CheMatAgent: Enhancing LLMs for Chemistry and Materials Science through Tree-Search Based Tool Learning

Binbin Lin; Dongzhan Zhou; Mengsong Wu; Tong Xie; Wenliang Chen; Yafei Wang; Yidong Ming; Yuqi An; Yuqiang Li; Yuwei Wan

REVIEW 3 major objections 5 minor 1 cited by

CheMatAgent: Enhancing LLMs for Chemistry and Materials Science through Tree-Search Based Tool Learning

T0 review · 3 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read CheMatAgent shows that decoupling tool planning from tool execution and self-training on tree-search trajectories lifts an LLM-based chemistry agent above GPT-4o on tool-calling benchmarks.

desk verdict Useful chemistry tool pool and benchmark, but the evaluation is too entangled with the training data to support the headline numbers. read the letter →

arxiv 2506.07551 v2 pith:32HO7WYF submitted 2025-06-09 cs.LG cs.AIcs.CEcs.CL

Mengsong Wu , YaFei Wang , Yidong Ming , Yuqi An , Yuwei Wan , Wenliang Chen , Binbin Lin , Yuqiang Li

show 2 more authors

Tong Xie Dongzhan Zhou

This is my paper

classification cs.LGcs.AIcs.CEcs.CL

keywords LLMagenttoollearningchemistrymaterialsscienceMonteCarloTreeSearchprocessrewardmodelbenchmarkself-training

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

The paper sets out to show that an LLM-based agent for chemistry and materials science becomes substantially more capable when it is given a large, curated toolbox of 137 external tools and trained to plan and execute tool calls separately. To support this, the authors build ChemToolBench, a dataset of single- and multi-tool queries with detailed step-level labels, and HE-MCTS, a hierarchical Monte Carlo Tree Search framework that decouples tool selection from parameter generation. Using self-generated search trajectories, they fine-tune the policy model and train process and outcome reward models, all without manual annotation. They report that the resulting agent beats a frontier commercial model and several open-source baselines on tool selection, parameter filling, and final-answer pass rate across the Comprehensive Chemistry and Materials Science splits. The claim matters because it points to a way of keeping chemistry LLMs current with specialized knowledge through tool integration instead of retraining on fresh data.

What carries the argument

The central mechanism is HE-MCTS (Hierarchical Evolutionary Monte Carlo Tree Search), a search framework in which each tree node is a partial tool-selection trajectory. A policy model proposes tool calls, a separate execution model fills in parameter values and returns results, and trained process and outcome reward models score intermediate states and final answers. The search data generated by this loop is then reused to fine-tune the policy and the critics, closing the cycle without manual annotation.

What would settle it

Construct a test set of chemistry queries that explicitly allow two or more distinct, chemically equivalent tool sequences. If the CheMatAgent advantage over GPT-4o shrinks substantially on such a test, or if the PRM-marked 'correct' steps are not the ones a domain expert would choose, then the reported gains largely follow from the single-chain labeling rather than from better chemistry tool use.

Watch

Extended reading notes

Core claim

The central claim is that decoupling tool planning from tool execution, and guiding both with self-trained reward models, makes an LLM-based chemistry agent reliably pick the right tools and fill in their parameters. The paper shows that a policy model that explores multiple tool-calling paths with HE-MCTS and is then fine-tuned on step-level data, including noisy and multi-path trajectories, surpasses a leading commercial model on ChemToolBench. The authors attribute the improvement to three mechanisms: the separate execution model operates in a much smaller action space and therefore fills parameters more accurately; the enhanced step-level fine-tuning teaches the policy to tolerate noise and to follow correct tool-selection strategies even after errors; and the task-adaptive process and outcome reward models cut off unproductive search branches. The reported gains appear on both single- and multiple-tool-calling tasks and generalize across the chemistry and materials splits.

Load-bearing premise

The entire training and evaluation rests on the assumption that every query in ChemToolBench has one correct 'standard tool invocation chain', so any step that does not match that chain is counted as an error, even if the alternative chain would still solve the problem.

Editorial extensions

If this is right

Chemistry and materials questions can be answered by an agent that looks up and invokes specialized tools, reducing the impact of outdated pretraining knowledge.
Decoupling tool selection from parameter generation suggests that execution accuracy improves when the action space is narrowed, a design that may transfer to other tool-using domains.
Self-generated tree-search trajectories can serve as a training signal for both the policy and the reward models, offering a path to autonomous agent improvement.
The ChemToolBench benchmark provides a fine-grained way to evaluate process-level reasoning (tool selection and parameter filling) alongside final answer quality, which could become a standard evaluation for scientific tool agents.

Reading between the lines

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

An implication the authors leave implicit is that the single-'standard chain' labeling in ChemToolBench may conflate 'the route the dataset happened to record' with 'the correct route'. If a query has several chemically valid tool sequences, both the PRM training labels and the P/R/F1 evaluation would penalize legitimate alternatives, so part of the reported improvement over baselines may come fro
The decoupling principle could be tested outside chemistry: the same hierarchical policy/execution split with self-trained reward models might improve tool agents in other specialized domains such as bioinformatics or materials synthesis, where parameter precision is often the bottleneck.
A concrete extension would be to release a version of ChemToolBench with multiple accepted tool chains per query and measure how much the agent's advantage shrinks; that would separate tool-use quality from dataset-conformity.

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 gives the community two genuinely useful things: a 137-tool chemistry/materials tool pool and ChemToolBench, a new benchmark with a described curation pipeline. The decoupled policy/execution design in HE-MCTS is a sensible engineering choice, and the code and data are released. That is real value.

The problem is the evaluation. The PRM labels, the step-level fine-tuning filters, and the P/R/F1 metrics in Section 3.2 all lean on the same "standard tool invocation chains" that were LLM-generated with manual checks. So the training signal and the scoring metric are aligned to the same single-solution path. If a query has more than one valid tool chain, or if a reference chain has an error, the model is being optimized toward the dataset's path, not toward chemically better tool use. The paper also does not compare against prior chemistry agents like ChemCrow or CACTUS, reports no error bars or significance tests, and the "no manual annotation" claim is contradicted by its own dataset construction details. The "surpass GPT-4o" claim is presented vaguely and is not directly evidenced by the tables.

That said, the core idea is plausible and the framework is not incoherent. The circularity is a serious flaw, but it is fixable. An external benchmark, or at least a held-out set of real chemistry tasks with independent ground truth, would resolve the main worry. Some of the smaller claims, like whether the PRM/ORM truly beat GPT-4o as critics, need clearer experimental framing.

This paper deserves a serious referee. The tool pool and benchmark are contributions worth engaging with, and the framework is reasonable. But I would not cite the performance numbers as they stand. The authors should be pushed to separate training and evaluation sources, add statistical rigor, and benchmark against existing chemistry agents. If they do that, this could become a solid reference point for LLM tool use in chemistry.

Referee Report

3 major / 5 minor

Summary. CheMatAgent integrates 137 chemistry/materials tools, introduces a new dataset ChemToolBench, and proposes a Hierarchical Evolutionary Monte Carlo Tree Search (HE-MCTS) framework with decoupled policy and execution models. The paper claims that step-level fine-tuning on self-generated HE-MCTS data and training task-adaptive PRM/ORM 'surpass GPT-4o' and 'significantly improve' performance in chemistry QA and discovery tasks. The evaluation is conducted entirely on ChemToolBench, using P/R/F1 against reference tool invocation chains and GPT-4o-judged pass rates, with no external benchmark or comparison to prior chemistry tool agents.

Significance. If the claims are validated independently, the paper would contribute a large, organized chemistry tool pool, a benchmark for tool learning, and a search-based framework that decouples tool planning from execution. The release of code and data is a concrete strength. The core significance, however, hinges on whether the reported gains reflect genuine improvements in tool use rather than overfitting to the dataset's reference chains, because the training labels, reward-model supervision, and evaluation metrics all share the same reference standard. The decoupled architecture and use of learned PRM/ORM are interesting and worth further study, but the current evidence is insufficient to establish the headline claim of surpassing GPT-4o.

major comments (3)

[Section 2.3.4, Section 3.2] The PRM labels are defined as vi=1 iff action ai aligns with the standard tool invocation chain, and Section 3.2 computes P/R/F1 against those same standard chains. Additionally, Section 2.3.3 filters HE-MCTS paths by consistency with the standard chain, and Table 5 shows that the policy for M2/M3 is trained on Dp derived from the meta-dataset. Thus the training signal for the policy, the PRM/ORM supervision, and the evaluation metric all share the same reference solution. This circularity means the improvements of -M2 and -M3 over GPT-4o-mini and Claude-3.5-S in Table 2 may reflect memorization of the reference chains rather than chemically better tool use. The Limitations section acknowledges only computational overhead and pretraining obsolescence, not this issue. To support the central claim, the authors should evaluate on an external benchmark (e.g., SciToolBench or ChemCrow tasks) or provide a human-expert evaluation on a held-out set of queries that permits multiple valid tool chains.
[Section 2.3.3, Section 3.2] The evaluation assumes a unique 'standard tool invocation chain' per query, but Section 2.3.3 explicitly states that LLMs can exhibit multiple valid reasoning paths and that multiplicity arises from parallel tools and interchangeable orderings. A correct but non-canonical tool sequence will be scored as incorrect in tool-selection P/R/F1. The paper does not report how often equivalent valid chains exist, nor does it adapt the scoring to accept any valid chain. This makes the quantitative comparisons in Tables 2-4 potentially unfair and the magnitude of the reported gains difficult to interpret. The authors should either provide evidence that the reference chains are exhaustive on a sample of queries or change the metric to be chain-agnostic.
[Section 3.3, Tables 2-4] The experimental section reports point estimates without error bars or significance tests, and it omits existing chemistry tool agents such as ChemCrow, CACTUS, and SciAgent as baselines, despite their discussion in Related Works. For example, in Table 2 the differences between Llama-3.1-8B-M2 and M3 in Tool F1 (90.79 vs 92.47) and Param F1 (86.36 vs 85.41) are small relative to likely run-to-run variance, so the claim that 'M3 models outperform M2 models' is not statistically substantiated. The authors should include multiple seeds or bootstrap confidence intervals and at least one prior-agent baseline to support the claim of significant improvement.

minor comments (5)

[Abstract] The phrase 'autonomous optimization without manual annotation' is overstated, because the meta-dataset construction involves manual checks (Sections 2.2.2 and 2.2.3) and the PRM/ORM training uses GPT-generated judgments in Section 2.3.4. Please rephrase to clarify that the self-training phase itself does not require manual annotation, while the initial dataset does.
[Appendix B.2] The adaptive pruning mechanism introduces hyperparameters tau_0, lambda, Dmax, Dearly, kappa, epsilon, alpha, and beta, but the paper does not report their values or a sensitivity analysis. This hinders reproducibility.
[Table 2] The table headers are difficult to parse because the columns are not clearly separated; the meaning of 'Format', 'Tool', 'Param', and 'Return' is ambiguous. Please restructure the table with clear subheaders and a legend.
[Section 2.1] The claim of 'largest tool pool' is based on counting pymatgen functions as separate tools; please clarify how the 137 tools are counted and whether this is a meaningful comparison to ChemCrow's 18 tools.
[Section 2.2.1] The description of parameter-example generation is terse; please provide more detail on how privacy-related parameters are handled and how diversity is ensured across generated examples.

Circularity Check

2 steps flagged · score 6.0 of 10

The training labels for step-level FT and the PRM/ORM, and the P/R/F1 benchmark metrics, are all defined by alignment with ChemToolBench's LLM-generated 'standard tool invocation chains,' so the reported gains over GPT-4o partly measure self-consistency with the training annotation rather than independent chemistry tool-use quality.

fitted input called prediction [Sections 2.3.3-2.3.4 and 3.2]
"Specifically, if ai aligns with the standard tool invocation chain, vi is 1; otherwise, vi is 0. ... (2) r2 L, derived from rule-based correctness evaluation of the sequence [q, a1, o1, . . . , aL−1, oL−1] using meta-dataset. ... To provide a fine-grained analysis of the agent's reasoning capability, we compute Precision, Recall, andF1-scoreof tool selection and parameter filling-in."

The PRM/ORM supervision labels and the evaluation metrics are both defined by the same ChemToolBench reference chains. Step-FT data are filtered to those chains (Section 2.3.3), PRM labels are 1 iff a step aligns with the standard chain, and ORM reward r2 is computed from the meta-dataset. The P/R/F1 scores in Tables 2-4 then measure how well the trained models reproduce that same annotation scheme. Because the optimization target and the reported benchmark are the same construct, the claim that PRM/ORM 'surpass GPT-4o' is partly a self-consistency result, not an independently verified chemistry tool-use improvement.
self definitional [Section 2.3.3, 'Multi-Path Reasoning and Noise Filtering Strategy']
"For multi-step tool invocation tasks, LLMs can exhibit multiple valid reasoning paths. ... To address this, we filter reasoning paths with meta-dataset, enforcing consistency between each node and the standard invocation chain, ensuring noise-free training labels."

The paper acknowledges multiple valid reasoning paths, then defines 'noise-free' as conformity to one LLM-generated 'standard invocation chain' and uses that same chain as the evaluation ground truth. Correctness is therefore defined as alignment with the dataset's single solution path, and any valid alternative tool sequence is scored as incorrect. This makes the benchmark's P/R/F1 a measure of agreement with the training annotation rather than an external measure of whether the agent solved the chemistry task.

full rationale

The central claim is partially circular, but not fully. CheMatAgent's policy, PRM, and ORM are optimized using labels derived from ChemToolBench's LLM-generated reference chains, and the reported P/R/F1 and Pass Rate are computed against the same benchmark construct. There is no external benchmark or independent ground truth to show that the learned tool-invocation behavior is chemically better, rather than better aligned with the dataset's canonical path. This is not a case of self-citation load-bearing circularity: the HE-MCTS components are adapted from external prior work, and the only self-citations (Seal-Tools, ChemLLM) are not used to justify the central result. The Limitations section acknowledges only computational overhead and pretraining obsolescence, not the shared-label issue. Because the test split is held out and the fine-tuned models must still generalize to unseen queries, the circularity is partial (score 6), not total.

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

The method rests on a large set of hand-chosen search hyperparameters and on domain assumptions about tool reliability, the uniqueness of the canonical tool chain, and the validity of GPT-4o as a judge. These are not independently validated, and the canonical-chain assumption is load-bearing for both training and evaluation.

free parameters (3)

UCT exploration coefficient C = not reported
Controls exploration-exploitation balance in tree search selection (Eq. 2).
Adaptive pruning thresholds (tau_0, lambda, Dmax, Dearly, kappa, epsilon) = not reported
Used in node score thresholding, soft pruning, and fast recovery (Appendix B.2). These hand-chosen hyperparameters affect search behavior and hence final results.
Node score weights alpha and beta = not reported
Combine value and information-gain in the node importance score I(s_i) (Eq. 6).

assumptions (4)

domain assumption The 137 tools collected from existing packages execute correctly in the local environment and return accurate results.
The entire evaluation depends on tool outputs being reliable. Mentioned in Section 2.1.3 after code refinement, but no testing or error analysis is provided.
domain assumption The manually verified 'standard tool invocation chain' in ChemToolBench is the unique correct solution for each query.
Used as ground truth for filtering training paths and computing P/R/F1 and PRM labels (Sections 2.3.4 and 3.2). If multiple valid chains exist, this assumption biases the benchmark.
domain assumption GPT-4o's judgement of answer consistency with the reference answer is a valid measure of final answer correctness.
Pass Rate relies entirely on GPT-4o as judge (Section 3.2).
standard math MCTS with the proposed heuristics improves reasoning efficiency and quality.
The framework assumes UCB and information-gain based pruning are appropriate for LLM decoding; prior works are cited but no theoretical justification is given.

how reviews work

0 comments

Cite this review

Pith. "Pith review of CheMatAgent: Enhancing LLMs for Chemistry and Materials Science through Tree-Search Based Tool Learning." pith.science (2026). https://pith.science/paper/32HO7WYF

@misc{pith2026250607551,
  author       = {Pith},
  title        = {Pith review of: CheMatAgent: Enhancing LLMs for Chemistry and Materials Science through Tree-Search Based Tool Learning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/32HO7WYF}},
  note         = {Machine review of arXiv:2506.07551}
}

read the original abstract

Large language models (LLMs) have recently demonstrated promising capabilities in chemistry tasks while still facing challenges due to outdated pretraining knowledge and the difficulty of incorporating specialized chemical expertise. To address these issues, we propose an LLM-based agent that synergistically integrates 137 external chemical tools created ranging from basic information retrieval to complex reaction predictions, and a dataset curation pipeline to generate the dataset ChemToolBench that facilitates both effective tool selection and precise parameter filling during fine-tuning and evaluation. We introduce a Hierarchical Evolutionary Monte Carlo Tree Search (HE-MCTS) framework, enabling independent optimization of tool planning and execution. By leveraging self-generated data, our approach supports step-level fine-tuning (FT) of the policy model and training task-adaptive PRM and ORM that surpass GPT-4o. Experimental evaluations demonstrate that our approach significantly improves performance in Chemistry QA and discovery tasks, offering a robust solution to integrate specialized tools with LLMs for advanced chemical applications. All datasets and code are available at https://github.com/AI4Chem/ChemistryAgent .

Figures

Figures reproduced from arXiv: 2506.07551 by the authors.

**Figure 1.** Overview of our CheMatAgent. cution feedbacks to enhance accuracy. Additionally, we leverage self-generated HE-MCTS data alongside the meta-dataset to perform step-level fine-tuning on the policy model, and train taskadaptive PRM and ORM as alternatives to GPT4o. Crucially, this training process requires no manual annotation or curation. The self-evolving agent, guided by HE-MCTS, autonomously optimizes its perf… view at source ↗

**Figure 2.** Domain-specific Tool Learning dataset construction pipeline. [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗

**Figure 3.** HE-MCTS pipeline.The left part presents the process of Search-Based Hierarchical inferring process. The [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗

Figures from the paper (3 more)

**Figure 4.** Figure 4: contrast of Dp and D˜ p graph(DAG). Leveraging this structure, an alternative approach is reordering interchangeable tools in meta-dataset, while using GPT to ensure coherent reasoning within each invocation chain. Robustness Reasoning and Noise Retention Strategy: In…

**Figure 5.** Figure 5: H-MCTS process Expansion: Policy Model generates k child nodes. To enhance the efficiency of exploration, T is constrained to retrieved tools. To mitigate redundancy, uniqueness is enforced among child nodes, ensuring parent node does not generate duplicate children. F…

**Figure 6.** Figure 6: Tool Integration Procedure in 3 steps [PITH_FULL_IMAGE:figures/full_fig_p015_6.png]

Discussion (0). Sign in 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

Open, Reliable, and Collective: A Community-Driven Framework for Tool-Using AI Agents
cs.AI 2026-03 conditional novelty 6.0 of 10

A community-maintained, schema-standardized toolbox with continuous intrinsic-tool evaluation yields 6–22% relative gains over a curated baseline across multiple LLM agent architectures.

Reference graph

Works this paper leans on

32 extracted references · 9 canonical work pages · cited by 1 Pith paper

[1]

online" 'onlinestring :=

ENTRY address archivePrefix author booktitle chapter edition editor eid eprint eprinttype howpublished institution journal key month note number organization pages publisher school series title type volume year doi pubmed url lastchecked label extra.label sort.label short.list INTEGERS output.state before.all mid.sentence after.sentence after.block STRING...
[2]

write newline

" write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION word.in bbl.in capitalize " " * FUNCT...
[3]

Andres M Bran, Sam Cox, Oliver Schilter, Carlo Baldassari, Andrew D White, and Philippe Schwaller. 2023. Chemcrow: Augmenting large-language models with chemistry tools. arXiv preprint arXiv:2304.05376

arXiv 2023
[4]

Guoxin Chen, Minpeng Liao, Chengxi Li, and Kai Fan. 2024 a . Alphamath almost zero: process supervision without process. arXiv preprint arXiv:2405.03553

arXiv 2024
[5]

Guoxin Chen, Minpeng Liao, Chengxi Li, and Kai Fan. 2024 b . Step-level value preference optimization for mathematical reasoning. arXiv preprint arXiv:2406.10858

arXiv 2024
[6]

Sijia Chen, Yibo Wang, Yi-Feng Wu, Qing-Guo Chen, Zhao Xu, Weihua Luo, Kaifu Zhang, and Lijun Zhang. 2024 c . Advancing tool-augmented large language models: Integrating insights from errors in inference trees. arXiv preprint arXiv:2406.07115

arXiv 2024
[7]

Sylvain Gelly and David Silver. 2011. Monte-carlo tree search and rapid action value estimation in computer go. Artificial Intelligence, 175(11):1856--1875

work page 2011
[8]

Shibo Hao, Tianyang Liu, Zhen Wang, and Zhiting Hu. 2024. Toolkengpt: Augmenting frozen language models with massive tools via tool embeddings. Advances in neural information processing systems, 36

work page 2024

Show all 32 references

[9]

Levente Kocsis and Csaba Szepesv \'a ri. 2006. Bandit based monte-carlo planning. In European conference on machine learning, pages 282--293. Springer

2006
[10]

Minghao Li, Yingxiu Zhao, Bowen Yu, Feifan Song, Hangyu Li, Haiyang Yu, Zhoujun Li, Fei Huang, and Yongbin Li. 2023. Api-bank: A comprehensive benchmark for tool-augmented llms. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, pages 3102--3116

2023
[11]

Weiwen Liu, Xu Huang, Xingshan Zeng, Xinlong Hao, Shuai Yu, Dexun Li, Shuai Wang, Weinan Gan, Zhengying Liu, Yuanqing Yu, et al. 2024. Toolace: Winning the points of llm function calling. arXiv preprint arXiv:2409.00920

2024 arXiv
[12]

Yubo Ma, Zhibin Gou, Junheng Hao, Ruochen Xu, Shuohang Wang, Liangming Pan, Yujiu Yang, Yixin Cao, Aixin Sun, Hany Awadalla, et al. 2024. Sciagent: Tool-augmented language models for scientific reasoning. arXiv preprint arXiv:2402.11451

2024 arXiv
[13]

Andrew D McNaughton, Gautham Krishna Sankar Ramalaxmi, Agustin Kruel, Carter R Knutson, Rohith A Varikoti, and Neeraj Kumar. 2024. Cactus: Chemistry agent connecting tool usage to science. ACS omega, 9(46):46563--46573

2024
[14]

Adrian Mirza, Nawaf Alampara, Sreekanth Kunchapu, Marti \ n o R \' os-Garc \' a, Benedict Emoekabu, Aswanth Krishnan, Tanya Gupta, Mara Schilling-Wilhelmi, Macjonathan Okereke, Anagha Aneesh, et al. 2024. Are large language models superhuman chemists? arXiv preprint arXiv:2404.01475

2024 arXiv
[15]

Shyue Ping Ong, William Davidson Richards, Anubhav Jain, Geoffroy Hautier, Michael Kocher, Shreyas Cholia, Dan Gunter, Vincent L Chevrier, Kristin A Persson, and Gerbrand Ceder. 2013. Python materials genomics (pymatgen): A robust, open-source python library for materials anal...

2013
[16]

Aaron Parisi, Yao Zhao, and Noah Fiedel. 2022. Talm: Tool augmented language models. arXiv preprint arXiv:2205.12255

2022 arXiv
[17]

Judea Pearl. 1984. Heuristics: intelligent search strategies for computer problem solving. Addison-Wesley Longman Publishing Co., Inc

1984
[18]

Yujia Qin, Shengding Hu, Yankai Lin, Weize Chen, Ning Ding, Ganqu Cui, Zheni Zeng, Xuanhe Zhou, Yufei Huang, Chaojun Xiao, et al. 2024. Tool learning with foundation models. ACM Computing Surveys, 57(4):1--40

2024
[19]

Yujia Qin, Shihao Liang, Yining Ye, Kunlun Zhu, Lan Yan, Yaxi Lu, Yankai Lin, Xin Cong, Xiangru Tang, Bill Qian, et al. 2023. Toolllm: Facilitating large language models to master 16000+ real-world apis. arXiv preprint arXiv:2307.16789

2023 arXiv
[20]

Timo Schick, Jane Dwivedi-Yu, Roberto Dess \` , Roberta Raileanu, Maria Lomeli, Eric Hambro, Luke Zettlemoyer, Nicola Cancedda, and Thomas Scialom. 2024. Toolformer: Language models can teach themselves to use tools. Advances in Neural Information Processing Systems, 36

2024
[21]

Yongliang Shen, Kaitao Song, Xu Tan, Dongsheng Li, Weiming Lu, and Yueting Zhuang. 2024. Hugginggpt: Solving ai tasks with chatgpt and its friends in hugging face. Advances in Neural Information Processing Systems, 36

2024
[22]

David Silver, Aja Huang, Chris J Maddison, Arthur Guez, Laurent Sifre, George Van Den Driessche, Julian Schrittwieser, Ioannis Antonoglou, Veda Panneershelvam, Marc Lanctot, et al. 2016. Mastering the game of go with deep neural networks and tree search. nature, 529(7587):484--489

2016
[23]

Yifan Song, Da Yin, Xiang Yue, Jie Huang, Sujian Li, and Bill Yuchen Lin. 2024. Trial and error: Exploration-based trajectory optimization for llm agents. arXiv preprint arXiv:2403.02502

2024 arXiv
[24]

Peter Spirtes and Clark Glymour. 1991. An algorithm for fast recovery of sparse causal graphs. Social science computer review, 9(1):62--72

1991
[25]

Ye Tian, Baolin Peng, Linfeng Song, Lifeng Jin, Dian Yu, Haitao Mi, and Dong Yu. 2024. Toward self-improvement of llms via imagination, searching, and criticizing. arXiv preprint arXiv:2404.12253

2024 arXiv
[26]

Ziyu Wan, Xidong Feng, Muning Wen, Stephen Marcus McAleer, Ying Wen, Weinan Zhang, and Jun Wang. 2024. Alphazero-like tree-search can guide large language model decoding and training. In Forty-first International Conference on Machine Learning

2024
[27]

Mengsong Wu, Tong Zhu, Han Han, Chuanyuan Tan, Xiang Zhang, and Wenliang Chen. 2024 a . Seal-tools: Self-instruct tool learning dataset for agent tuning and detailed benchmark. In CCF International Conference on Natural Language Processing and Chinese Computing, pages 372--384...

2024
[28]

Qinzhuo Wu, Wei Liu, Jian Luan, and Bin Wang. 2024 b . Toolplanner: A tool augmented llm for multi granularity instructions with path planning and feedback. arXiv preprint arXiv:2409.14826

2024 arXiv
[29]

Dongyu Xue, Han Zhang, Dongling Xiao, Yukang Gong, Guohui Chuai, Yu Sun, Hao Tian, Hua Wu, Yukun Li, and Qi Liu. 2020. X-mol: large-scale pre-training for molecular understanding and diverse molecular analysis. bioRxiv, pages 2020--12

2020
[30]

Dan Zhang, Sining Zhoubian, Yisong Yue, Yuxiao Dong, and Jie Tang. 2024 a . Rest-mcts*: Llm self-training via process reward guided tree search. arXiv preprint arXiv:2406.03816

2024 arXiv
[31]

Di Zhang, Wei Liu, Qian Tan, Jingdan Chen, Hang Yan, Yuliang Yan, Jiatong Li, Weiran Huang, Xiangyu Yue, Dongzhan Zhou, et al. 2024 b . Chemllm: A chemical large language model. arXiv preprint arXiv:2402.06852

2024 arXiv
[32]

Xuan Zhang, Chao Du, Tianyu Pang, Qian Liu, Wei Gao, and Min Lin. 2024 c . Chain of preference optimization: Improving chain-of-thought reasoning in llms. arXiv preprint arXiv:2406.09136

2024 arXiv

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.

[1] [1]

online" 'onlinestring :=

ENTRY address archivePrefix author booktitle chapter edition editor eid eprint eprinttype howpublished institution journal key month note number organization pages publisher school series title type volume year doi pubmed url lastchecked label extra.label sort.label short.list INTEGERS output.state before.all mid.sentence after.sentence after.block STRING...

[2] [2]

write newline

" write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION word.in bbl.in capitalize " " * FUNCT...

[3] [3]

Andres M Bran, Sam Cox, Oliver Schilter, Carlo Baldassari, Andrew D White, and Philippe Schwaller. 2023. Chemcrow: Augmenting large-language models with chemistry tools. arXiv preprint arXiv:2304.05376

arXiv 2023

[4] [4]

Guoxin Chen, Minpeng Liao, Chengxi Li, and Kai Fan. 2024 a . Alphamath almost zero: process supervision without process. arXiv preprint arXiv:2405.03553

arXiv 2024

[5] [5]

Guoxin Chen, Minpeng Liao, Chengxi Li, and Kai Fan. 2024 b . Step-level value preference optimization for mathematical reasoning. arXiv preprint arXiv:2406.10858

arXiv 2024

[6] [6]

Sijia Chen, Yibo Wang, Yi-Feng Wu, Qing-Guo Chen, Zhao Xu, Weihua Luo, Kaifu Zhang, and Lijun Zhang. 2024 c . Advancing tool-augmented large language models: Integrating insights from errors in inference trees. arXiv preprint arXiv:2406.07115

arXiv 2024

[7] [7]

Sylvain Gelly and David Silver. 2011. Monte-carlo tree search and rapid action value estimation in computer go. Artificial Intelligence, 175(11):1856--1875

work page 2011

[8] [8]

Shibo Hao, Tianyang Liu, Zhen Wang, and Zhiting Hu. 2024. Toolkengpt: Augmenting frozen language models with massive tools via tool embeddings. Advances in neural information processing systems, 36

work page 2024