. Since the dataset was made publicly available only in July (long after the NeurIPS submission deadline in May), we couldn’t include it in our original experiments. We acknowledge that it offers a useful testbed for open-ended reasoning.

We would also like to note that our paper reports results across 5 datasets spanning 3 domains (medical diagnosis, trouble shooting, and 20-questions). Table 1 of this benchmark paper cites all of these domains (we used the same datasets as Hu et al. 2024) under the Active Reasoning paradigm. This overlap reinforces that our experiments reflect the relevance and coverage with respect to recent work on information seeking.

We value your suggestion and have started running experiments on the new benchmark. Given the careful setup required and the limited time left in the discussion phase, we plan to include the results in the final version and will share preliminary findings here as soon as they are ready.

We are grateful for your support and positive recognition of our work.

Sincerely,

The Authors

Dear Reviewer m8dK,

As conveyed in our previous comment, we report the results below. We began with the Situation Puzzles (SP) data from the benchmark you suggested, as it is the closest match to our problem setup. SP directly tests open-ended reasoning for information seeking:- given the surface description of a mystery puzzle, the player/detective LLM must ask yes/no questions to collect the clues in order to come up with the target explanation that determines what happened, why, and how. This is a challenging setting because -

• Unlike domains with well-defined options in the space of possible outcomes (e.g., a fixed set of disease names or technical issues) that can be mapped to various problems, SP requires asking logical case-specific questions to identify free-text explanations that differ for every mystery puzzle, making the target space highly variable and open-ended.
• This makes it an inherently open-set scenario, where we initialize a set of possible outcomes $\Omega$ that can serve as an explanation and narrow down that space to arrive at the real answer.

We compared DP, UoT (Hu et al., 2024), and MISQ (ours) under maximum allowed turns ($T$) limited to 10 and 25 turns, reporting Success Rate (SR), Mean Conversation Length in Successful Cases (MSC), and Question Generation Calls (QGC) on the test set of SP containing 100 mystery puzzles. Definitions of these metrics are elaborated in our paper.

To handle the open-set nature of SP, we initialize the candidate set $\Omega$ with 5 possibilities from the puzzle description and two initial question-answer turns. During the information-seeking phase, $\Omega$ is updated dynamically whenever two candidates remain, refining the set based on new information to guide the questioning strategy. In the targeting phase, deductions are drawn from the conversation history to generate the final explanation, which the judge LLM verifies against the ground truth. Our adaptive open-set approach helps MISQ efficiently explore diverse free-text explanations and improve question utility and overall success rate.

The experimental configuration is the same as described in Section 4.4 of our paper. We used Llama 3.3 70B Instruct for the following results.

In this setting, we used MISQ rather than MISQ-HF, as direct reuse of prior search paths is less meaningful when each case demands a unique questioning path shaped by different possible explanations. MISQ still offers clear advantages over UoT, as its feedback-driven scoring within a single episode (UCT considers the number of visits across K simulations) prioritizes high-utility questions, reducing redundancy and improving SR and MSC while significantly lowering QGC.

We are happy to add the relevance of this benchmark in our paper and include the results in appendix. We appreciate your suggestion, as this evaluation further validates that our method is effective and makes an important contribution to the field. We trust it provides additional evidence to support your favorable evaluation of our work.

Sincerely,

The Authors