Pith. sign in

REVIEW 4 major objections 6 minor 1 cited by

Teaching Language Models To Gather Information Proactively

T0 review · 4 major / 6 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 Asking the right questions lifts a 7B model past o3-mini

desk verdict Solid training idea, but the empirical case is undermined by a missing train/test split and an abstract that credits the model for outlines actually written by GPT-4o. read the letter →

arxiv 2507.21389 v1 pith:BLHNVTAC submitted 2025-07-28 cs.AI cs.CL

Tenghao Huang , Sihao Chen , Muhao Chen , Jonathan May , Longqi Yang , Mengting Wan , Pei Zhou This is my paper
classification cs.AIcs.CL
keywords proactiveinformationgatheringclarificationquestionsreinforcementlearningproximalpolicyoptimizationsyntheticconversationengineimplicitDOLOMITEShumanevaluation
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 tries to establish that large language models can be trained to be proactive information gatherers: instead of passively answering an under-specified prompt, the model should detect what is missing and ask the user targeted questions before producing a final response. To study this, the authors build a scalable synthetic pipeline from existing domain-specific writing tasks, masking the procedural and output-format information that a real user would often leave unstated. Their core claim is that reinforcement fine-tuning with a reward for questions that elicit genuinely new hidden information turns a Qwen-2.5-7B model into a better question-asker and writer, improving automatic task-completion scores by 18% over o3-mini and winning human preference by 42% for questions and 28% for outlines. If true, the result matters because it offers a practical route to making smaller, open-weight models behave like collaborative thought partners in ambiguous real-world work, without expensive human dialogue data.

What carries the argument

The load-bearing mechanism is the evidence-sentence reward: the hidden implicit information is split into sentences, a question is answered by an LLM oracle that cites the sentences it would use, and the immediate reward is positive exactly when any cited sentence belongs to the masked fields. This reward is what makes dense supervision unnecessary: it directly incentivizes questions that uncover previously unspoken domain expertise and fine-grained requirements, and it is applied in a PPO actor-critic loop over synthetic multi-turn conversations built from the DOLOMITES task templates.

What would settle it

Run the trained model with real human users who hold the masked information, compare their answers and outline preferences against the oracle-based reward, and check whether questions that receive reward 0 are judged unhelpful or whether rewarded questions fail to transfer. A low correlation between human helpfulness ratings and the evidence-sentence reward would refute the claim that this reward captures genuinely useful proactive clarification.

Watch

Extended reading notes

Core claim

The central discovery is that a binary, outcome-based reward — one that pays the model only when its question makes the simulated user reveal a sentence drawn from the hidden task information — is a sufficient training signal to teach proactive clarification. The authors formalize the setting as a partially observable dialogue in which the assistant sees only explicit information and must ask up to five questions; a user oracle answers from the masked implicit information, and the reward is 1 if the cited evidence overlaps the hidden fields and 0 otherwise. Using PPO on this reward, the trained Qwen-2.5-7B model outperforms vanilla in-context question-asking by stronger models, outperforms supervised fine-tuning on emulated conversations, and produces questions that target the procedural and output-requirement parts of the task rather than surface details already present in the prompt. Human annotators preferred the trained model's clarification questions and resulting outlines over o3-mini's.

Load-bearing premise

The load-bearing premise is that the simulated LLM user oracle answers questions by faithfully revealing exactly the hidden task information, so that the binary evidence-sentence reward measures what a real user would find useful; if real users answer differently or gain little from the rewarded questions, the learned policy may not transfer.

Editorial extensions

If this is right

  • A 7B model trained this way can ask clarification questions that a stronger, larger reasoning model does not ask, and those questions lead to outlines that human readers prefer.
  • Supervised fine-tuning on synthetic clarification dialogues alone is not enough; the reward signal, not just imitation of question-asking behavior, drives the gains.
  • The method generalizes across open-ended domains: the largest improvements over a direct-answer baseline appear in social science and humanities tasks, where hidden procedural knowledge and output requirements matter most.
  • The evidence-sentence reward creates a training signal for open-ended writing tasks where step-level rewards and outcome verification do not exist, extending reinforcement learning beyond math and code.
  • Because the user oracle is an LLM and the reward is computed automatically, the whole pipeline can be scaled without collecting human clarification data.

Reading between the lines

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

  • Our inference: the evidence-sentence reward could be sharpened to weight which hidden sentences matter, since a question that reveals a trivial fact and one that reveals a decisive requirement currently receive the same reward; a graded reward might further improve question quality.
  • Our inference: the same synthetic masking-plus-reward recipe could transfer to other collaborative settings, such as coding with hidden test requirements or medical note-taking with hidden patient preferences, whenever the implicit information can be decomposed into citable sentences.
  • Our inference: a direct testable extension would be to measure whether the learned questioning policy transfers to human users who have not seen the oracle prompt, and whether human-answer quality correlates with the reward assigned during training.
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 / 6 minor

Summary. The paper introduces a task paradigm called proactive information gathering, in which an LLM must ask clarifying questions to elicit implicit information from a user before producing a long-form writing output. The authors adapt the DOLOMITES dataset by masking procedure and output-specification fields, simulate a user oracle with another LLM, and train a Qwen-2.5-7B policy with PPO using a reward that fires when the oracle's answer cites a hidden sentence. Experiments report an 18% improvement over o3-mini on an LLM-judged coverage metric, domain-wise gains in Table 1, and human preferences for RFT-Qwen's questions (62% win) and outlines (50% win) over o3-mini. The main claim is that reinforcement finetuning with an evidence-sentence reward teaches a small model to ask strategically useful questions and thereby improves downstream writing quality.

Significance. If the reported results are valid and reproducible, the paper would make a useful contribution: it proposes a tractable reward signal for proactive clarification in open-ended writing tasks, provides a synthetic training pipeline that does not require human dialogue logs, and shows across both automatic and human evaluation that a 7B model can surpass a stronger reasoning model when given the right questioning policy. The task formulation, the masking scheme grounded in DOLOMITES, and the evidence-sentence reward are all concrete and externally anchored rather than purely self-referential. However, the experimental validation currently lacks a disclosed train/test split, conflates the model that generates questions with the model that writes the final outline, and relies on an LLM judge and an LLM simulated user whose fidelity to real users is not established. These issues are load-bearing for the abstract's central claim that the trained model generalizes to unseen tasks and outperforms o3-mini.

major comments (4)
  1. [§3.2, §5.1, §5.3] The paper never reports a train/test split. Section 3.2 states that DOLOMITES contains only 519 task templates, Section 5.1 describes PPO training on this data, and Section 5.3 reports evaluation on the same benchmark, but no dataset statistics are given: no number of training episodes, no number of evaluation instances, and no template-level split. With only 519 templates, if the reported scores are computed on instances used in training, the claimed 18% improvement over o3-mini and the +0.37/+0.31 domain gains in Table 1 could reflect memorization rather than learned proactive clarification. This is the central claim of the paper, so the authors must disclose exactly how training and evaluation instances are separated, ideally at the template level, and report performance on held-out templates or domains.
  2. [§5.1, Abstract, Table 4] Section 5.1 states 'We fix vanilla GPT-4o as the writer at draft phase throughout our experiments,' yet the abstract and Section 5.4 attribute the human-evaluated outlines to 'our model' and Table 4 is described as comparing 'model-generated task outlines.' Because GPT-4o is the writer for all methods, the 28% outline preference and the 50% win in Table 4 measure GPT-4o's outline generation conditioned on RFT-Qwen's questions, not outlines written by RFT-Qwen. This conflation obscures what the human evaluation actually demonstrates; the paper should either train and evaluate the same model as the writer or rephrase the claims to state precisely that the questions come from RFT-Qwen and the outlines are generated by a fixed GPT-4o writer.
  3. [§4.1, §4.2, Appendix B] The training reward and the human evaluation both rely on the simulated user oracle responding to questions by citing hidden sentences. The reward in §4.2 is a binary indicator of whether any hidden sentence is cited, which does not distinguish between trivially relevant and strategically important hidden information. More fundamentally, there is no validation that the oracle's citation behavior approximates how real users would answer clarification questions, nor that the binary reward correlates with the quality of the final outline. The human evaluation in Appendix B is static side-by-side comparisons of questions and outlines, not interactive user trials, so it does not directly validate the simulated-dialogue premise. Adding a small human-human clarification study, or at least an analysis of whether reward-improving questions lead to better final outlines in a human interaction setting, would substantially strengthen the external-validity claim.
  4. [§3.3] The automatic evaluation metric is computed by an LLM judge, but the paper reports no validation of this judge against human judgments or against a known reference. The headline 18% improvement over o3-mini is based entirely on this automatic score, and the judge prompt in Fig. 8 is a binary checklist match without any reported inter-annotator agreement, calibration, or error analysis. Without evidence that the judge's scores correlate with human preferences, the automatic metric alone cannot support the claim that RFT-Qwen significantly outperforms o3-mini.
minor comments (6)
  1. [Abstract] The phrase 'a reinforcement finetuning strategy rewards questions' is missing a 'that' before 'rewards'; it should read 'a reinforcement finetuning strategy that rewards questions.'
  2. [§4.1] The word 'unavailble' should be 'unavailable,' and 'needed-to' in §3.2 should be 'needed to.'
  3. [§5.4] The heading 'Demonstraed' should be 'Demonstrated.'
  4. [Appendix A] The text says 'Figure 4 illustrates the learning dynamics of our PPO critic,' but Figure 4 is already used for the main writing-task results; the appendix figure should be numbered separately.
  5. [Appendix B] The human evaluation used only three annotators and 30 outputs, and no inter-annotator agreement is reported; please add agreement statistics and clarify whether the annotators were blind to model identity.
  6. [§5.2] The GPT-4o Indirectly Supervised baseline is described as trained 'in a DPO fashion' on QMSUM, but no hyperparameters, filtering details, or number of preference pairs are given; a brief description would improve reproducibility.

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity: the reward and evaluation are anchored to external DOLOMITES data, and no fitted parameter is renamed as a prediction.

full rationale

The paper's derivation chain is: DOLOMITES supplies external task instances with hidden procedure and output specification; a simulated user oracle answers clarification questions using that hidden information; the evidence-sentence reward is a binary signal based on whether the oracle's cited sentences belong to the hidden set; PPO trains the policy on this signal; final drafts are written by a fixed GPT-4o using the accumulated dialogue; evaluation uses a frozen LLM judge checking final outlines against the output-spec checklist; human annotators compare RFT-Qwen questions and outlines against those of o3-mini. None of these steps reduces to its inputs by construction: the reward is a training proxy, not identical to the judge's checklist score; no fitted parameter is reused as a prediction; and the DOLOMITES ground truth is external to the trained policy. The paper does contain self-citations (e.g., Spangher et al., 2025; Zhou et al., 2023), but they are contextual and not load-bearing for the central claim. The main threats—undisclosed train/test split, the oracle's complete cooperativeness, and the binary reward's inability to distinguish trivial from strategically important hidden information—are validity and generalization limitations, not circularity. Therefore no specific circular step is identified; the score reflects minor methodological concerns rather than a by-construction reduction.

Assumptions & free parameters 1 free parameters · 4 assumptions · 0 invented entities

The central claim rests on four domain assumptions about the benchmark, the simulated oracle, the reward signal, and the judge. None of these are derived from first principles, and the paper does not provide independent evidence for them outside the DOLOMITES setting.

free parameters (1)
  • dialogue turn budget = 5
    Maximum number of clarification questions per episode; chosen by the authors as a practical limit, not fitted to data.
assumptions (4)
  • domain assumption DOLOMITES task templates and expert-curated checklists are representative of real-world professional writing tasks.
    All training and evaluation rest on this benchmark; the paper defends the choice in the Limitations section but does not validate against other benchmarks.
  • domain assumption The simulated user oracle faithfully answers clarification questions using only the hidden implicit information.
    The reward signal is derived from the oracle's cited evidence sentences; unrealistically cooperative behavior would inflate measured usefulness. See Section 4.1.
  • domain assumption The binary evidence-sentence reward provides a sufficient learning signal for good proactive clarification.
    The reward is 1 if any hidden sentence is cited, which does not distinguish between more and less strategically important information. See Section 4.2.
  • domain assumption The LLM judge's binary match against checklist items is a valid proxy for writeup quality.
    Automatic scores rely on this judge without a reported validation against human judgments. See Section 3.3.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Teaching Language Models To Gather Information Proactively." pith.science (2026). https://pith.science/paper/BLHNVTAC

@misc{pith2026250721389,
  author       = {Pith},
  title        = {Pith review of: Teaching Language Models To Gather Information Proactively},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/BLHNVTAC}},
  note         = {Machine review of arXiv:2507.21389}
}
read the original abstract

Large language models (LLMs) are increasingly expected to function as collaborative partners, engaging in back-and-forth dialogue to solve complex, ambiguous problems. However, current LLMs often falter in real-world settings, defaulting to passive responses or narrow clarifications when faced with incomplete or under-specified prompts, falling short of proactively gathering the missing information that is crucial for high-quality solutions. In this work, we introduce a new task paradigm: proactive information gathering, where LLMs must identify gaps in the provided context and strategically elicit implicit user knowledge through targeted questions. To systematically study and train this capability, we design a scalable framework that generates partially specified, real-world tasks, masking key information and simulating authentic ambiguity. Within this setup, our core innovation is a reinforcement finetuning strategy that rewards questions that elicit genuinely new, implicit user information -- such as hidden domain expertise or fine-grained requirements -- that would otherwise remain unspoken. Experiments demonstrate that our trained Qwen-2.5-7B model significantly outperforms o3-mini by 18% on automatic evaluation metrics. More importantly, human evaluation reveals that clarification questions and final outlines generated by our model are favored by human annotators by 42% and 28% respectively. Together, these results highlight the value of proactive clarification in elevating LLMs from passive text generators to genuinely collaborative thought partners.

Figures

Figures reproduced from arXiv: 2507.21389 by the authors.

Figure 1
Figure 1. Distribution of LLM Interaction Styles Across [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Proactive Clarification Enables Optimal LLM [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. An example of our task input. Contents marked by red boxes are not visible to LLMs. Only contents marked by green boxes are fed to LLMs as input. everyday conversation (Chen et al., 2024b) within fully unobservable environments and reasoning tasks (Wu et al., 2025). In contrast, our work targets open-ended writing tasks and employs a partially observable environment, reflecting real-world sce￾narios where users inpu… view at source ↗
Figures from the paper (5 more)
Figure 4
Figure 4. Figure 4: Writing Task Performance Across LLM Variants. Bar chart comparing average writing task scores [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]
Figure 6
Figure 6. Figure 6: Writing task scores across clarification ques [PITH_FULL_IMAGE:figures/full_fig_p007_6.png]
Figure 8
Figure 8. Figure 8: LLM-as-Judge style prompt for response eval [PITH_FULL_IMAGE:figures/full_fig_p010_8.png]
Figure 7
Figure 7. Figure 7: Critic return rewards average per step. Validation Reward [PITH_FULL_IMAGE:figures/full_fig_p010_7.png]
Figure 9
Figure 9. Figure 9: Prompt details for proactive clarification question generation. [PITH_FULL_IMAGE:figures/full_fig_p012_9.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

  1. Clarify Before Executing: A Self-Evolving Agent for Resolving Intent Asymmetry in 3D Tool Orchestration

    cs.CV 2026-07 conditional novelty 6.0 of 10

    A clarification-first 3D agent, trained by simulated multi-turn dialogue, reaches 60.4% and 43.3% success on single- and multi-step 3D tool tasks, more than doubling prior baselines.

Reference graph

Works this paper leans on

26 extracted references · 8 canonical work pages · cited by 1 Pith paper

  1. [1]

    Bruce Croft

    Keping Bi, Qingyao Ai, and W. Bruce Croft. 2021. https://doi.org/10.1145/3471158.3472232 Asking clarifying questions based on negative feedback in conversational search . In Proceedings of the 2021 ACM SIGIR International Conference on Theory of Information Retrieval, ICTIR '21, page 157–166, New York, NY, USA. Association for Computing Machinery

    arXiv 2021

  2. [2]

    Hudson, Ehsan Adeli, Russ Altman, Simran Arora, Sydney von Arx, Michael S

    Rishi Bommasani, Drew A. Hudson, Ehsan Adeli, Russ Altman, Simran Arora, Sydney von Arx, Michael S. Bernstein, Jeannette Bohg, Antoine Bosselut, Emma Brunskill, Erik Brynjolfsson, Shyamal Buch, Dallas Card, Rodrigo Castellon, Niladri Chatterji, Annie Chen, Kathleen Creel, Jared Quincy Davis, Dora Demszky, and 95 others. 2022. https://arxiv.org/abs/2108.07...

    arXiv 2022

  3. [3]

    Le, Christopher Ré, and Azalia Mirhoseini

    Bradley Brown, Jordan Juravsky, Ryan Ehrlich, Ronald Clark, Quoc V. Le, Christopher Ré, and Azalia Mirhoseini. 2024. https://doi.org/10.48550/arXiv.2407.21787 Large language monkeys: Scaling inference compute with repeated sampling

    arXiv doi:10.48550/arxiv.2407.21787 2024

  4. [4]

    Guoxin Chen, Minpeng Liao, Chengxi Li, and Kai Fan. 2024 a . https://arxiv.org/abs/2405.03553 Alphamath almost zero: Process supervision without process . Preprint, arXiv:2405.03553

    arXiv 2024

  5. [5]

    Maximillian Chen, Ruoxi Sun, Sercan \"O Ar k, and Tomas Pfister. 2024 b . Learning to clarify: Multi-turn conversations with action-based contrastive self-training. arXiv preprint arXiv:2406.00222

    arXiv 2024

  6. [6]

    Yang Deng, Wenqiang Lei, Wenxuan Zhang, Wai Lam, and Tat-Seng Chua. 2022. Pacific: towards proactive conversational question answering over tabular and textual data in finance. arXiv preprint arXiv:2210.08817

    arXiv 2022

  7. [7]

    Meiqi Guo, Mingda Zhang, Siva Reddy, and Malihe Alikhani. 2021. Abg-coqa: Clarifying ambiguity in conversational question answering. In 3rd Conference on Automated Knowledge Base Construction

    work page 2021

  8. [8]

    Kunal Handa, Drew Bent, Alex Tamkin, Miles McCain, Esin Durmus, Michael Stern, Mike Schiraldi, Saffron Huang, Stuart Ritchie, Steven Syverud, Kamya Jagadish, Margaret Vo, Matt Bell, and Deep Ganguli. 2025. https://www.anthropic.com/news/anthropic-education-report-how-university-students-use-claude Anthropic education report: How university students use claude

    work page 2025

Show all 26 references
  1. [9]

    Philipp Koehn and Christof Monz. 2006. Manual and automatic evaluation of machine translation between european languages. In Proceedings of the workshop on statistical machine translation, pages 102--121. Association for Computational Linguistics

    2006

  2. [10]

    Hunter Lightman, Vineet Kosaraju, Yura Burda, Harri Edwards, Bowen Baker, Teddy Lee, Jan Leike, John Schulman, Ilya Sutskever, and Karl Cobbe. 2023. https://arxiv.org/abs/2305.20050 Let's verify step by step . Preprint, arXiv:2305.20050

    2023 arXiv

  3. [11]

    Yaxi Lu, Shenzhi Yang, Cheng Qian, Guirong Chen, Qinyu Luo, Yesai Wu, Huadong Wang, Xin Cong, Zhong Zhang, Yankai Lin, and 1 others. 2024. Proactive agent: Shifting llm agents from reactive responses to active assistance. arXiv preprint arXiv:2410.12361

    2024 arXiv

  4. [12]

    Liangchen Luo, Yinxiao Liu, Rosanne Liu, Samrat Phatale, Meiqi Guo, Harsh Lara, Yunxuan Li, Lei Shu, Yun Zhu, Lei Meng, Jiao Sun, and Abhinav Rastogi. 2024. https://arxiv.org/abs/2406.06592 Improve mathematical reasoning in language models by automated process supervision . Pr...

    2024 arXiv

  5. [13]

    Chaitanya Malaviya, Priyanka Agrawal, Kuzman Ganchev, Pranesh Srinivasan, Fantine Huot, Jonathan Berant, Mark Yatskar, Dipanjan Das, Mirella Lapata, and Chris Alberti. 2024. https://arxiv.org/abs/2405.05938 Dolomites: Domain-specific long-form methodical tasks . Preprint, arXi...

    2024 arXiv

  6. [14]

    Jing-Cheng Pang, Heng-Bo Fan, Pengyuan Wang, Jia-Hao Xiao, Nan Tang, Si-Hang Yang, Chengxing Jia, Sheng-Jun Huang, and Yang Yu. 2024. Empowering language models with active inquiry for deeper understanding. arXiv preprint arXiv:2402.03719

    2024 arXiv

  7. [15]

    Manning, and Chelsea Finn

    Rafael Rafailov, Archit Sharma, Eric Mitchell, Stefano Ermon, Christopher D. Manning, and Chelsea Finn. 2024. https://arxiv.org/abs/2305.18290 Direct preference optimization: Your language model is secretly a reward model . Preprint, arXiv:2305.18290

    2024 arXiv

  8. [16]

    Xuhui Ren, Hongzhi Yin, Tong Chen, Hao Wang, Zi Huang, and Kai Zheng. 2021. https://doi.org/10.1145/3404835.3462839 Learning to ask appropriate questions in conversational recommendation . In Proceedings of the 44th International ACM SIGIR Conference on Research and Developmen...

    2021

  9. [17]

    Nicholas Roth, Christopher Hidey, Lucas Spangher, William F Arnold, Chang Ye, Nick Masiewicki, Jinoo Baek, Peter Grabowski, and Eugene Ie. 2025. Factored agents: Decoupling in-context learning and memorization for robust tool use. arXiv preprint arXiv:2503.22931

    2025 arXiv

  10. [18]

    John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, and Oleg Klimov. 2017. https://arxiv.org/abs/1707.06347 Proximal policy optimization algorithms . Preprint, arXiv:1707.06347

    2017 arXiv

  11. [19]

    Guangming Sheng, Chi Zhang, Zilingfeng Ye, Xibin Wu, Wang Zhang, Ru Zhang, Yanghua Peng, Haibin Lin, and Chuan Wu. 2024. Hybridflow: A flexible and efficient rlhf framework. arXiv preprint arXiv: 2409.19256

    2024 arXiv

  12. [20]

    Alexander Spangher, Tenghao Huang, Philippe Laban, and Nanyun Peng. 2025. https://aclanthology.org/2025.naacl-tutorial.1/ Creative planning with language models: Practice, evaluation and applications . In Proceedings of the 2025 Annual Conference of the Nations of the Americas...

    2025

  13. [21]

    Peiyi Wang, Lei Li, Zhihong Shao, R. X. Xu, Damai Dai, Yifei Li, Deli Chen, Y. Wu, and Zhifang Sui. 2024. https://arxiv.org/abs/2312.08935 Math-shepherd: Verify and reinforce llms step-by-step without human annotations . Preprint, arXiv:2312.08935

    2024 arXiv

  14. [22]

    Shirley Wu, Michel Galley, Baolin Peng, Hao Cheng, Gavin Li, Yao Dou, Weixin Cai, James Zou, Jure Leskovec, and Jianfeng Gao. 2025. https://arxiv.org/abs/2502.00640 Collabllm: From passive responders to active collaborators . Preprint, arXiv:2502.00640

    2025 arXiv

  15. [23]

    Ming Zhong, Da Yin, Tao Yu, Ahmad Zaidi, Mutethia Mutuma, Rahul Jha, Ahmed Hassan Awadallah, Asli Celikyilmaz, Yang Liu, Xipeng Qiu, and 1 others. 2021. Qmsum: A new benchmark for query-based multi-domain meeting summarization. arXiv preprint arXiv:2104.05938

    2021 arXiv

  16. [24]

    Pei Zhou, Andrew Zhu, Jennifer Hu, Jay Pujara, Xiang Ren, Chris Callison-Burch, Yejin Choi, and Prithviraj Ammanabrolu. 2023. I cast detect thoughts: Learning to converse and guide with intents and theory-of-mind in dungeons and dragons. In Proceedings of the 61st Annual Meeti...

    2023

  17. [25]

    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...

  18. [26]

    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 gl...

Pith tools

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