Pith. sign in

REVIEW 4 major objections 5 minor 34 references

Beyond Passive Critical Thinking: Fostering Proactive Questioning to Enhance Human-AI Collaboration

T0 review · 4 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 Proactive critical thinking can be trained into small language models: reinforcement learning with an answerability reward lifts Qwen3-1.7B from 0.15% to 73.98% accuracy on incomplete math problems.

desk verdict A well-motivated training recipe for proactive questioning whose headline numbers rest on a self-referential user-simulator and need verification. read the letter →

arxiv 2507.23407 v1 pith:V3EUF5BX submitted 2025-07-31 cs.CL

Ante Wang , Yujie Lin , Jingyao Liu , Suhang Wu , Hao Liu , Xinyan Xiao , Jinsong Su This is my paper
classification cs.CL
keywords proactivecriticalthinkingclarificationquestionsreinforcementlearningGRPOanswerabilityrewardGSM8Kmathematicalreasoninghuman-AIcollaboration
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 introduces proactive critical thinking: a model that detects when a question is unanswerable, asks the user a targeted clarifying question, and then solves the problem once the user supplies the missing detail. To measure this ability, the authors build GSM-MC and GSM-MCE from GSM8K by removing a key variable (and injecting distractors in GSM-MCE). They find that off-the-shelf Qwen3 and Llama models, especially small ones, rarely ask for the missing information and instead hallucinate answers. The central result is that reinforcement learning with a heuristic answerability reward—a $+0.5$ bonus for requesting information on unanswerable questions and a $-0.5$ penalty for unnecessary requests on answerable ones—converts this behavior: Qwen3-1.7B's accuracy on GSM-MC rises from 0.15% to 73.98% while its standard GSM8K accuracy is preserved. A sympathetic reader would take this as evidence that proactive questioning is a trainable skill that does not require model scale.

What carries the argument

The mechanism that carries the argument is an answerability-conditioned reward heuristic added to GRPO, an on-policy RL algorithm in which a group of rollouts are scored and advantages are normalized per group. Each rollout is a two-turn interaction: model request, simulated user reply, and final answer. The heuristic gives $+0.5$ for requesting missing information on unanswerable questions and $-0.5$ for making requests on answerable questions, making the sparse correctness signal denser and accelerating convergence. The same answerability information is used as a prompt-level hint during SFT data collection. The benchmarks GSM-MC (1,368 problems) and GSM-MCE (1,134 with injected distractors) are the measuring instruments, and the DeepSeek-V3-based user agent is what turns the interactive setting into a scalable automatic evaluation.

What would settle it

Present the trained model with a user agent or human who does not reply with the missing value verbatim—for example, answering 'I don't know' or providing a related but different quantity—and measure whether the model still reaches the correct answer. If accuracy collapses under such replies, the capability is an artifact of the scripted user agent rather than robust proactive critical thinking.

Watch

Extended reading notes

Core claim

The paper's central claim is that proactive critical thinking—identifying a missing piece of information and asking for it—is a distinct skill that can be instilled through training. The authors remove a key variable from each GSM8K problem, rephrase the question fluently, and filter for quality using DeepSeek-V3, producing GSM-MC and GSM-MCE. On top of standard SFT and on-policy GRPO reinforcement learning, they add a heuristic answerability reward: for unanswerable questions the model receives $+0.5$ for making a clarification request, and for answerable questions it receives $-0.5$ for making an unnecessary request, layered on the sparse $+1/0$ final-answer correctness reward. Trained models learn to identify the missing variable, phrase a request, incorporate the user's reply, and produce the correct final answer, reaching 73.98% ACC on GSM-MC for Qwen3-1.7B with thinking enabled, up from 0.15% for the vanilla model. The paper also reports that RL turns the thinking mode from a liability into an asset, preserves GSM8K accuracy, improves out-of-distribution performance on MIP-MATH, and extends to multi-turn dialogues.

Load-bearing premise

The load-bearing premise is that the scripted DeepSeek-V3 user agent—prompted with the original, unmodified question—always replies with exactly the missing value the model asks for, so the measured accuracy reflects genuine proactive collaboration rather than imitation of that scripted exchange.

Editorial extensions

If this is right

  • Small models (1.7B–3B) can acquire proactive critical thinking through RL, narrowing the gap with much larger models on this capability.
  • The answerability reward accelerates RL convergence compared with sparse outcome rewards, so similar reward shaping could be applied to other interactive tasks that require a question-asking step.
  • RL-trained models retain and sometimes improve standard GSM8K accuracy, suggesting proactive questioning does not trade away basic math competence.
  • The trained skill transfers to an out-of-distribution unanswerable benchmark (MIP-MATH) and to multi-turn conversations, indicating generalization beyond the training distribution.
  • After RL, enabling the thinking mode improves performance, reversing the pattern in vanilla models where thinking mode hurt performance.

Reading between the lines

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

  • If the measured behavior is driven by imitation of the scripted user agent, real human-AI collaboration may require training against more varied and less cooperative user responses; the paper's single-turn scripted agent likely understates that challenge.
  • The answerability-reward heuristic should transfer to other domains, such as clinical triage or tool use, where the cost of an unnecessary clarification differs; a natural test is to apply the same shaping to a medical question-answering dataset.
  • The magnitude of the gain (0.15% to 73.98%) suggests the benchmark is gameable: a model that always asks for more information and then echoes the user agent's reply would score well, so a harder benchmark should penalize requests that a real user cannot answer or that do not narrow down the problem.
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.

Referee Report

4 major / 5 minor

Summary. The paper proposes 'proactive critical thinking' as a paradigm in which an LLM detects missing information in a user query and asks a targeted clarifying question instead of merely refusing or producing a wrong answer. It introduces two benchmarks, GSM-MC and GSM-MCE, derived from GSM8K by removing one key variable and (for GSM-MCE) adding irrelevant details, with a sampling-based filtering pipeline built on DeepSeek-V3. The authors then train small language models with supervised fine-tuning and GRPO-based reinforcement learning, augmented with answerability heuristics as dense rewards and SFT data conditioning. Their central quantitative claim is that Qwen3-1.7B's accuracy on GSM-MC rises from 0.15% (vanilla) to 73.98% (SFT+RL), with consistent gains across model families and thinking modes. They also report a transfer experiment on MIP-MATH and a multi-turn extension.

Significance. If the reported gains hold under human-like interaction, the paper would be a useful step toward making LLMs genuinely collaborative when input is incomplete. The benchmarks are a concrete resource, and the ablations on data filtering and heuristic guidance provide useful evidence about which design choices matter. The MIP-MATH transfer and multi-turn results are encouraging attempts to test generalization. However, the headline metric relies on a simulated user agent that has access to the original full question, which may overstate the model's ability to identify missing information from a real user's perspective. The absence of error bars further weakens the quantitative claims. Overall, the contribution is significant but requires substantial additional validation before the central claim can be accepted.

major comments (4)
  1. [Section 3, Appendix A] The user-agent protocol supplies the original unmodified question and instructs the agent to output only the relevant parts about the requested information, returning 'None' only if the information is absent. Since every GSM-MC item has exactly one removed variable, any clarification request that names the variable role (e.g., 'the price of the first glass') will yield the exact value with no reasoning on the model's part. The reported ACC may therefore measure the model's ability to elicit a value from a permissive extractor rather than its ability to identify the missing premise and ask a human-understandable, sufficiently specific question. I recommend adding a human evaluation, or a user model without access to the original question, on a random subset to validate the metric and the headline 73.98% figure.
  2. [Section 4, Sampling-Based Filtering] The filtering thresholds (removing items where more than 12 of 16 sampled solutions answer directly, removing items where more than 12 of 16 are resolved in the second turn, and discarding items that fail all 16 attempts) appear arbitrary, and no sensitivity analysis is reported. More importantly, the filter is run with DeepSeek-V3 as both the solver and the user agent, and the same model family is later used as the test-time user agent. This creates a closed calibration loop: benchmark items are retained precisely when DeepSeek-V3's simulated interaction can resolve them. The authors should report robustness to the thresholds and should ideally use a different model for filtering than for test-time simulation.
  3. [Table 2 and Section 6.2] All results in Table 2 are from single runs with no standard deviations or multiple seeds. This undermines comparative claims such as the statement that for Qwen3-8B 'RL alone surpasses the performance of two-stage training,' which rests on gaps like 81.73 vs. 79.24 (w/o think) and 85.53 vs. 83.11 (w/ think) on GSM-MC; these are within the range of typical RL run-to-run variation. Please report means and standard deviations over at least three seeds for all core configurations.
  4. [Section 6.3.3 and 6.3.4] The MIP-MATH generalization experiment and the multi-turn expansion experiment use the same DeepSeek-V3 user-agent protocol in which the user agent has the original unmodified question in context. They therefore do not break the circularity identified above; the multi-turn improvement may simply reflect more chances to elicit a value from the permissive extractor. A transfer evaluation with a different user model, or with human annotators, is needed to support the claim that the learned skill generalizes to realistic human-AI collaboration.
minor comments (5)
  1. [References] In the reference to Rahman et al. (2024), 'logical interity' should be 'logical integrity'.
  2. [Equation (2)] The symbol q is overloaded: earlier in the paper q denotes the clarification question, but in the RL objective the policy input is the original user question. Please use separate symbols to avoid confusion.
  3. [Table 2] The caption should define 'w/ think' and 'w/o think' (thinking mode enabled/disabled) and state that ACC for GSM-MC and GSM-MCE is measured on the second-turn response, as described in Section 6.1.1.
  4. [Section 4] The relationship between GSM-MC (1,368 items) and GSM-MCE (1,134 items) is not explained; after irrelevant-information injection, some items presumably fail the second filtering stage. Please state the pass rate and criteria that reduce the size.
  5. [Overall availability] The paper does not mention a release plan or URL for the GSM-MC and GSM-MCE datasets or the training code. Since the benchmarks are a central contribution, a release statement should be included.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the trained-model gains are empirical benchmark results, and the DeepSeek-V3 user-agent coupling is an external-validity caveat, not a circular derivation.

full rationale

Walking the paper's derivation chain, I find no step where a prediction or first-principles result reduces to its inputs by construction. The central claim, improving Qwen3-1.7B accuracy on GSM-MC from 0.15% to 73.98%, is an empirical outcome measured on a held-out benchmark after training; it is not a fitted parameter renamed as a prediction, nor is the benchmark score defined in terms of the model's own outputs. The benchmark construction pipeline uses DeepSeek-V3 for variable removal, filtering, and user-agent simulation, which makes the task easier and raises legitimate external-validity concerns about transfer to real human users, but it does not make the trained model's accuracy equivalent to the construction inputs. The user-agent prompt (Appendix A) provides the unmodified original question so that the simulated user can answer clarification requests, and the filtering criterion retains cases resolvable by DeepSeek-V3; however, vanilla Qwen3-1.7B still scores 0.15% on this same setup, showing the metric is not forced by construction. The out-of-distribution MIP-MATH evaluation is an external benchmark from other authors and partially breaks the dependency. Self-citations in the paper (e.g., Lin et al. 2025, Wang et al. 2025a/b, Shao et al. 2025) appear only as related-work context and are not load-bearing for the central claim. Under the specified standards, this is a non-circular empirical study with a benchmark-validity caveat rather than a circular derivation.

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

The central claim rests on several domain assumptions about the user agent and the filtering pipeline, plus hand-chosen reward weights and thresholds. No new physical or conceptual entities are introduced beyond the 'proactive critical thinking' task framing.

free parameters (4)
  • Heuristic reward weights = +0.5 for requests on unanswerable, -0.5 for unnecessary requests, 0 otherwise
    Hand-chosen in Section 5 to shape the RL reward; they strongly influence the trade-off between asking and answering.
  • Filtering thresholds = 12 out of 16 cases
    In Section 4, questions are removed if answered first-turn in >12/16 samples, kept if resolved second-turn in >12/16, and discarded if unresolved in all 16. These thresholds are arbitrary and shape the benchmark composition.
  • Number of samples for filtering = 16
    Section 4 samples 16 solutions per question for filtering; this affects the precision of the quality estimates.
  • Training hyperparameters = SFT lr=5e-6, RL lr=1e-6, rollouts=8, 1 epoch, warmup 10%
    Section 6.1.2; these are standard but not tuned systematically, and results may vary with different choices.
assumptions (4)
  • domain assumption The DeepSeek-V3 user agent correctly extracts the exact missing value from the original question when the model asks for it, and never infers or calculates the value.
    Appendix A prompt explicitly instructs the user agent to output only relevant parts and never infer. This assumption is critical for the ACC metric.
  • domain assumption The filtering pipeline using DeepSeek-V3 yields a set of questions that are truly unanswerable first-turn and resolvable in one clarification request.
    Section 4 describes the filtering rules; the benchmark quality depends entirely on DeepSeek-V3's judgments.
  • standard math The original GSM8K answers are correct ground truth.
    ACC on benchmarks is computed by comparing to the original answer of the unmodified GSM8K question.
  • domain assumption Second-turn response accuracy, after the user agent provides the missing detail, measures proactive critical thinking quality.
    The paper defines ACC this way in Section 6.1.1; it assumes that a correct final answer implies the model successfully asked for what was needed.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Beyond Passive Critical Thinking: Fostering Proactive Questioning to Enhance Human-AI Collaboration." pith.science (2026). https://pith.science/paper/V3EUF5BX

@misc{pith2026250723407,
  author       = {Pith},
  title        = {Pith review of: Beyond Passive Critical Thinking: Fostering Proactive Questioning to Enhance Human-AI Collaboration},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/V3EUF5BX}},
  note         = {Machine review of arXiv:2507.23407}
}
read the original abstract

Critical thinking is essential for building robust AI systems, preventing them from blindly accepting flawed data or biased reasoning. However, prior work has primarily focused on passive critical thinking, where models simply reject problematic queries without taking constructive steps to address user requests. In this work, we introduce proactive critical thinking, a paradigm where models actively seek missing or clarifying information from users to resolve their queries better. To evaluate this capability, we present GSM-MC and GSM-MCE, two novel benchmarks based on GSM8K for assessing mathematical reasoning under incomplete or misleading conditions. GSM-MC contains 1,368 math problems with a key variable deliberately removed, requiring models to identify and request the missing information. GSM-MCE further increases the difficulty by introducing irrelevant details to test robustness against distractions. Experiments on Qwen3 and Llama series models show that, while these models excel in traditional reasoning tasks due to extensive post-training and inference-time scaling, they struggle with proactive critical thinking, especially smaller ones. However, we demonstrate that reinforcement learning (RL) can significantly improve this ability. Using our enhanced RL algorithm, we achieve substantial gains, boosting the Qwen3-1.7B's accuracy from 0.15% to 73.98% on GSM-MC. We hope this work advances models that collaborate more effectively with users in problem-solving through proactive critical thinking.

Figures

Figures reproduced from arXiv: 2507.23407 by the authors.

Figure 1
Figure 1. When faced with flawed questions, Left: Vanilla reasoning models fall into repetitive speculation and eventual hallucination. Despite identifying the missing conditions during the thinking process, they still attempt to force an answer, ultimately providing an incorrect solution. Middle: Models with passive critical thinking are more cautious, refusing to answer upon identifying flaws in the question. However, this … view at source ↗
Figure 2
Figure 2. Reward curves in the RL stage following SFT, [PITH_FULL_IMAGE:figures/full_fig_p008_2.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

34 extracted references · 15 canonical work pages

  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]

    Azza Alkaabi and Deena Elsori. 2025. Navigating digital frontiers in uae healthcare: A qualitative exploration of healthcare professionals’ and patients’ experiences with ai and telemedicine. PLOS Digital Health, 4(4):e0000586

    work page 2025

  4. [4]

    Chinmaya Andukuri, Jan-Philipp Fr \"a nken, Tobias Gerstenberg, and Noah D Goodman. 2024. Star-gate: Teaching language models to ask clarifying questions. arXiv preprint arXiv:2403.19154

    arXiv 2024

  5. [5]

    Karl Cobbe, Vineet Kosaraju, Mohammad Bavarian, Mark Chen, Heewoo Jun, Lukasz Kaiser, Matthias Plappert, Jerry Tworek, Jacob Hilton, Reiichiro Nakano, Christopher Hesse, and John Schulman. 2021. Training verifiers to solve math word problems. arXiv preprint arXiv:2110.14168

    arXiv 2021

  6. [6]

    Yihong Dong, Xue Jiang, Zhi Jin, and Ge Li. 2024. Self-collaboration code generation via chatgpt. ACM Transactions on Software Engineering and Methodology, 33(7):1--38

    work page 2024

  7. [7]

    Chenrui Fan, Ming Li, Lichao Sun, and Tianyi Zhou. 2025. Missing premise exacerbates overthinking: Are reasoning models losing critical thinking skill? arXiv preprint arXiv:2504.06514

    arXiv 2025

  8. [8]

    Yunhe Feng, Sreecharan Vanam, Manasa Cherukupally, Weijian Zheng, Meikang Qiu, and Haihua Chen. 2023. Investigating code generation performance of chatgpt with crowdsourcing social data. In 2023 IEEE 47th Annual Computers, Software, and Applications Conference (COMPSAC), pages 876--885. IEEE

    2023

Show all 34 references
  1. [9]

    Dan Hendrycks, Collin Burns, Saurav Kadavath, Akul Arora, Steven Basart, Eric Tang, Dawn Song, and Jacob Steinhardt. 2021. https://datasets-benchmarks-proceedings.neurips.cc/paper_files/paper/2021/file/be83ab3ecd0db773eb2dc1b0a17836a1-Paper-round2.pdf Measuring mathematical pr...

    2021

  2. [10]

    Shima IMANI, Harsh Shrivastava, and Liang Du. 2024. Mathematical reasoning using large language models. US Patent App. 18/144,802

    2024

  3. [11]

    Yoonsu Kim, Jueon Lee, Seoyoung Kim, Jaehyuk Park, and Juho Kim. 2024. Understanding users’ dissatisfaction with chatgpt responses: Types, resolving tactics, and the effect of knowledge level. In Proceedings of the 29th International Conference on Intelligent User Interfaces, ...

    2024

  4. [12]

    Polina Kirichenko, Mark Ibrahim, Kamalika Chaudhuri, and Samuel J Bell. 2025. Abstentionbench: Reasoning llms fail on unanswerable questions. arXiv preprint arXiv:2506.09038

    2025 arXiv

  5. [13]

    Lorenz Kuhn, Yarin Gal, and Sebastian Farquhar. 2022. Clam: Selective clarification for ambiguous questions with generative language models. arXiv preprint arXiv:2212.07769

    2022 arXiv

  6. [14]

    Qintong Li, Leyang Cui, Xueliang Zhao, Lingpeng Kong, and Wei Bi. 2024. Gsm-plus: A comprehensive benchmark for evaluating the robustness of llms as mathematical problem solvers. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume...

    2024

  7. [15]

    Shuyue Stella Li, Jimin Mun, Faeze Brahman, Jonathan Ilgen, Yulia Tsvetkov, and Maarten Sap. 2025. https://doi.org/10.48550/arXiv.2502.14860 Aligning llms to ask good questions a case study in clinical reasoning . CoRR, abs/2502.14860

    2025 doi arXiv

  8. [16]

    Yujie Lin, Ante Wang, Moye Chen, Jingyao Liu, Hao Liu, Jinsong Su, and Xinyan Xiao. 2025. https://aclanthology.org/2025.findings-acl.808/ Investigating inference-time scaling for chain of multi-modal thought: A preliminary study . In Findings of the Association for Computation...

    2025

  9. [17]

    Aixin Liu, Bei Feng, Bing Xue, Bingxuan Wang, Bochao Wu, Chengda Lu, Chenggang Zhao, Chengqi Deng, Chenyu Zhang, Chong Ruan, et al. 2024. Deepseek-v3 technical report. arXiv preprint arXiv:2412.19437

    2024 arXiv

  10. [18]

    Jingyuan Ma, Damai Dai, Zihang Yuan, Weilin Luo, Bin Wang, Qun Liu, Lei Sha, Zhifang Sui, et al. 2024. Large language models struggle with unreasonability in math problems. arXiv preprint arXiv:2403.19346

    2024 arXiv

  11. [19]

    AI Meta. 2024. Llama 3.2: Revolutionizing edge ai and vision with open, customizable models. Meta AI Blog. Retrieved December, 20:2024

    2024

  12. [20]

    AM Rahman, Junyi Ye, Wei Yao, Sierra S Liu, Jesse Yu, Jonathan Yu, Wenpeng Yin, and Guiling Wang. 2024. From blind solvers to logical thinkers: Benchmarking llms' logical interity on faulty mathematical problems. arXiv preprint arXiv:2410.18921

    2024 arXiv

  13. [21]

    Liangying Shao, Yanfu Yan, Denys Poshyvanyk, and Jinsong Su. 2025. https://doi.org/10.1109/ICSE-NIER66352.2025.00020 Unigencoder: Merging seq2seq and seq2tree paradigms for unified code generation . In 2025 IEEE/ACM 47th International Conference on Software Engineering: New Id...

    2025

  14. [22]

    Zhihong Shao, Peiyi Wang, Qihao Zhu, Runxin Xu, Junxiao Song, Xiao Bi, Haowei Zhang, Mingchuan Zhang, YK Li, Yang Wu, et al. 2024. Deepseekmath: Pushing the limits of mathematical reasoning in open language models. arXiv preprint arXiv:2402.03300

    2024 arXiv

  15. [23]

    Linxin Song, Taiwei Shi, and Jieyu Zhao. 2025. The hallucination tax of reinforcement finetuning. arXiv preprint arXiv:2505.13988

    2025 arXiv

  16. [24]

    YuHong Sun, Zhangyue Yin, Qipeng Guo, Jiawen Wu, Xipeng Qiu, and Hui Zhao. 2024. Benchmarking hallucination in large language models based on unanswerable math word problem. In Proceedings of the 2024 Joint International Conference on Computational Linguistics, Language Resour...

    2024

  17. [25]

    Ante Wang, Linfeng Song, Ye Tian, Baolin Peng, Dian Yu, Haitao Mi, Jinsong Su, and Dong Yu. 2025 a . Litesearch: Efficient tree search with dynamic exploration budget for math reasoning. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 39, pages 25318--25326

    2025

  18. [26]

    Wenxuan Wang, Juluan Shi, Zixuan Ling, Yuk-Kit Chan, Chaozheng Wang, Cheryl Lee, Youliang Yuan, Jen-tse Huang, Wenxiang Jiao, and Michael R Lyu. 2024. Learning to ask: When llm agents meet unclear instruction. arXiv preprint arXiv:2409.00557

    2024 arXiv

  19. [27]

    Yaoxiang Wang, Haoling Li, Xin Zhang, Jie Wu, Xiao Liu, Wenxiang Hu, Zhongxin Guo, Yangyu Huang, Ying Xin, Yujiu Yang, Jinsong Su, Qi Chen, and Scarlett Li. 2025 b . https://openreview.net/forum?id=RAxe7nF4Oz Epicoder: Encompassing diversity and complexity in code generation ....

    2025

  20. [28]

    Shirley Wu, Michel Galley, Baolin Peng, Hao Cheng, Gavin Li, Yao Dou, Weixin Cai, James Zou, Jure Leskovec, and Jianfeng Gao. 2025. https://openreview.net/forum?id=DmH4HHVb3y Collab LLM : From passive responders to active collaborators . In Forty-second International Conferenc...

    2025

  21. [29]

    Ryutaro Yamauchi, Sho Sonoda, Akiyoshi Sannai, and Wataru Kumagai. 2023. Lpml: llm-prompting markup language for mathematical reasoning. arXiv preprint arXiv:2309.13078

    2023 arXiv

  22. [30]

    An Yang, Anfeng Li, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Gao, Chengen Huang, Chenxu Lv, et al. 2025. Qwen3 technical report. arXiv preprint arXiv:2505.09388

    2025 arXiv

  23. [31]

    Shunyu Yao, Dian Yu, Jeffrey Zhao, Izhak Shafran, Tom Griffiths, Yuan Cao, and Karthik Narasimhan. 2023. Tree of thoughts: Deliberate problem solving with large language models. Advances in neural information processing systems, 36:11809--11822

    2023

  24. [32]

    J Diego Zamfirescu-Pereira, Richmond Y Wong, Bjoern Hartmann, and Qian Yang. 2023. Why johnny can’t prompt: how non-ai experts try (and fail) to design llm prompts. In Proceedings of the 2023 CHI conference on human factors in computing systems, pages 1--21

    2023

  25. [33]

    Bradley Knox, and Eunsol Choi

    Michael JQ Zhang, W. Bradley Knox, and Eunsol Choi. 2025. https://openreview.net/forum?id=cwuSAR7EKd Modeling future conversation turns to teach LLM s to ask clarifying questions . In The Thirteenth International Conference on Learning Representations

    2025

  26. [34]

    Zhehua Zhou, Jiayang Song, Kunpeng Yao, Zhan Shu, and Lei Ma. 2024. Isr-llm: Iterative self-refined large language model for long-horizon sequential task planning. In 2024 IEEE International Conference on Robotics and Automation (ICRA), pages 2081--2088. IEEE

    2024

Pith tools

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