Enhancing Test-Time Scaling of Large Language Models with Hierarchical Retrieval-Augmented MCTS

Alex ZH Dou; Chaofan Tao; Dongfei Cui; Haokun Lin; Jing Xiong; Mi Zhang; Shen Yan; Xin Wang; Zhongwei Wan

REVIEW 3 major objections 5 minor 1 cited by

Enhancing Test-Time Scaling of Large Language Models with Hierarchical Retrieval-Augmented MCTS

T0 review · 3 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read R2-LLMs claims that a training-free, dual-level retrieval scheme—similar problems first, then similar solution steps during tree search—raises the math reasoning accuracy of 7B- to 8B-scale LLMs by up to 16% relative to ICL and…

desk verdict The hierarchical retrieval idea is worth a look, but the main result is exposed to a leakage risk that the paper never checks, and the out-of-domain control actually suggests the gains mostly come from retrieving near-duplicates of test problems. read the letter →

arxiv 2507.05557 v1 pith:YDIOJKHG submitted 2025-07-08 cs.CL

Alex ZH Dou , Zhongwei Wan , Dongfei Cui , Xin Wang , Jing Xiong , Haokun Lin , Chaofan Tao , Shen Yan

show 1 more author

Mi Zhang

This is my paper · ORCID

classification cs.CL

keywords test-timescalingretrieval-augmentedgenerationMonteCarloTreeSearchmathematicalreasoningin-contextlearningprocessrewardmodellargelanguagemodels

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

This paper tries to establish that test-time math reasoning of instruction-tuned LLMs can be improved by retrieving external reference material at two scales: whole similar problems before search, and similar intermediate steps during Monte Carlo Tree Search. The claimed payoff is accuracy gains of up to 16% relative on MATH500, GSM8K, and OlympiadBench-TO using LLaMA-3.1-8B, without distillation or any additional training of the policy model. A careful reader would care because the method is a plug-and-play inference-time wrapper: it only changes prompts and retrieval, so it can sit on top of existing base and process reward models. The load-bearing idea is that retrieval gives the PRM external analogical evidence, making its step-level scores more informative than scores computed from the problem alone.

What carries the argument

The mechanism that carries the argument is the combination of a conceptual unit and two retrieval-enhanced stages. The conceptual unit $T = (\texttt{ttype}, \texttt{tkey}, \texttt{tstrategy})$ is a triplet summarizing the problem's type, key terms, and a solution strategy, extracted by prompting an LLM; it is what lets the system match problems by abstract logic rather than surface wording. Deep Logical Retrieval then selects a coarse reference set via BM25 on the question-plus-type query and refines it by cosine similarity of SentenceBERT embeddings of the key terms and strategies. The second machinery is Hierarchical Augmented Reasoning MCTS: during expansion the policy model generates next steps with the retrieved reference problems in its prompt, and during evaluation the PRM scores each candidate node given a fine-grained retrieval of similar intermediate steps. The paper attributes the gain to PRM evaluations becoming more contextually informed, which reduces reward sparsity during tree search.

What would settle it

Run a contamination audit: remove from the candidate set any problem that overlaps or near-duplicates a MATH500 or GSM8K test question, re-run R2-LLMs, and compare. If the accuracy gains vanish, the reported improvement is retrieval memorization; if they persist, the analogical in-context mechanism is confirmed.

Watch

Extended reading notes

Core claim

The central claim, stated as the authors would state it, is that a hierarchical retrieval-augmented reasoning framework, R2-LLMs, improves test-time scaling of LLMs on mathematical reasoning benchmarks. At the coarse level, the framework extracts a conceptual unit from the question—its problem type, key terms, and a proposed solving strategy—and uses a two-stage retrieval process (BM25 filtering followed by SentenceBERT cosine re-ranking) to find four analogous problems whose solutions and strategies are placed in the policy model's context. At the fine level, during MCTS the framework takes the partial solution path as a query, retrieves similar reasoning steps from a reference dataset, and hands both the path and those steps to a process reward model through a meta-prompt so the node scores reflect external analogical evidence. The paper reports that this raises LLaMA-3.1-8B accuracy from 46.6% to 52.5% on MATH500 and from 82.5% to 87.4% on GSM8K, with relative gains up to 16% over ICL and tree-search baselines.

Load-bearing premise

The results assume the reference problem bank is in-domain with the test set and contains no leaked test answers or near-duplicate questions, since out-of-domain retrieval can fall below no retrieval at all.

Editorial extensions

If this is right

If the central claim is right, test-time math reasoning can be improved without distillation: any existing policy model plus a process reward model can be wrapped with retrieval and gain several absolute points on MATH500 and GSM8K.
Because the framework is training-free, gains and search costs trade off separately from the model's parameters; the paper reports only a 5–8% runtime overhead on 4 A100 GPUs.
Stronger PRMs amplify the effect: swapping Mistral-7B for Qwen2.5-7B Math PRM raises R2-LLMs on GSM8K from 87.4% to 89.7%, suggesting the retrieval works by making step evaluation more accurate, not by bypassing it.
The in-domain requirement is a practical constraint: on MATH, switching the DLR reference set from PRM800K to MATHQA drops accuracy from 52.5% to 43.5%, below the 47.5% no-retrieval baseline, so the method's value is conditional on reference-set quality.

Reading between the lines

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

A natural testable extension is applying the same two-level retrieval pattern to non-math reasoning, such as code debugging or multi-step inference in medicine or law, where 'conceptual unit' could be replaced by domain taxonomies.
The paper does not check overlap between the retrieved pool and the test set; a contamination audit would determine whether part of the gain is memorization rather than analogical generalization.
Since the fine-grained enhancement alone contributes less than the DLR set, an efficiency-minded variant could retrieve steps only when the PRM's confidence is low, saving compute on easy nodes.
The framework could be made self-improving by adding solved instances generated by the policy model itself back into the reference pool, turning static retrieval into a growing memory.

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 paper combines two known ideas—retrieval-augmented prompting and MCTS with a process reward model—into a genuinely new arrangement: coarse-level retrieval of abstract problem templates plus fine-level retrieval of solution steps during tree search. That integration is sensible, and the authors did some things right: they report ablations showing both levels contribute, they include an out-of-domain control (Section 4.4), they give runtime overhead, and they released code. The conceptual-unit extraction is a reasonable engineering contribution, not a conceptual breakthrough.

The problem is the evaluation. MATH500 is a subset of MATH, and PRM800K is built from MATH; the paper samples the retrieval candidate set from PRM800K without deduplicating against the test set. Since MATH problems are template-generated, near-duplicate variants are plausible. The fine-grained retrieval goes one step further: it retrieves solution steps from PRM800K during search, so a leaked near-duplicate can directly supply future reasoning steps. The paper never quantifies overlap. Worse, Table 3 undercuts the versatility claim: swapping in an out-of-domain pool (MATHQA for MATH) drops accuracy below the no-retrieval MCTS baseline (43.5 vs 47.5), while the in-domain PRM800K pool gives 52.5. That pattern is exactly what you would expect if most of the benefit comes from retrieving near-duplicates rather than from generic logical analogies.

The baselines are also weaker than they should be. Table 2 compares against only four tree-based methods on a different base model (LLaMA-3-8B) and has dashes for missing results; there is no comparison with large-N sampling under matched compute, nor with more recent MCTS+PRM work. No error bars anywhere. The appendix contains a red flag: Qwen-2 14B scores 62.6 on GSM8K with R2-LLMs, while Qwen-2 7B scores 89.1 on the same benchmark—either a reporting bug or an inconsistency that needs explanation.

Is the central idea dead? No. The dual-level retrieval could still help in settings with a clean, truly disjoint reference database. But as presented, the headline 16% improvement is not trustworthy until the leakage question is resolved. The paper deserves a serious referee who can push for a deduplication analysis, stronger baselines, and error bars; it should not be accepted as-is. For me, this is a conditional reject with a clear path to revision.

Referee Report

3 major / 5 minor

Summary. The paper proposes R2-LLMs, a training-free test-time scaling framework that augments Monte Carlo Tree Search with retrieval at two levels: coarse-level Deep Logical Retrieval (DLR) of similar problems via conceptual units, and fine-grained retrieval of solution steps used to guide a process reward model (PRM) during MCTS. The method is evaluated on MATH500, GSM8K, and OlympiadBench-TO with LLaMA-3.1-8B and Qwen2-7B policy models, reporting gains over zero-shot CoT, few-shot CoT, and CoT+SC@4, with additional comparisons against tree-based baselines and ablation studies.

Significance. If the central claim holds, the paper makes a useful contribution: a plug-and-play, distillation-free method that improves mathematical reasoning at inference time by combining hierarchical retrieval with MCTS, with modest runtime overhead (Table 6) and consistent gains across two policy models and three benchmarks. The paper ships a public code repository, includes ablations for each component (Table 4 and Table 8), sensitivity analyses for set sizes (Figure 2), and a control experiment with out-of-domain retrieval (Table 3). These are real strengths. However, the significance is conditional on ruling out test-set leakage from the retrieval pools, because the main gains appear precisely in the condition where contamination is most plausible.

major comments (3)

[§4.1, Table 3] The MATH500 candidate set is sampled from PRM800K, and PRM800K is constructed from the MATH dataset (Lightman et al., 2023), from which MATH500 is also drawn; the paper does not check for exact or near-duplicate overlap between the 2,500 retrieved candidates and the 500 test problems, nor does it deduplicate template-generated variants. Because Table 3 shows that replacing this pool with an out-of-domain pool (MATHQA) drops MATH accuracy below the no-retrieval baseline (43.5 vs 47.5), the observed in-domain gain (47.5 to 52.5) could be explained by answer leakage rather than by hierarchical retrieval improving generalization. I request an explicit overlap analysis and a rerun with a deduplicated, disjoint reference pool before the central claim is accepted.
[§3.4, Eq. (6), §4.1] The process reward model is Mistral-7B trained on PRM800K, and the fine-grained retrieval pool Q_fin is also drawn from PRM800K. Consequently, the increased scores R(vi,j) may reflect the PRM recognizing solution steps it was trained on rather than the quality of the current reasoning path. This should be controlled by evaluating with a PRM whose training data is disjoint from the retrieval pool, or by retrieving from a pool the PRM has not seen.
[§4.3, Table 2, §4.1] The tree-search comparison is incomplete and not compute-matched: ReST-MCTS* is missing on GSM8K and LiteSearch is missing on MATH500, and no baseline with large-N sampling (e.g., Best-of-N at a matched inference budget) is reported, so the claim that R2-LLMs improves test-time scaling over standard scaling procedures is not yet supported. Please complete the table and add a compute-controlled comparison against Best-of-N / CoT+SC at comparable total generated tokens.

minor comments (5)

[§4.1, Evaluation metrics] The evaluation metrics paragraph states both 'A solution is correct only if the model’s final answer exactly matches the ground truth' and 'A solution is deemed correct only if the final reasoning process is fully aligned with the ground truth'; these are mutually inconsistent, so please clarify the actual metric.
[§4.6] The heading 'Sensitively Analysis' should be 'Sensitivity Analysis'.
[§3.4, Eq. (6)] The notation R(vi,j) is introduced but not linked to Q(vi,j); please clarify how the retrieval-enhanced score is combined with the PRM score in node selection and backpropagation.
[§4.1] Hyperparameters such as the number of MCTS rollouts/iterations, the candidate count U, the exploration constant c in Eq. (1), and the BM25 top-N threshold used before refined selection are not reported; adding these to the appendix would improve reproducibility.
[§4.4, Table 3] The AMC 12 reference set is used in Table 3 but is not defined or cited; please add a reference and describe how it was constructed.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the paper reports an empirical retrieval-augmented MCTS pipeline with no fitted quantity that is later renamed as a prediction and no load-bearing self-citation chain.

full rationale

R2-LLMs is an inference-time algorithm, not a derivation from first principles. The conceptual-unit extraction (Eq. 2), BM25 and SentenceBERT retrieval (Eqs. 3-4), MCTS step generation conditioned on retrieved references (Eq. 5), and PRM scoring augmented by retrieved steps (Eq. 6) are all empirical components. No quantity in the method is defined in terms of the reported accuracy numbers, no parameter is fitted to a subset of the evaluation data and then 'predicted' on a closely related quantity, and no load-bearing premise rests on a self-citation or on an imported uniqueness theorem. The comparisons are against external ICL and tree-search baselines. The most plausible concern is experimental rather than circular: the MATH500 candidate pool is sampled from PRM800K while the PRM itself is trained on PRM800K, and Table 3 shows that the benefit largely disappears with an out-of-domain retrieval pool on MATH (43.5% vs. 47.5% without retrieval). That is a potential data-contamination or validity threat, but it is not a reduction of the method's output to its inputs by construction. Under the stated criteria, no circular step can be exhibited with a specific equation-level reduction, so the circularity score is 0.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The central claim rests on evaluation techniques rather than on mathematical derivation. Key assumptions include the cleanliness of in-domain retrieval pools, the reliability of the PRM as a reward signal, and the accuracy of LLM-extracted problem templates.

free parameters (3)

DLR reference set size M = 4
Selected by default; ablation shows 4 works well but no principled choice.
Selection enhancement set size K = 3
Selected by default; ablation shows modest gains at 3.
Candidate set size N = 2500
Chosen for all benchmarks; sensitivity analysis shows plateau after 2000.

assumptions (4)

domain assumption In-domain reference sets do not contain test answers or near-duplicates of test questions.
The retrieval pool for MATH500 is PRM800K, a public dataset of MATH-style questions; the paper does not check overlap between retrieved questions and the test set. If near-duplicates are retrieved, the accuracy gains could be leakage rather than reasoning improvement.
domain assumption The PRM trained on PRM800K is a reliable reward signal for guiding MCTS when augmented with retrieved steps.
The method relies on PRM scores to select nodes; if the PRM is biased toward its training distribution, retrieved steps from that distribution could inflate scores without improving genuine reasoning.
domain assumption LLM-extracted conceptual units (problem type, keywords, strategy) are accurate and useful for retrieval.
The DeepSeek-70B generated abstract templates are used to find similar questions; errors in extraction would propagate to retrieval quality.
domain assumption MCTS with a fixed rollout budget is a valid baseline and the comparison protocols match across methods.
The paper does not report compute budgets (rollouts, simulations, temperatures) for baselines, so 'fair comparison' is assumed.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Enhancing Test-Time Scaling of Large Language Models with Hierarchical Retrieval-Augmented MCTS." pith.science (2026). https://pith.science/paper/YDIOJKHG

@misc{pith2026250705557,
  author       = {Pith},
  title        = {Pith review of: Enhancing Test-Time Scaling of Large Language Models with Hierarchical Retrieval-Augmented MCTS},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/YDIOJKHG}},
  note         = {Machine review of arXiv:2507.05557}
}

read the original abstract

Test-time scaling has emerged as a promising paradigm in language modeling, leveraging additional computational resources at inference time to enhance model performance. In this work, we introduce R2-LLMs, a novel and versatile hierarchical retrieval-augmented reasoning framework designed to improve test-time scaling in large language models (LLMs) without requiring distillation from more advanced models to obtain chain-of-thought (CoT) training data. R2-LLMs enhances inference-time generalization by integrating dual-level retrieval-based in-context learning: (1) At the coarse level, our approach extracts abstract templates from complex reasoning problems and retrieves similar problem-answer pairs to facilitate high-level in-context learning; (2) At the fine level, during Monte Carlo Tree Search (MCTS), R2-LLMs efficiently retrieves analogous intermediate solution steps from reference mathematical problem datasets, refining step-wise reasoning with the aid of a process reward model (PRM) for scoring. R2-LLMs is a robust hierarchical reasoning-augmentation method that enhances in-context-level reasoning while seamlessly integrating with step-level tree search methods. Utilizing PRM, it refines both candidate generation and decision-making for improved reasoning accuracy. Empirical evaluations on the MATH500, GSM8K, and OlympiadBench-TO datasets achieve substantial relative improvement with an increase of up to 16% using LLaMA-3.1-8B compared to the baselines, showcasing the effectiveness of our approach in complex reasoning tasks.

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

MILES: Modular Instruction Memory with Learnable Selection for Self-Improving LLM Reasoning
cs.CL 2026-07 unverdicted novelty 5.0 of 10

MILES dynamically expands step-wise memory with learnable selection heads that rerank candidates and guide reasoning, improving LLM test-time performance under limited supervision.

Reference graph

Works this paper leans on

10 extracted references · 8 linked inside Pith · cited by 1 Pith paper

[3]

arXiv preprint arXiv:2305.11738

Critic: Large language models can self-correct with tool-interactive critiquing. arXiv preprint arXiv:2305.11738. Xinyu Guan, Li Lyna Zhang, Yifei Liu, Ning Shang, Youran Sun, Yi Zhu, Fan Yang, and Mao Yang

arXiv
[5]

arXiv preprint arXiv:2305.20050

Let’s verify step by step. arXiv preprint arXiv:2305.20050. Che Liu, Cheng Ouyang, Zhongwei Wan, Haozhe Wang, Wenjia Bai, and Rossella Arcucci. 2025a. Knowledge-enhanced multimodal ecg representation learning with arbitrary-lead inputs. arXiv preprint arXiv:2502.17900. Che Liu, Zhongwei Wan, Sibo Cheng, Mi Zhang, and Rossella Arcucci. 2024a. Etp: Learning...

arXiv 2024
[7]

In European Conference on Computer Vision, pages 268–278

Famba-v: Fast vision mamba with cross-layer token fusion. In European Conference on Computer Vision, pages 268–278. Springer. Hui Shen, Taiqiang Wu, Qi Han, Yunta Hsieh, Jizhou Wang, Yuyue Zhang, Yuxin Cheng, Zijian Hao, Yuan- sheng Ni, Xin Wang, et al. 2025a. Phyx: Does your model have the" wits" for physical reasoning? arXiv preprint arXiv:2505.15929. H...

arXiv 2024
[8]

arXiv preprint arXiv:2502.14317

Parallelcomp: Parallel long-context com- pressor for length extrapolation. arXiv preprint arXiv:2502.14317. Jing Xiong, Zhongwei Wan, Xiping Hu, Min Yang, and Chengming Li. 2022. Self-consistent reason- ing for solving math word problems. arXiv preprint arXiv:2210.15373. Wendong Xu, Jing Xiong, Chenyang Zhao, Qiujiang Chen, Haoran Wang, Hui Shen, Zhongwei...

arXiv 2022
[9]

Ling Yang, Zhaochen Yu, Tianjun Zhang, Shiyi Cao, Minkai Xu, Wentao Zhang, Joseph E Gonzalez, and Bin Cui

Reasonflux: Hierarchical llm reasoning via scaling thought templates. Ling Yang, Zhaochen Yu, Tianjun Zhang, Shiyi Cao, Minkai Xu, Wentao Zhang, Joseph E Gonzalez, and Bin Cui. 2024b. Buffer of thoughts: Thought- augmented reasoning with large language models. NeurIPS 2024. Shunyu Yao, Dian Yu, Jeffrey Zhao, Izhak Shafran, Tom Griffiths, Yuan Cao, and Kar...

2024
[10]

Advances in neural information processing systems, 36:11809–11822

Tree of thoughts: Deliberate problem solving with large language models. Advances in neural information processing systems, 36:11809–11822. Xiang Yue, Xingwei Qu, Ge Zhang, Yao Fu, Wen- hao Huang, Huan Sun, Yu Su, and Wenhu Chen
[2019]

arXiv preprint arXiv:1905.13319

Mathqa: Towards interpretable math word problem solving with operation-based formalisms. arXiv preprint arXiv:1905.13319. Bradley Brown, Jordan Juravsky, Ryan Ehrlich, Ronald Clark, Quoc V Le, Christopher Ré, and Azalia Mirho- seini. 2024. Large language monkeys: Scaling infer- ence compute with repeated sampling. arXiv preprint arXiv:2407.21787. Canyu Ch...

arXiv 1905
[2023]

arXiv preprint arXiv:2309.17179

Alphazero-like tree-search can guide large lan- guage model decoding and training. arXiv preprint arXiv:2309.17179. Charles R Fletcher. 1985. Understanding and solving arithmetic word problems: A computer simulation. Behavior Research Methods, Instruments, & Comput- ers, 17(5):565–571. Zixuan Gong, Guangyin Bao, Qi Zhang, Zhongwei Wan, Duoqian Miao, Shouj...

arXiv 1985

Show all 10 references

[2024]

arXiv preprint arXiv:2410.08146

Rewarding progress: Scaling automated pro- cess verifiers for llm reasoning. arXiv preprint arXiv:2410.08146. Zhihong Shao, Peiyi Wang, Qihao Zhu, Runxin Xu, Junxiao Song, Xiao Bi, Haowei Zhang, Mingchuan Zhang, YK Li, Y Wu, et al. 2024. Deepseekmath: Pushing the limits of mat...

arXiv
[2025]

arXiv preprint arXiv:2501.04519

rstar-math: Small llms can master math reason- ing with self-evolved deep thinking. arXiv preprint arXiv:2501.04519. Daya Guo, Dejian Yang, Haowei Zhang, Junxiao Song, Ruoyu Zhang, Runxin Xu, Qihao Zhu, Shirong Ma, Peiyi Wang, Xiao Bi, et al. 2025. Deepseek-r1: In- centivizing...

2025 arXiv

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.

[1] [3]

arXiv preprint arXiv:2305.11738

Critic: Large language models can self-correct with tool-interactive critiquing. arXiv preprint arXiv:2305.11738. Xinyu Guan, Li Lyna Zhang, Yifei Liu, Ning Shang, Youran Sun, Yi Zhu, Fan Yang, and Mao Yang

arXiv

[2] [5]

arXiv preprint arXiv:2305.20050

Let’s verify step by step. arXiv preprint arXiv:2305.20050. Che Liu, Cheng Ouyang, Zhongwei Wan, Haozhe Wang, Wenjia Bai, and Rossella Arcucci. 2025a. Knowledge-enhanced multimodal ecg representation learning with arbitrary-lead inputs. arXiv preprint arXiv:2502.17900. Che Liu, Zhongwei Wan, Sibo Cheng, Mi Zhang, and Rossella Arcucci. 2024a. Etp: Learning...

arXiv 2024

[3] [7]

In European Conference on Computer Vision, pages 268–278

Famba-v: Fast vision mamba with cross-layer token fusion. In European Conference on Computer Vision, pages 268–278. Springer. Hui Shen, Taiqiang Wu, Qi Han, Yunta Hsieh, Jizhou Wang, Yuyue Zhang, Yuxin Cheng, Zijian Hao, Yuan- sheng Ni, Xin Wang, et al. 2025a. Phyx: Does your model have the" wits" for physical reasoning? arXiv preprint arXiv:2505.15929. H...

arXiv 2024

[4] [8]

arXiv preprint arXiv:2502.14317

Parallelcomp: Parallel long-context com- pressor for length extrapolation. arXiv preprint arXiv:2502.14317. Jing Xiong, Zhongwei Wan, Xiping Hu, Min Yang, and Chengming Li. 2022. Self-consistent reason- ing for solving math word problems. arXiv preprint arXiv:2210.15373. Wendong Xu, Jing Xiong, Chenyang Zhao, Qiujiang Chen, Haoran Wang, Hui Shen, Zhongwei...

arXiv 2022

[5] [9]

Ling Yang, Zhaochen Yu, Tianjun Zhang, Shiyi Cao, Minkai Xu, Wentao Zhang, Joseph E Gonzalez, and Bin Cui

Reasonflux: Hierarchical llm reasoning via scaling thought templates. Ling Yang, Zhaochen Yu, Tianjun Zhang, Shiyi Cao, Minkai Xu, Wentao Zhang, Joseph E Gonzalez, and Bin Cui. 2024b. Buffer of thoughts: Thought- augmented reasoning with large language models. NeurIPS 2024. Shunyu Yao, Dian Yu, Jeffrey Zhao, Izhak Shafran, Tom Griffiths, Yuan Cao, and Kar...

2024

[6] [10]

Advances in neural information processing systems, 36:11809–11822

Tree of thoughts: Deliberate problem solving with large language models. Advances in neural information processing systems, 36:11809–11822. Xiang Yue, Xingwei Qu, Ge Zhang, Yao Fu, Wen- hao Huang, Huan Sun, Yu Su, and Wenhu Chen

[7] [2019]

arXiv preprint arXiv:1905.13319

Mathqa: Towards interpretable math word problem solving with operation-based formalisms. arXiv preprint arXiv:1905.13319. Bradley Brown, Jordan Juravsky, Ryan Ehrlich, Ronald Clark, Quoc V Le, Christopher Ré, and Azalia Mirho- seini. 2024. Large language monkeys: Scaling infer- ence compute with repeated sampling. arXiv preprint arXiv:2407.21787. Canyu Ch...

arXiv 1905

[8] [2023]

arXiv preprint arXiv:2309.17179

Alphazero-like tree-search can guide large lan- guage model decoding and training. arXiv preprint arXiv:2309.17179. Charles R Fletcher. 1985. Understanding and solving arithmetic word problems: A computer simulation. Behavior Research Methods, Instruments, & Comput- ers, 17(5):565–571. Zixuan Gong, Guangyin Bao, Qi Zhang, Zhongwei Wan, Duoqian Miao, Shouj...

arXiv 1985