pith. sign in

arxiv: 2605.22195 · v1 · pith:FGPIMARWnew · submitted 2026-05-21 · 💻 cs.LG

Reinforced Graph of Thoughts: RL-Driven Adaptive Prompting for LLMs

Manuel Noah Riesen , Peter Alfred von Niederh\"ausern This is my paper

Pith reviewed 2026-05-22 07:25 UTC · model grok-4.3

classification 💻 cs.LG
keywords graph of thoughtsreinforcement learningadaptive promptinglarge language modelsprompt engineeringautomated reasoningRL for LLMs
0
0 comments X

The pith

Reinforcement learning can automatically generate task-specific graphs of operations for large language model prompting.

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

The paper proposes Reinforced Graph of Thoughts, which applies reinforcement learning to build graphs of operations adaptively from a fixed human-defined set instead of requiring manual construction for each problem. This targets the limitation that static graphs lack flexibility and demand expert insight into solution structures. A sympathetic reader would care because the method promises to make structured prompting more practical for problems whose complexity varies, reducing the need for repeated human design of the graph layout.

Core claim

Reinforced Graph of Thoughts (RGoT) is an automated approach to the Graph of Thoughts prompting paradigm that leverages reinforcement learning to adaptively generate a graph of operations from a human-defined set. Results indicate that, under certain constraints, it is possible to construct graphs of operations adaptively to the task's complexity in an automated way.

What carries the argument

Reinforced Graph of Thoughts (RGoT), an RL agent that selects and connects operations from the human-defined set to form a graph whose structure varies with the input task.

If this is right

  • Graphs of operations no longer require manual redesign for each new problem instance.
  • The generated graph structure can change automatically according to detected task complexity.
  • Elaborate LLM problem solving becomes feasible without in-depth knowledge of solution paths for every task.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same RL selection process could be tested on other structured prompting formats such as trees or chains to check transferability.
  • Experiments that systematically vary the size and coverage of the operation set would clarify the boundary between workable and intractable constraint regimes.

Load-bearing premise

The human-defined set of operations must contain good solutions for the target tasks yet remain small enough for reinforcement learning to explore the space of possible graphs effectively.

What would settle it

On a high-complexity task solvable by a manually designed graph, measure whether the RL agent fails to discover any viable graph or produces solutions whose accuracy falls below that of the static baseline.

Figures

Figures reproduced from arXiv: 2605.22195 by Manuel Noah Riesen, Peter Alfred von Niederh\"ausern.

Figure 1
Figure 1. Figure 1: Prompting Paradigms with increasing complexity (from left to right). [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Example GoO, GoT, and their relationship for the task [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Operation success probability Pˆ(c) as a function of complexity c for the operations of the tasks sum list and count keywords, evaluated with gpt-3.5-turbo-0125. 4.2 Reinforced Graph of Thoughts Agents We employ simple PPO agents with 2 layers of 64 units (for both actor and critic network) ( [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Solved rate per complexity for all five tasks. RGoT agents (teal) are compared against the IO [PITH_FULL_IMAGE:figures/full_fig_p009_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Project Structure 13 [PITH_FULL_IMAGE:figures/full_fig_p013_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Pure Graph of Thoughts Types The related code can be found in the following repository: https://github.com/mriesen/ pure-graph-of-thoughts 14 [PITH_FULL_IMAGE:figures/full_fig_p014_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Operations of task sum list with input/output arities and prompt descriptions. Operation sum For a given list of single-digit integers, the sum is to be calculated (Listing 8). The scoring is based on whether the sum is correct. op_sum = PromptOperation ( name = ’ sum ’, n_outputs =1 , n_inputs =1 , type = OperationType . GENERATE , output_complexity = absolute_complexity (1) , prompt = Prompt ( instructio… view at source ↗
Figure 8
Figure 8. Figure 8: Definition of Operation sum Operation split A list is to be split into two equally sized sublists. The scoring is based on whether the output contains two similarly sized lists, with some lenience (sublist ratio ≥ 0.5). Operation merge Two given lists are to be merged into a single list. The result is scored based on whether the sum of the resulting list equals the sum of the input sublists and on the corr… view at source ↗
Figure 9
Figure 9. Figure 9: Operations of task sort list with input/output arities and prompt descriptions. Operation sort A given list is to be sorted in ascending order. The number of sort errors is calculated and the (binary) scoring strictly allows no errors. Operation split Identical to the Operation split of Task sum list. Operation merge Identical to the Operation merge of Task sum list. Operation branch 5 A programmatic opera… view at source ↗
Figure 10
Figure 10. Figure 10: Operations of task intersect set with input/output arities and prompt descriptions. Operation intersect The intersection of a set is to be created. The scoring is based on whether the intersection of the sets is correct. Operation split set Given two sets as input (a single state), two states are created: The first set is split in half, the subsets are fed as first set into the respective output states, i… view at source ↗
Figure 11
Figure 11. Figure 11: Operations of task count keywords with input/output arities and prompt descriptions. Operation count Given a text, the number of specific words (here: countries) is to be counted. The scoring is based on whether the words are counted correctly. Operation split Given a text, it is to be split into two substrings of similar size. Here, the scoring is dependent on whether the text is split correctly and in t… view at source ↗
Figure 12
Figure 12. Figure 12: Operations of task merge docs with input/output arities and prompt descriptions. Operation merge Several documents are to be merged into a single one, while minimizing redundancy and maximizing information retention. The scoring is based on the ROUGE-L (retention) and ROUGE-1 (redundancy) score and the resulting F1 score Lin (2004). To decide whether a result is valid, we used the threshold 0.6, which was… view at source ↗
Figure 13
Figure 13. Figure 13: Operation success probability Pˆ(c) as a function of complexity c for each task, evaluated with gpt-3.5-turbo-0125. G.1 Initial LLM Evaluation The operation sum of the task sum list was initially evaluated on different language models. The variants of gpt-3.5-turbo are quite on-par in solving the specified task. The model gpt-4-turbo-2024-04-09 maintains a higher success rate before degrading similarly ( … view at source ↗
Figure 14
Figure 14. Figure 14: LLM Evaluation Results for Operation sum of Task sum list (with n = 100) 23 [PITH_FULL_IMAGE:figures/full_fig_p023_14.png] view at source ↗
Figure 15
Figure 15. Figure 15: LLM Evaluation Results for Operation sum of Task sum list (with n = 100) 24 [PITH_FULL_IMAGE:figures/full_fig_p024_15.png] view at source ↗
Figure 16
Figure 16. Figure 16: Number of operations per complexity for all five tasks. The vertical line marks the training front, [PITH_FULL_IMAGE:figures/full_fig_p025_16.png] view at source ↗
read the original abstract

Graph of Thoughts (GoT), a generalized form of recent prompting paradigms for large language models (LLMs), has been shown to be useful for elaborate problem solving. By executing a graph of operations, thoughts of the LLM are structured as an arbitrary graph, forming the actual graph of thoughts. Originally, the graph of operations is defined manually, which requires in-depth knowledge about the solution of the problem to solve. Such a static graph of operations is rigid and therefore lacks adaptability. We propose Reinforced Graph of Thoughts (RGoT), an automated approach to the GoT prompting paradigm that leverages reinforcement learning (RL) to adaptively generate a graph of operations from a human-defined set. Results indicate that, under certain constraints, it is possible to construct graphs of operations adaptively to the task's complexity in an automated way.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 2 minor

Summary. The manuscript introduces Reinforced Graph of Thoughts (RGoT), which applies reinforcement learning to automatically construct adaptive graphs of operations for LLM prompting, starting from a human-defined set of operations. It claims that under certain constraints this yields graphs tailored to task complexity, overcoming the rigidity of manually specified static graphs in the Graph of Thoughts paradigm.

Significance. If the empirical demonstration holds with proper validation and baselines, the integration of RL for dynamic operation-graph construction could meaningfully reduce the expert knowledge required for effective LLM reasoning on complex tasks and extend structured prompting techniques. The approach directly addresses adaptability limitations in prior GoT work.

major comments (2)
  1. [Abstract] Abstract: The statement 'Results indicate that, under certain constraints, it is possible to construct graphs of operations adaptively to the task's complexity in an automated way' supplies no quantitative metrics, error bars, baseline comparisons (e.g., to static GoT or other RL-free methods), or description of the RL setup (policy, reward, state representation). This leaves the central empirical claim without visible support.
  2. [Method] The human-defined operation set is asserted to enable effective RL exploration, yet no characterization of its cardinality, diversity, or task coverage appears; without this, it is impossible to verify that the set is simultaneously rich enough to contain high-quality solutions and compact enough to avoid prohibitive sample complexity, which is load-bearing for the 'certain constraints' qualifier.
minor comments (2)
  1. [Method] Clarify the precise RL algorithm (e.g., PPO, REINFORCE) and how the graph-construction action space is represented to avoid ambiguity in reproducibility.
  2. [Experiments] Add a table summarizing the evaluated tasks, operation-set sizes, and key performance deltas versus baselines.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive feedback on our manuscript. We address each major comment point by point below, indicating where revisions will be incorporated to strengthen the presentation of our results and method.

read point-by-point responses

  1. Referee: [Abstract] Abstract: The statement 'Results indicate that, under certain constraints, it is possible to construct graphs of operations adaptively to the task's complexity in an automated way' supplies no quantitative metrics, error bars, baseline comparisons (e.g., to static GoT or other RL-free methods), or description of the RL setup (policy, reward, state representation). This leaves the central empirical claim without visible support.

    Authors: We agree that the abstract is currently high-level and does not include quantitative metrics, error bars, or explicit details on the RL setup. The full manuscript reports experimental results with baseline comparisons (including static GoT) and describes the RL components (policy gradient optimization with a reward based on task accuracy and graph efficiency) in the Experiments and Method sections. To better support the central claim, we will revise the abstract to include a concise reference to key performance gains and the RL formulation while maintaining length limits. revision: yes

  2. Referee: [Method] The human-defined operation set is asserted to enable effective RL exploration, yet no characterization of its cardinality, diversity, or task coverage appears; without this, it is impossible to verify that the set is simultaneously rich enough to contain high-quality solutions and compact enough to avoid prohibitive sample complexity, which is load-bearing for the 'certain constraints' qualifier.

    Authors: The referee correctly notes that the manuscript does not characterize the human-defined operation set in terms of cardinality, diversity, or task coverage. We will add this information in the revised Method section, including the number of operations, their coverage of reasoning primitives, and how the set size balances expressiveness against sample complexity to support the 'certain constraints' under which adaptive graph construction is feasible. revision: yes

Circularity Check

0 steps flagged

No circularity: empirical RL method proposal with no derivations or self-referential reductions

full rationale

The paper introduces RGoT as an RL-based automation of the GoT prompting paradigm, where a human-defined operation set is an explicit external input and the central claim is an empirical observation that adaptive graph construction is possible under certain constraints. No equations, derivations, fitted parameters renamed as predictions, or load-bearing self-citations appear in the abstract or described content. The method is presented as a practical extension relying on standard RL exploration rather than any closed-form reduction to its own inputs, rendering the derivation chain self-contained.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

The approach assumes a human-provided operation set is both sufficient and tractable for RL search; no free parameters, axioms, or invented entities are described in the abstract.

pith-pipeline@v0.9.0 · 5670 in / 1052 out tokens · 34509 ms · 2026-05-22T07:25:23.272466+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

41 extracted references · 41 canonical work pages · 2 internal anchors

  1. [2]

    and Le, Quoc V

    Wei, Jason and Wang, Xuezhi and Schuurmans, Dale and Bosma, Maarten and Ichter, Brian and Xia, Fei and Chi, Ed H. and Le, Quoc V. and Zhou, Denny , title =. Proceedings of the 36th International Conference on Neural Information Processing Systems , articleno =. 2024 , isbn =

    work page 2024

  2. [3]

    Advances in Neural Information Processing Systems , author =

    Tree of thoughts: Deliberate problem solving with large language models , volume =. Advances in Neural Information Processing Systems , author =

    work page

  3. [5]

    Watkins, Christopher , title =

    work page

  4. [6]

    Reinforcement Learning, Second Edition , isbn =

    Sutton, Richard S and Barto, Andrew G , year =. Reinforcement Learning, Second Edition , isbn =

    work page

  5. [7]

    Large Language Models for Mathematical Reasoning: Progresses and Challenges

    Ahn, Janice and Verma, Rishu and Lou, Renze and Liu, Di and Zhang, Rui and Yin, Wenpeng. Large Language Models for Mathematical Reasoning: Progresses and Challenges. Proceedings of the 18th Conference of the European Chapter of the Association for Computational Linguistics: Student Research Workshop. 2024

    work page 2024

  6. [8]

    , year =

    Mnih, Volodymyr and Kavukcuoglu, Koray and Silver, David and Graves, Alex and Antonoglou, Ioannis and Wierstra, Daan and Riedmiller, Martin A. , year =. Playing Atari with Deep Reinforcement Learning , volume =

    work page

  7. [9]

    Asynchronous methods for deep reinforcement learning , pages =

    Mnih, Volodymyr and Badia, Adria Puigdomenech and Mirza, Mehdi and Graves, Alex and Lillicrap, Timothy and Harley, Tim and Silver, David and Kavukcuoglu, Koray , year =. Asynchronous methods for deep reinforcement learning , pages =. International conference on machine learning , publisher =

    work page

  8. [10]

    Proximal policy optimization algorithms , journal =

    Schulman, John and Wolski, Filip and Dhariwal, Prafulla and Radford, Alec and Klimov, Oleg , year =. Proximal policy optimization algorithms , journal =

    work page

  9. [11]

    2024 , urldate =

    Models -. 2024 , urldate =

    work page 2024

  10. [12]

    Retrieval-augmented generation for knowledge-intensive NLP tasks , year =

    Lewis, Patrick and Perez, Ethan and Piktus, Aleksandra and Petroni, Fabio and Karpukhin, Vladimir and Goyal, Naman and K\". Retrieval-augmented generation for knowledge-intensive NLP tasks , year =. Proceedings of the 34th International Conference on Neural Information Processing Systems , articleno =

    work page

  11. [13]

    Reward Shaping , isbn =

    Wiewiora, Eric , editor =. Reward Shaping , isbn =. Encyclopedia of Machine Learning , publisher =. 2010 , doi =

    work page 2010

  12. [14]

    1998 , month = jan, doi =

    Matsumoto, Makoto and Nishimura, Takuji , title =. 1998 , month = jan, doi =

    work page 1998

  13. [15]

    , title =

    O'Neill, Melissa E. , title =

    work page

  14. [17]

    Welcome to Python.org , howpublished =

    work page

  15. [18]

    Gymnasium Documentation , howpublished =

    work page

  16. [19]

    Journal of Machine Learning Research , author =

    Stable-Baselines3: Reliable Reinforcement Learning Implementations , volume =. Journal of Machine Learning Research , author =

    work page

  17. [21]

    ROUGE : A Package for Automatic Evaluation of Summaries

    Lin, Chin-Yew. ROUGE : A Package for Automatic Evaluation of Summaries. Text Summarization Branches Out. 2004

    work page 2004

  18. [22]

    2504.02670 , title =

    Besta, Maciej and Paleari, Lorenzo and Jiang, Jia and Gerstenberger, Robert and Wu, You and Iff, Patrick and Kubicek, Ales and Nyczyk, Piotr and Khimey, Diana and Hannesson, Jón and Kwasniewski, Grzegorz and Copik, Marcin and Niewiadomski, Hubert and Hoefler, Torsten , year =. 2504.02670 , title =

    work page arXiv

  19. [23]

    Large language models for mathematical reasoning: Progresses and challenges

    Janice Ahn, Rishu Verma, Renze Lou, Di Liu, Rui Zhang, and Wenpeng Yin. Large language models for mathematical reasoning: Progresses and challenges. In Neele Falk, Sara Papi, and Mike Zhang (eds.), Proceedings of the 18th Conference of the European Chapter of the Association for Computational Linguistics: Student Research Workshop, pp.\ 225--237, St. Juli...

    work page 2024

  20. [24]

    Graph of thoughts: Solving elaborate problems with large language models

    Maciej Besta, Nils Blach, Ales Kubicek, Robert Gerstenberger, Lukas Gianinazzi, Joanna Gajda, Tomasz Lehmann, Michał Podstawski, Hubert Niewiadomski, Piotr Nyczyk, and Torsten Hoefler. Graph of thoughts: Solving elaborate problems with large language models. Proceedings of the AAAI Conference on Artificial Intelligence , 38 0 (16): 0 17682--17690, 2024. d...

    work page doi:10.1609/aaai.v38i16.29720 2024

  21. [25]

    Affordable ai assistants with knowledge graph of thoughts, 2025

    Maciej Besta, Lorenzo Paleari, Jia Jiang, Robert Gerstenberger, You Wu, Patrick Iff, Ales Kubicek, Piotr Nyczyk, Diana Khimey, Jón Hannesson, Grzegorz Kwasniewski, Marcin Copik, Hubert Niewiadomski, and Torsten Hoefler. Affordable ai assistants with knowledge graph of thoughts, 2025

    work page 2025

  22. [26]

    Gymnasium documentation

    Gymnasium. Gymnasium documentation. https://gymnasium.farama.org/, 2026. Accessed: Mar. 12, 2026

    work page 2026

  23. [27]

    Auto graph of thoughts: A hands-free and cost effective method for using graph of thoughts

    Thien-Loc Ha, Trong-Bao Ho, Long Nguyen, and Dien Dinh. Auto graph of thoughts: A hands-free and cost effective method for using graph of thoughts. In Proceedings of the 2024 10th International Conference on Computer Technology Applications, ICCTA '24, pp.\ 116–121, New York, NY, USA, 2024. Association for Computing Machinery. ISBN 9798400716386. doi:10.1...

    work page doi:10.1145/3674558.3674574 2024

  24. [28]

    Hancock and Taghi M

    John T. Hancock and Taghi M. Khoshgoftaar. Survey on categorical data for neural networks. Journal of Big Data, 7 0 (1): 0 28, 2020. ISSN 2196-1115. doi:10.1186/s40537-020-00305-w. URL https://doi.org/10.1186/s40537-020-00305-w

    work page doi:10.1186/s40537-020-00305-w 2020

  25. [29]

    u ttler, Mike Lewis, Wen-tau Yih, Tim Rockt\

    Patrick Lewis, Ethan Perez, Aleksandra Piktus, Fabio Petroni, Vladimir Karpukhin, Naman Goyal, Heinrich K\" u ttler, Mike Lewis, Wen-tau Yih, Tim Rockt\" a schel, Sebastian Riedel, and Douwe Kiela. Retrieval-augmented generation for knowledge-intensive nlp tasks. In Proceedings of the 34th International Conference on Neural Information Processing Systems,...

    work page 2020

  26. [30]

    ROUGE : A package for automatic evaluation of summaries

    Chin-Yew Lin. ROUGE : A package for automatic evaluation of summaries. In Text Summarization Branches Out, pp.\ 74--81, Barcelona, Spain, July 2004. Association for Computational Linguistics. URL https://aclanthology.org/W04-1013/

    work page 2004

  27. [31]

    Mersenne Twister: A 6 23-dimensionally equidistributed uniform pseudorandom number generator

    Makoto Matsumoto and Takuji Nishimura. Mersenne twister: A 623-dimensionally equidistributed uniform pseudo-random number generator. ACM Transactions on Modeling and Computer Simulation , 8 0 (1): 0 3--30, January 1998. doi:10.1145/272991.272995

    work page doi:10.1145/272991.272995 1998

  28. [32]

    Playing Atari with Deep Reinforcement Learning

    Volodymyr Mnih, Koray Kavukcuoglu, David Silver, Alex Graves, Ioannis Antonoglou, Daan Wierstra, and Martin A. Riedmiller. Playing atari with deep reinforcement learning. CoRR , abs/1312.5602, 2013. URL http://arxiv.org/abs/1312.5602

    work page internal anchor Pith review Pith/arXiv arXiv 2013

  29. [33]

    Asynchronous methods for deep reinforcement learning

    Volodymyr Mnih, Adria Puigdomenech Badia, Mehdi Mirza, Alex Graves, Timothy Lillicrap, Tim Harley, David Silver, and Koray Kavukcuoglu. Asynchronous methods for deep reinforcement learning. In International conference on machine learning, pp.\ 1928--1937. PMLR , 2016

    work page 1928

  30. [34]

    Melissa E. O'Neill. PCG : A family of simple fast space-efficient statistically good algorithms for random number generation. Technical Report HMC-CS-2014-0905, Harvey Mudd College, Claremont, CA, September 2014

    work page 2014

  31. [35]

    Models - OpenAI API

    OpenAPI Models. Models - OpenAI API . https://platform.openai.com/docs/models/overview, 2024

    work page 2024

  32. [36]

    Welcome to python.org

    Python. Welcome to python.org. https://www.python.org/, 2026. Accessed: Mar. 12, 2026

    work page 2026

  33. [37]

    Stable-baselines3: Reliable reinforcement learning implementations

    Antonin Raffin, Ashley Hill, Adam Gleave, Anssi Kanervisto, Maximilian Ernestus, and Noah Dormann. Stable-baselines3: Reliable reinforcement learning implementations. Journal of Machine Learning Research, 22 0 (268): 0 1--8, 2021. URL http://jmlr.org/papers/v22/20-1364.html

    work page 2021

  34. [38]

    Salmon, Mark A

    John K. Salmon, Mark A. Moraes, Ron O. Dror, and David E. Shaw. Parallel random numbers: as easy as 1, 2, 3. In Proceedings of 2011 International Conference for High Performance Computing, Networking, Storage and Analysis, pp.\ 1--12. ACM , 2011. ISBN 978-1-4503-0771-0. doi:10.1145/2063384.2063405. URL https://dl.acm.org/doi/10.1145/2063384.2063405

    work page doi:10.1145/2063384.2063405 2011

  35. [39]

    Proximal Policy Optimization Algorithms

    John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, and Oleg Klimov. Proximal policy optimization algorithms. arXiv preprint arXiv :1707.06347 , 2017

    work page internal anchor Pith review Pith/arXiv arXiv 2017

  36. [40]

    Developer guide --- Stable Baselines3 2.3.2 documentation

    Stable Baselines3. Developer guide --- Stable Baselines3 2.3.2 documentation. https://stable-baselines3.readthedocs.io/en/master/guide/developer.html, 2026. Accessed: Mar. 15, 2026

    work page 2026

  37. [41]

    Reinforcement Learning, Second Edition

    Richard S Sutton and Andrew G Barto. Reinforcement Learning, Second Edition. The MIT Press, 2018. ISBN 978-0-262-36401-0

    work page 2018

  38. [42]

    Learning from Delayed Rewards

    Christopher Watkins. Learning from Delayed Rewards. PhD thesis, King's College, Cambridge, 1989

    work page 1989

  39. [43]

    Chi, Quoc V

    Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Brian Ichter, Fei Xia, Ed H. Chi, Quoc V. Le, and Denny 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, Red Hook, NY, USA, 2024. Curran Associates Inc. ISBN 9781713871088

    work page 2024

  40. [44]

    Reward shaping

    Eric Wiewiora. Reward shaping. In Claude Sammut and Geoffrey I. Webb (eds.), Encyclopedia of Machine Learning, pp.\ 863--865. Springer US , 2010. ISBN 978-0-387-30164-8. doi:10.1007/978-0-387-30164-8_731. URL https://doi.org/10.1007/978-0-387-30164-8_731

    work page doi:10.1007/978-0-387-30164-8_731 2010

  41. [45]

    Tree of thoughts: Deliberate problem solving with large language models

    Shunyu Yao, Dian Yu, Jeffrey Zhao, Izhak Shafran, Tom Griffiths, Yuan Cao, and Karthik Narasimhan. Tree of thoughts: Deliberate problem solving with large language models. Advances in Neural Information Processing Systems, 36, 2024

    work page 2024