Feedback-Aware Monte Carlo Tree Search for Efficient Information Seeking in Goal-Oriented Conversations

Chirag Shah; Harshita Chopra

REVIEW 1 major objections 6 minor 25 references

Feedback-Aware Monte Carlo Tree Search for Efficient Information Seeking in Goal-Oriented Conversations

T0 review · 1 major / 6 minor · reviewed 2026-08-10 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read A feedback-aware Monte Carlo tree search planner raises goal-oriented conversation success by about 12% while cutting planning LLM calls roughly tenfold.

desk verdict Plausible incremental extension of UoT with a useful feedback mechanism, but the headline 12%/10x numbers rest on an online/transductive evaluation and are not established without order-randomized repetitions or a held-out split. read the letter →

arxiv 2501.15056 v2 pith:4AL3XLGG submitted 2025-01-25 cs.AI cs.CLcs.HCcs.LG

Harshita Chopra , Chirag Shah This is my paper · ORCID

classification cs.AIcs.CLcs.HCcs.LG

keywords MonteCarloTreeSearchhierarchicalfeedbackinformationseekinggoal-orientedconversationLLMplanningmedicaldiagnosistroubleshootingquestionselection

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

Goal-oriented conversation systems must ask questions that shrink a large space of possible answers as quickly as possible. This paper tries to establish that a Monte Carlo tree search planner with a cluster-based memory of past successful questions can do that more accurately and far more cheaply than existing tree-planning methods. On medical-diagnosis and technical-troubleshooting benchmarks, it reports an average 12% success-rate improvement and about ten times fewer LLM calls for planning per conversation, plus an additional 8% success gain when the start is a constrained set of possibilities. If the result holds, inference-time question planning becomes practical for cost- and latency-sensitive assistants.

What carries the argument

The central object is a cached decision tree of questions in which each node stores a subset of possible outcomes, and an LLM generates candidate questions that split that subset into balanced affirmative and negative parts. A normalized entropy reward scores information gain, and Monte Carlo tree search selects questions with the UCT formula, where a cluster-specific bonus $B_k(v)$ is added to bias selection toward historically successful paths. After a successful conversation the bonus is updated along the path from the target back to the root with exponential decay $\gamma^{d_v}$, while depth-limited rollouts and a single shared tree keep the number of LLM calls low.

What would settle it

Run MISQ-HF and UoT against human answerers on DX, MedDG, and FloDial, or replay a fixed corpus of human responses; if the ~12% success advantage and ~10x question-generation-call reduction do not appear, the central efficiency claim is refuted.

Watch

Extended reading notes

Core claim

The central claim is that the UCT score in Monte Carlo tree search can be augmented with a cluster-specific bonus learned from past successful conversations, and that this hierarchical feedback is what improves success while cutting planning cost. Each new problem is embedded and assigned to a semantic cluster; after a successful interaction, a bonus reward is propagated up the shared decision tree along the successful question path, decaying exponentially with depth, so later similar problems prefer questions that worked before. On medical diagnosis and technical troubleshooting, the paper reports an average 12% success-rate improvement over the UoT baseline and roughly a tenfold reduction in question-generation LLM calls, with a further 8% gain when the root starts from a constrained possibility set.

Load-bearing premise

The load-bearing premise is that the LLM pretending to be the patient or customer answers the way real people would, so every reported success rate and feedback signal rests on that simulation.

Editorial extensions

If this is right

Conversational agents can plan with tree search at a fraction of the inference cost, since selective expansion and cached trees cut question-generation LLM calls by about an order of magnitude.
A deployed system improves with use, because successful conversations reinforce cluster-specific bonuses in a shared decision tree that later similar problems reuse.
Giving the planner a constrained set of likely outcomes derived from the user's description adds about 8% success on average, so cheap upfront classification pays for itself.
The efficiency gains hold in open-set settings where the outcome space is unknown and must be generated on the fly.
The planner scales to large possibility spaces: on the 20 Questions datasets, selective MCTS without feedback already beats exhaustive tree planning on both success rate and question-generation cost.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

An extension the paper does not test is learning the decay factor and bonus scale per cluster rather than fixing them, which would directly test whether early broad questions should always dominate.
Because the bonus is a scalar added to UCT, the same feedback layer could be paired with any base question generator, so the gains may transfer to LLMs beyond the three evaluated.
The simulated answerer is the main unstated risk: a human replay of the same tasks would show whether real user wording and non-cooperative answers preserve the reported gains.
The cluster assignments used for feedback could also serve as cache keys for retrieving successful question trajectories, connecting this planner to retrieval-augmented question generation.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

The stress-test concern is real. Algorithm 1 processes samples sequentially, caching one decision tree and updating cluster bonuses B_k(v) after every success. So the questions asked for sample i depend on all earlier samples, and the reported SR and QGC are averages over this online process. No order shuffling, learning curves, or train/test split are reported. That makes the headline 12% SR gain and ~10x QGC reduction order-dependent, and the comparison with UoT is ambiguous if UoT did not run under the same shared-tree/online-update protocol. Part of the claimed gain could be amortization and test-time learning rather than the algorithm itself. The authors do not flag this in the limitations section, which only mentions not learning from failures.

What is genuinely new: the cluster-specific bonus term in the UCT formula and the selective expansion that cuts LLM calls. The ablation (MISQ vs MISQ-HF) shows the feedback term changes results, which is a real point in favor. The paper is clearly written, the method is described well enough to reimplement, and the QGC metric is useful. Those are solid contributions.

Other soft spots are more minor but real: single runs without error bars or significance tests; many hand-set hyperparameters (C, lambda, gamma, tau, beta, ds, m, delta) with no sensitivity analysis; and the code link in the footnote is not actually present in the text. The simulated user (Llama 3.3 70B) is acknowledged and is a known limitation, but it does mean success rates may not transfer to humans.

The central idea is plausible and the ablation helps, but the evaluation as reported does not establish the headline claims. This is fixable in revision: run multiple random orders, report learning curves, and compare against UoT under the same online protocol or with a proper split for tree and bonus learning.

The paper is for researchers working on inference-time planning in goal-oriented dialogue, especially those building on UoT. It deserves a serious referee because the method is nontrivial and the ablation is informative, but it needs substantial rework on evaluation before I would trust the numbers.

Referee Report

1 major / 6 minor

Summary. The paper introduces MISQ-HF, a Monte Carlo Tree Search framework for information-seeking question selection in goal-oriented dialogues. Questions are generated by an LLM, selected via a UCT-style search with an information-gain reward, and augmented by a cluster-based bonus that propagates feedback from successful conversations through a shared, cached decision tree. The system is evaluated across medical diagnosis (DX, MedDG), troubleshooting (FloDial), and 20 Questions datasets, using two baselines (Direct Prompting and Uncertainty of Thoughts) and three LLM questioners, with a simulated LLM user. The headline claims are an average 12% improvement in success rate and about 10x reduction in planning LLM calls over UoT, plus an additional 8% success-rate gain when initializing from a constrained possibility set.

Significance. If the reported gains are robust, the paper makes a useful practical contribution: a cheaper, more adaptive questioning policy for goal-oriented conversational agents, backed by an ablation (MISQ vs MISQ-HF) that supports the value of the feedback term. The authors also provide code, concrete prompts, and a tractable algorithmic description, which are strengths. However, the central quantitative claims rest on an evaluation protocol whose validity is not yet established: the method is evaluated in an online, transductive mode where the tree and cluster bonuses are updated on the test set itself, with no held-out evaluation or order-controlled repetitions. The simulated user further limits the strength of any deployment-oriented claim. The paper's contribution is therefore currently more suggestive than conclusive.

major comments (1)

[§5, Table 3] Table 3 is used to argue that MISQ (without feedback) outperforms UoT on the 20 Questions domain, but the table does not include MISQ-HF. The paper explains that MISQ-HF cannot be evaluated without a problem description, yet the conclusion in §5 states that “MISQ-HF outperforms other baselines” on datasets including this one. This is an overreach: the feedback component is not tested in the 20 Questions setting, and the claim should be restricted to MISQ. Please also note that the 12%/10x headline from the abstract is computed only on MD and TS domains, which are the domains where the feedback mechanism is active.

minor comments (6)

[Algorithm 1, line 8] The condition `|Ωv ≥ 2|` should read `|Ωv| ≥ 2`; as written it is not well-formed.
[§3.4, Eq. (8)] The backpropagation equation is written inline after the algorithm description, but it is not numbered like the other equations; consider numbering it for consistency.
[§4.2] The omission of Chain-of-Thought and Tree-of-Thought baselines is justified by reference to [7], but the paper should state explicitly that UoT is the state-of-the-art comparator in this setting, since the abstract claims comparison to “state of the art.”
[§5, “In the medical domain ...”] The sentence “the cost of misdiagnosis far outweighs the overhead of a few additional clarifying questions” is a value judgment that is not supported by the data; rephrase as a discussion of the SR-MSC trade-off.
[Appendix A, Table 4] The table caption says “Improvement in Success Rate ... when using MISQ-HF-c,” but the body text in §5 states “Figure 2 shows the notable gain in SR using GPT-4o,” and Figure 2 is not shown inline. Please make the figure/table cross-referencing consistent.
[Appendix G] The QGC efficiency analysis assumes a full (2m)^ds expansion for UoT, but it does not account for UoT’s own pruning or caching behavior; the 10x claim should be qualified accordingly.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the reported gains are empirical comparisons against an external simulated answerer, and no equation reduces the central claims to fitted constants or self-citations.

full rationale

The paper's derivation chain is not circular. The information-gain reward RIG(v) in Eq. 2-3 is computed from the LLM's own proposed partition of the possibility set, but this is a search heuristic, not the success criterion; success is determined by whether the Questioner identifies the hidden target held by a separately simulated Answerer (§4.4). The cluster bonus Bk(v) in Eq. 12 is updated after conversations that already succeeded, and Eq. 13 feeds it back into UCT selection; this is an online-learning mechanism and does not make the reported success rate equal to an input by construction. The 10x QGC reduction and 12% SR gain are empirical measurements, not derived identities, and no equation in the paper forces them. The shared-tree and cluster-bonus updates during evaluation (Algorithm 1, §3.2, §3.5.2) raise a legitimate evaluation-protocol concern about transductive or order-dependent reporting, but that is a validity limitation, not a definitional circularity. The paper contains no load-bearing self-citations; the UoT baseline [7] is prior independent work by other authors and is used for comparison and prompting conventions, not to prove the central claim. The limitations paragraph acknowledges missing failure-case learning, redundant-question penalties, and confidence metrics, but these are scope limitations rather than circular steps.

Assumptions & free parameters 10 free parameters · 6 assumptions · 0 invented entities

The central claim relies on a cluster-based feedback loop whose transfer assumption is not independently validated, plus a set of hand-set hyperparameters. No new physical or ontological entities are introduced.

free parameters (10)

MCTS exploration constant C = 0.2
Hand-set in §4.4; controls exploration-exploitation trade-off, no sensitivity analysis.
MCTS iterations K = 10
Hand-set; higher K would improve search at greater cost.
Simulation depth ds = 3
Hand-set rollout depth in §4.4.
Candidate questions per node m = 3
LLM asked to generate 3 questions per node; affects tree width.
Reward scaling parameter lambda = 0.4
Normalizes information gain in Eq. 3; no sensitivity analysis.
Question-asking turn ratio delta = 0.6
Sets fraction of turns used for information seeking vs targeting.
Bonus decay factor gamma = 0.9
Controls exponential decay of cluster bonus with depth in Eq. 12.
Cluster similarity threshold tau = 0.9
Cosine threshold for assigning a case to a cluster in Eq. 10.
Bonus scaling factor beta = 0.2
Scales cluster bonus update in Eq. 12.
Open-set initial possibility size = 5
In Open Set experiments, the LLM is asked to propose 5 initial possibilities; affects all open-set results.

assumptions (6)

domain assumption The LLM can generate valid binary questions and partition Omega_v into yes/no subsets that respect the true target.
Central to tree construction and reward computation; appears in §3.2 and the prompts in Appendix C-D. If partitions are wrong, information gain rewards are miscalibrated.
domain assumption The probabilities pA_v and pN_v, computed from LLM partition sizes, are good proxies for the probability that a real user answers yes or no.
Used in Eq. 2-3; if real user answer rates differ, the question rewards do not reflect information gain.
domain assumption The answerer LLM simulates human behavior faithfully.
All empirical results use Llama 3.3 70B as the user; stated in §4.4.
domain assumption Cosine similarity of problem descriptions at threshold tau identifies cases that share successful questioning strategies.
The whole feedback mechanism in §3.5 assumes transfer within clusters; no validation of this transfer is provided.
domain assumption A single cached tree, shared across all samples in a dataset, remains a valid search structure as new problems stream in.
The method reuses the decision tree across data points (§3.2); questions generated for early samples may bias later planning.
standard math UoT reward propagation equations (Eq. 4-6) are accepted from [7].
The paper reuses UoT's recursive reward definitions as a starting point; no new derivation is attempted.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Feedback-Aware Monte Carlo Tree Search for Efficient Information Seeking in Goal-Oriented Conversations." pith.science (2026). https://pith.science/paper/4AL3XLGG

@misc{pith2026250115056,
  author       = {Pith},
  title        = {Pith review of: Feedback-Aware Monte Carlo Tree Search for Efficient Information Seeking in Goal-Oriented Conversations},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/4AL3XLGG}},
  note         = {Machine review of arXiv:2501.15056}
}

read the original abstract

Effective decision-making and problem-solving in conversational systems require the ability to identify and acquire missing information through targeted questioning. A key challenge lies in efficiently narrowing down a large space of possible outcomes by posing questions that minimize uncertainty. To address this, we introduce a novel framework that leverages Large Language Models (LLMs) to generate information-seeking questions, with Monte Carlo Tree Search (MCTS) to strategically select questions that maximize information gain, as a part of inference-time planning. Our primary contribution includes a hierarchical feedback mechanism that exploits past interaction patterns to guide future strategy. Specifically, each new problem is mapped to a cluster based on semantic similarity, and our UCT (Upper Confidence bound for Trees) formulation employs a cluster-specific bonus reward to prioritize successful question trajectories that have proven effective for similar problems in the past. Extensive empirical evaluation across medical diagnosis and technical troubleshooting domains shows that our method achieves an average of 12% improvement in success rates and about 10x reduction in the number of LLM calls made for planning per conversation, compared to the state of the art. An additional 8% gain in success rate is observed on average when we start with a constrained set of possibilities. Our results underscore the efficacy of feedback-aware MCTS in enhancing information-seeking in goal-oriented dialogues.

Figures

Figures reproduced from arXiv: 2501.15056 by the authors.

**Figure 2.** Gain in SR when using MISQHF initialized with a constrained set. Across both datasets in the MD domain, MISQ-HF achieves superior performance, with an overall average reduction of 6.77 times in Mean QGC over UoT. Notably, using Llama 3.3 70B, MISQ-HF required only 0.04 Mean QGC compared to UoT’s 0.77 on DX, showing a 19.25 times reduction. Similarly, a reduction of 8.97 times is observed on MedDG. This efficiency s… view at source ↗

**Figure 3.** Improvement in Success Rate on MD and TS Domain in a Closed Set scenario, when [PITH_FULL_IMAGE:figures/full_fig_p012_3.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

25 extracted references · 12 canonical work pages

[1]

A. B. API. Build Generative AI Applications with Foundation Models - Amazon Bedrock - AWS, 2023. URL https://aws.amazon.com/bedrock/

work page 2023
[2]

Chaslot, S

G. Chaslot, S. Bakkes, I. Szita, and P. Spronck. Monte-carlo tree search: A new framework for game ai. In Proceedings of the AAAI Conference on Artificial Intelligence and Interactive Digital Entertainment, volume 4, pages 216–217, 2008

work page 2008
[3]

T. Dam, G. Chalvatzaki, J. Peters, and J. Pajarinen. Monte-Carlo Robot Path Planning. IEEE Robotics and Automation Letters , 7(4):11213–11220, Oct. 2022. ISSN 2377-3766. doi: 10.1109/LRA.2022.3199674. URL https://ieeexplore.ieee.org/document/9860036

work page arXiv 2022
[4]

Grattafiori, A

A. Grattafiori, A. Dubey, A. Jauhri, A. Pandey, A. Kadian, A. Al-Dahle, A. Letman, A. Mathur, A. Schelten, A. Vaughan, A. Goyal, A. Hartshorn, A. Yang, A. Sravankumar, A. Korenev, A. Hinsvark, A. Rao, A. Zhang, A. Rodriguez, and A. Gregerson et al. The Llama 3 Herd of Models, Nov. 2024. URL http://arxiv.org/abs/2407.21783. arXiv:2407.21783

arXiv 2024
[5]

A. Guez, T. Weber, I. Antonoglou, K. Simonyan, O. Vinyals, D. Wierstra, R. Munos, and D. Silver. Learning to search with MCTSnets. In Proceedings of the 35th International Conference on Machine Learning , pages 1822–1831. PMLR, July 2018. URL https:// proceedings.mlr.press/v80/guez18a.html

work page 2018
[6]

M. N. Hebart, A. H. Dickter, A. Kidder, W. Y . Kwok, A. Corriveau, C. V . Wicklin, and C. I. Baker. THINGS: A database of 1,854 object concepts and more than 26,000 natu- ralistic object images. PLOS ONE, 14(10):e0223792, Oct. 2019. ISSN 1932-6203. doi: 10.1371/journal.pone.0223792. URL https://journals.plos.org/plosone/article? id=10.1371/journal.pone.0223792

doi:10.1371/journal.pone.0223792 2019
[7]

Z. Hu, C. Liu, X. Feng, Y . Zhao, S.-K. Ng, A. T. Luu, J. He, P. W. Koh, and B. Hooi. Uncertainty of Thoughts: Uncertainty-Aware Planning Enhances Information Seeking in LLMs. In The Thirty-eighth Annual Conference on Neural Information Processing Systems, Nov. 2024. URL https://openreview.net/forum?id=CVpuVe1N22&noteId=aTI8PGpO47

work page 2024
[8]

Hui and K

W. Hui and K. Tu. RoT: Enhancing Large Language Models with Reflection on Search Trees, July 2024. URL http://arxiv.org/abs/2404.05449. arXiv:2404.05449 [cs]

arXiv 2024

Show all 25 references

[9]

A. Q. Jiang, A. Sablayrolles, A. Roux, A. Mensch, B. Savary, C. Bamford, D. S. Chaplot, D. d. l. Casas, E. B. Hanna, F. Bressand, G. Lengyel, G. Bour, G. Lample, L. R. Lavaud, L. Saulnier, M.-A. Lachaux, P. Stock, S. Subramanian, S. Yang, S. Antoniak, T. L. Scao, T. Gervet, T....

2024 arXiv
[10]

S. S. Li, V . Balachandran, S. Feng, J. S. Ilgen, E. Pierson, P. W. Koh, and Y . Tsvetkov. MediQ: Question-Asking LLMs and a Benchmark for Reliable Interactive Clinical Reasoning. In The Thirty-eighth Annual Conference on Neural Information Processing Systems, Nov. 2024

2024
[11]

Achiam, S

OpenAI, J. Achiam, S. Adler, S. Agarwal, L. Ahmad, I. Akkaya, F. L. Aleman, D. Almeida, J. Altenschmidt, S. Altman, S. Anadkat, R. Avila, I. Babuschkin, S. Balaji, V . Balcom, P. Bal- tescu, H. Bao, M. Bavarian, J. Belgum, and I. Bello et al. GPT-4 Technical Report, Mar. 2024....

2024 arXiv
[12]

Putta, E

P. Putta, E. Mills, N. Garg, S. Motwani, C. Finn, D. Garg, and R. Rafailov. Agent Q: Advanced Reasoning and Learning for Autonomous AI Agents, Aug. 2024. URL http://arxiv.org/ abs/2408.07199. arXiv:2408.07199

2024 arXiv
[13]

Raghu, S

D. Raghu, S. Agarwal, S. Joshi, and Mausam. End-to-End Learning of Flowchart Grounded Task- Oriented Dialogs. In M.-F. Moens, X. Huang, L. Specia, and S. W.-t. Yih, editors, Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 4348– 436...

2021 doi
[14]

V . Sanh, L. Debut, J. Chaumond, and T. Wolf. Distilbert, a distilled version of bert: smaller, faster, cheaper and lighter. ArXiv, abs/1910.01108, 2019

1910 arXiv
[15]

Silver, A

D. Silver, A. Huang, C. J. Maddison, A. Guez, L. Sifre, G. van den Driessche, J. Schrit- twieser, I. Antonoglou, V . Panneershelvam, M. Lanctot, S. Dieleman, D. Grewe, J. Nham, N. Kalchbrenner, I. Sutskever, T. Lillicrap, M. Leach, K. Kavukcuoglu, T. Graepel, and D. Hassabis. ...

2016 doi
[16]

Silver, T

D. Silver, T. Hubert, J. Schrittwieser, I. Antonoglou, M. Lai, A. Guez, M. Lanctot, L. Sifre, D. Kumaran, T. Graepel, T. Lillicrap, K. Simonyan, and D. Hassabis. A general reinforcement learning algorithm that masters chess, shogi, and Go through self-play. Science, 362(6419):...

2018 doi
[17]

G. Wang, X. Liu, Z. Ying, G. Yang, Z. Chen, Z. Liu, M. Zhang, H. Yan, Y . Lu, Y . Gao, K. Xue, X. Li, and Y . Chen. Optimized glycemic control of type 2 diabetes with reinforcement learning: a proof-of-concept trial. Nature Medicine, 29(10):2633–2642, Oct. 2023. ISSN 1546- 170...

2023 doi
[18]

J. Wei, X. Wang, D. Schuurmans, M. Bosma, B. Ichter, F. Xia, E. H. Chi, Q. V . Le, and D. Zhou. Chain-of-thought prompting elicits reasoning in large language models. In Proceedings of the 36th International Conference on Neural Information Processing Systems , NIPS ’22, pages...

2022
[19]

L. Xu, Q. Zhou, K. Gong, X. Liang, J. Tang, and L. Lin. End-to-End Knowledge-Routed Relational Dialogue System for Automatic Diagnosis. Proceedings of the AAAI Conference on Artificial Intelligence, 33(01):7346–7353, July 2019. ISSN 2374-3468. doi: 10.1609/aaai. v33i01.3301734...

2019 doi
[20]

S. Yao, D. Yu, J. Zhao, I. Shafran, T. Griffiths, Y . Cao, and K. Narasimhan. Tree of thoughts: Deliberate problem solving with large language models. In A. Oh, T. Naumann, A. Globerson, K. Saenko, M. Hardt, and S. Levine, editors, Ad- vances in Neural Information Processing S...
[21]

X. Yu, M. Chen, and Z. Yu. Prompt-Based Monte-Carlo Tree Search for Goal-oriented Dialogue Policy Planning. In H. Bouamor, J. Pino, and K. Bali, editors, Proceedings of the 2023 Con- ference on Empirical Methods in Natural Language Processing, pages 7101–7125, Singapore, Dec. ...

2023 doi
[22]

Z. Zhao, W. S. Lee, and D. Hsu. Large language models as commonsense knowledge for large-scale task planning. In Proceedings of the 37th International Conference on Neural Information Processing Systems, NIPS ’23, pages 31967–31987, Red Hook, NY , USA, May
[23]

X is possibly one of the following: Ω

Y . Zhu, L. Pang, Y . Lan, H. Shen, and X. Cheng. Adaptive Information Seeking for Open- Domain Question Answering. In M.-F. Moens, X. Huang, L. Specia, and S. W.-t. Yih, editors, Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 361...

2021 doi
[2023]

URL https://proceedings.neurips.cc/paper_files/paper/2023/file/ 271db9922b8d1f4dd7aaef84ed5ac703-Paper-Conference.pdf

2023
[2024]

Curran Associates Inc

Pith tools

Reviewed August 10, 2026 · model on record in the stance chip above.

[1] [1]

A. B. API. Build Generative AI Applications with Foundation Models - Amazon Bedrock - AWS, 2023. URL https://aws.amazon.com/bedrock/

work page 2023

[2] [2]

Chaslot, S

G. Chaslot, S. Bakkes, I. Szita, and P. Spronck. Monte-carlo tree search: A new framework for game ai. In Proceedings of the AAAI Conference on Artificial Intelligence and Interactive Digital Entertainment, volume 4, pages 216–217, 2008

work page 2008

[3] [3]

T. Dam, G. Chalvatzaki, J. Peters, and J. Pajarinen. Monte-Carlo Robot Path Planning. IEEE Robotics and Automation Letters , 7(4):11213–11220, Oct. 2022. ISSN 2377-3766. doi: 10.1109/LRA.2022.3199674. URL https://ieeexplore.ieee.org/document/9860036

work page arXiv 2022

[4] [4]

Grattafiori, A

A. Grattafiori, A. Dubey, A. Jauhri, A. Pandey, A. Kadian, A. Al-Dahle, A. Letman, A. Mathur, A. Schelten, A. Vaughan, A. Goyal, A. Hartshorn, A. Yang, A. Sravankumar, A. Korenev, A. Hinsvark, A. Rao, A. Zhang, A. Rodriguez, and A. Gregerson et al. The Llama 3 Herd of Models, Nov. 2024. URL http://arxiv.org/abs/2407.21783. arXiv:2407.21783

arXiv 2024

[5] [5]

A. Guez, T. Weber, I. Antonoglou, K. Simonyan, O. Vinyals, D. Wierstra, R. Munos, and D. Silver. Learning to search with MCTSnets. In Proceedings of the 35th International Conference on Machine Learning , pages 1822–1831. PMLR, July 2018. URL https:// proceedings.mlr.press/v80/guez18a.html

work page 2018

[6] [6]

M. N. Hebart, A. H. Dickter, A. Kidder, W. Y . Kwok, A. Corriveau, C. V . Wicklin, and C. I. Baker. THINGS: A database of 1,854 object concepts and more than 26,000 natu- ralistic object images. PLOS ONE, 14(10):e0223792, Oct. 2019. ISSN 1932-6203. doi: 10.1371/journal.pone.0223792. URL https://journals.plos.org/plosone/article? id=10.1371/journal.pone.0223792

doi:10.1371/journal.pone.0223792 2019

[7] [7]

Z. Hu, C. Liu, X. Feng, Y . Zhao, S.-K. Ng, A. T. Luu, J. He, P. W. Koh, and B. Hooi. Uncertainty of Thoughts: Uncertainty-Aware Planning Enhances Information Seeking in LLMs. In The Thirty-eighth Annual Conference on Neural Information Processing Systems, Nov. 2024. URL https://openreview.net/forum?id=CVpuVe1N22&noteId=aTI8PGpO47

work page 2024

[8] [8]

Hui and K

W. Hui and K. Tu. RoT: Enhancing Large Language Models with Reflection on Search Trees, July 2024. URL http://arxiv.org/abs/2404.05449. arXiv:2404.05449 [cs]

arXiv 2024