Pith. sign in

REVIEW 3 major objections 5 minor 68 references

Restructuring rationale supervision into dependency-conditioned QA records improves final-answer accuracy across four reasoning benchmarks.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

Dependency-conditioned intermediate QA supervision improves fine-tuned LLM final-answer accuracy over flat chain-of-thought and answer-only baselines on four reasoning benchmarks.

T0 review reviewed 2026-08-01 challenge →

load-bearing objection DAIS is a carefully controlled, modestly useful SFT-supervision paper, but Table 3's internal inconsistency on MedQA/Qwen3-4B IndepQA must be fixed before the dependency-validity claim is credible. the 3 major comments →

arxiv 2607.19088 v1 pith:QUT26SZP submitted 2026-07-21 cs.CL cs.AI

DAIS: Dependency-Aware Intermediate QA Supervision for Complex Reasoning

Yu Wang , Ming Fan , Xicheng Zhang , Zhiyong Li , Zhihu Wang , Caiyue Xu , Dahai Hu , Ting Liu This is my paper
classification cs.CL cs.AI
keywords dependency-aware supervisionchain-of-thoughtintermediate QAsupervised fine-tuningrationale distillationpolicy compliance reasoningsubtask decompositionreasoning benchmarks
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The reading

DAIS claims that the way intermediate reasoning is written as training targets matters as much as whether it is shown. Instead of training on a flat chain-of-thought, the paper converts filtered teacher rationales into stage-level question-answer records in which each later question's input includes the earlier sub-answers it needs. Trained with ordinary supervised fine-tuning and evaluated with direct final-answer prompting, this format beats answer-only SFT, flat CoT, and same-QA-without-dependencies across all four tested benchmarks, with the largest gain (5.6 points) on policy-compliance reasoning. Controlled ablations indicate the improvements come from valid previous-state conditioning, not from longer targets, ordering markers, or QA formatting alone. If right, DAIS shows that dependency structure can be injected at training time at no inference-time cost.

Core claim

DAIS's central claim is that dependency-conditioned intermediate QA records — each predicting a local answer with the preceding sub-answers serialized into the input — provide a better supervised fine-tuning signal for final-answer accuracy than flat rationales or independent sub-questions. The final-answer record stays in the original task format, so the gain is achieved purely through target construction. Across GDPR, AIACT, MedQA, and FOLIO and four backbone models, DAIS achieves the best average accuracy in every benchmark and the best result in 15 of 16 benchmark–model settings; on policy-compliance tasks it improves over the strongest non-DAIS baseline by an average of 4.2 points and u

What carries the argument

The mechanism is a stage-level QA record with serialized previous states. Each teacher rationale is decomposed into 2–5 subtasks, normalized under a dataset-level subtask schema, and materialized as instruction–input–output records: the first record answers a local question from the original input alone, each later record receives a 'Previous states:' block containing the earlier subtask questions and answers, and a separate final record uses only the original input and gold answer. This makes support relations between intermediate conclusions and later decisions visible during training without any inference-time decomposition, controller, or verifier.

Load-bearing premise

The framework assumes that a single fixed dataset-level subtask schema correctly represents the reasoning structure of every retained training instance, and that the rationale-filtering step does not secretly keep only the examples where dependency-conditioned supervision is easiest.

What would settle it

Train the same backbone with DAIS-style records but replace each 'Previous states:' block with the previous states of a different, randomly paired instance (or with corrupted answers) while keeping everything else identical; if final-answer accuracy matches DAIS, the dependency link itself is not the source of the gain. The paper's own CORR. control reports a gap of up to 8.6 points, so a replication where that gap vanishes would refute the central claim.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • Standard supervised fine-tuning pipelines can gain from dependency-ordered target construction, with no change to model architecture or inference procedure.
  • Policy-compliance reasoning, where final judgments depend on chaining facts to applicable rules, benefits most from this supervision (up to 5.6 points).
  • Fewer original training instances may suffice when records carry dependency structure: on one policy benchmark, 25% of the data already beats full-data final-answer SFT.
  • DAIS is complementary to inference-time decomposition; the two could be combined for further gains.
  • The approach does not claim faithful internal reasoning; gains are about supervision signal, not about the model genuinely tracking a dependency graph.

Where Pith is reading between the lines

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

  • If the gain indeed comes from teaching the model to condition later predictions on its own earlier outputs, DAIS is a training-time analogue of inference-time self-consistency or scratchpad use, suggesting a possible synergy with multi-step decoding or iterative self-correction.
  • A testable extension is varying schema granularity: if a generic, domain-agnostic subtask schema performs as well as the task-specific induced schema, then the dependency conditioning, not the schema content, is doing the work.
  • The retained-set bias flagged in the paper implies a measurable risk: if filtering discards examples where dependency-conditioned QA is hard, DAIS may overstate its advantage on easy-to-decompose instances; reporting label-wise retention rates would settle how much this matters.
  • The same record-construction idea could transfer to agentic or tool-use tasks, where subgoal outputs (tool results, intermediate decisions) form natural dependency chains.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

3 major / 5 minor

Summary. The paper proposes DAIS, a training-time data-construction framework that converts filtered teacher-generated chain-of-thought rationales into stage-level QA records. Later intermediate records are conditioned on serialized previous-state answers, while the final-answer record retains the original task format, so inference requires no decomposition module, teacher rationales, or gold intermediate states. The method is evaluated with standard SFT on GDPR, AIACT, MedQA, and FOLIO across four Qwen backbones, with controlled comparisons against final-only SFT, flat CoT SFT, and independent-QA SFT, as well as additional ablations (length-matched, order-marked, corrupted-state, subtask-coverage, and component-isolation controls). The paper reports consistent average gains, largest on policy-compliance benchmarks, and interprets the ablations as evidence that valid previous-state conditioning contributes beyond longer targets or QA formatting alone.

Significance. If the reported results hold, DAIS is a practically attractive auxiliary supervision signal: it requires only standard SFT, does not change inference-time behavior, introduces no fitted constants, and is not circular, since the teacher rationales come from DeepSeek models and evaluation uses gold held-out final answers. The controlled experimental design is a genuine strength: all target-format variants share the same retained instances, optimization budget, decoding settings, and answer matcher, with scores averaged over three runs. The main weakness is an internal inconsistency in the key dependency-validity table, which currently prevents acceptance of the mechanistic interpretation. The schema-induction and filtering steps are also acknowledged limitations that deserve fuller reporting. Overall, the headline accuracy gains are credible, but the paper's distinguishing contribution—dependency conditioning—rests on evidence that needs to be corrected or clarified.

major comments (3)
  1. [§5.2, Table 3] Table 3 reports INDEPQA accuracy on MedQA with Qwen3-4B as 0.574, but Table 1 and Table 9 report the same condition (MedQA, Qwen3-4B, INDEPQA) as 0.750, with DAIS at 0.769 in both. If 0.750 is correct, the DAIS-vs-INDEPQA margin collapses from 19.5 to 1.9 points, and the Section 5.2 conclusion that valid previous-state conditioning contributes beyond independent QA is far weaker. If 0.574 is correct, Tables 1 and 9 contain an error. Either way, the central ablation evidence is internally inconsistent, and because no code or data are released, the discrepancy cannot be resolved externally. Please provide the raw per-run results and reconcile the two tables.
  2. [§5.2, Table 3] The dependency-validity table gives no standard deviations or per-run values, although Section A.5 states that all experiments are repeated three times and Table 9 reports standard deviations. The Table 3 gaps that support the mechanistic claim (e.g., DAIS vs. CORR. on AIACT/Qwen2.5-7B, 8.6 points) could be within run-to-run variation, especially on small test sets such as FOLIO. Since this table carries the claim that valid dependencies matter, the paper should report the same uncertainty information as the main table or explain why it is omitted.
  3. [§3.4 and Limitations] The fixed schema mapping z_it = g(s_it) in Eq. (3) is induced by LLM-assisted clustering plus manual audit, but no coverage statistics, inter-annotator agreement, or assignment-consistency numbers are given. The filtering step in §3.2 discards instances whose teacher rationales fail validation, and the paper itself notes that retention rates are not reported. Because all controlled variants share the same retained set, the internal comparisons are fair, but the claim that DAIS encodes the dependency structure of the original tasks is only as strong as the schema's validity and the retained set's representativeness. Please report retention rates, label-wise retention, and a summary of the schema audit results.
minor comments (5)
  1. [Table 3 header] Table 3 uses 'Qwen-7B' and 'Qwen-4B', while Table 1 uses 'Qwen2.5-7B' and 'Qwen3-4B'. If these are the same models, the naming should be consistent; if not, the reader cannot tell which backbones are ablated.
  2. [Table 4] The 'Subtask1', 'Subtask2', and 'Subtask3' variants are not defined. It would help to state which subtask subset each variant retains and how it is selected from the DAIS record set.
  3. [Table 5 and §5.4] The single-component variants P-only, A-only, and Q-only are described briefly. Please specify how 'plan', 'answer components', and 'questions' are separated from the full DAIS records, since this is the basis for the component-isolation conclusion.
  4. [Table 8] The reference table lists GLM-4.7 with a MedQA score of 0.565, which is far below the other reference models. If this is a typo or a different evaluation setting, it should be corrected; otherwise it may mislead readers about reference-system performance.
  5. [Appendix A.4] The deployment section lists models not used in the main experiments (e.g., Qwen2.5-14B, Qwen3-32B, QwQ-32B). This is not a problem, but a sentence clarifying that these are for a broader deployment suite would avoid confusion.

Circularity Check

0 steps flagged

No significant circularity: DAIS is a supervised target-construction method with held-out gold-label evaluation; the Table 3 inconsistency is an internal-evidence problem, not a circular derivation.

full rationale

DAIS's construction chain is: external teacher rationales (DeepSeek) are filtered by gold-answer match and consistency, decomposed into subtasks, mapped to a dataset-level schema (Eq. 3), converted into dependency-conditioned QA records (Eq. 4), and used as standard SFT targets. Final-answer accuracy is evaluated on held-out benchmark labels with the original task format only; no parameter is fitted to test labels and no reported constant is solved from the evaluation. The IndepQA/CoT/final baselines share the same retained instances and backbones, so the headline gain is a controlled empirical comparison rather than a consequence of construction. The dependency-validity ablation (Table 3) compares target-format variants; even though the states derive from the same teacher rationales, the experimental question is whether training on that format transfers to direct final-answer inference, so the comparison is not an identity. There are no self-citations and no imported uniqueness theorem; the GDPR/AIACT schema is an explicitly declared design choice rather than an appeal to prior work by the same authors. The Limitations section honestly flags possible noise in generated artifacts, retained-set bias, lack of token-budget matching, and the non-faithfulness of intermediate traces—validity/generalizability caveats, not circular reductions. The one notable evidence problem is internal inconsistency: Table 3 reports MedQA/Qwen3-4B IndepQA as 0.574 while Tables 1 and 9 report 0.750 (and AIACT/Qwen3-4B IndepQA differs as 0.534 vs 0.574), weakening the dependency-specific mechanistic claim and requiring a correction or raw results; however, an inconsistent table is not a derivation that equals its input. The absence of released code/data further limits external audit but is a reproducibility concern, not circularity. I therefore find no circular step.

Axiom & Free-Parameter Ledger

0 free parameters · 4 axioms · 0 invented entities

No free parameters are fitted in the conventional sense (no numeric constants tuned on test data). The main non-standard premises are practical: teacher rationale quality, schema validity, and retained-set unbiasedness. The paper explicitly acknowledges the retained-set and teacher-artifact caveats. No invented entities like new forces or particles.

axioms (4)
  • domain assumption Teacher rationales from DeepSeek-V4 models, filtered for final-answer correctness and consistency, are reliable enough to serve as supervision for subtask decomposition.
    Section 3.2 and A.2 rely on DeepSeek-generated rationales as the source of all intermediate QA states; if the teacher rationales contain systematic errors not caught by filtering, the constructed supervision is wrong in ways that could affect results.
  • domain assumption The dataset-level subtask schema induction (LLM clustering plus manual audit) produces a valid mapping from free-form subtasks to a small set of categories, and each instance subtask is correctly mapped to one schema type.
    Section 3.4: schema induction uses LLM-based clustering with consistency checks and manual audit but no objective criterion for schema correctness; errors here propagate to every dependency-conditioned record. The paper does not report agreement or audit statistics.
  • domain assumption The filtering step discarding instances whose teacher rationale fails validation does not systematically bias the retained training set in a way that favors DAIS over baselines.
    Limitations explicitly state the retained set may be biased and that future work should report retention rates; the paper does not report how many instances were discarded or their properties. While all controlled variants use the same retained set, the bias could interact with target format, so it is an unverified assumption.
  • standard math The standard autoregressive instruction-tuning objective with loss only on output tokens is an appropriate and sufficient training objective for the constructed multi-record supervision.
    Section 3.6 uses the usual SFT objective; this is standard practice, not an ad hoc assumption.

reviewed 2026-08-01 · how reviews work

0 comments
Cite this review

Pith. "Pith review of DAIS: Dependency-Aware Intermediate QA Supervision for Complex Reasoning." pith.science (2026). https://pith.science/paper/QUT26SZP

@misc{pith2026260719088,
  author       = {Pith},
  title        = {Pith review of: DAIS: Dependency-Aware Intermediate QA Supervision for Complex Reasoning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/QUT26SZP}},
  note         = {Machine review of arXiv:2607.19088}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Chain-of-thought (CoT) supervision exposes intermediate rationales, but flat rationale targets usually optimize a single reasoning sequence and provide limited supervision on how local conclusions should support later decisions. We introduce Dependency-Aware Intermediate QA Supervision (DAIS), a training-time framework that converts filtered teacher rationales into stage-level QA records. Each intermediate record predicts a local answer conditioned on the previous states needed for that decision, while the final-answer record keeps the original task format; evaluation therefore uses only the original input and optional context. Across GDPR, AIACT, MedQA, and FOLIO with multiple Qwen backbones, DAIS improves average final-answer accuracy over answer-only, flat chain-of-thought, and independent-QA baselines. On policy-compliance benchmarks, it achieves a largest gain of 5.6% and an average gain of 4.2% over the strongest non-DAIS baseline. Controlled ablations show that valid previous-state conditioning contributes beyond longer targets or additional intermediate text, supporting dependency-conditioned intermediate QA as a lightweight auxiliary supervision signal for standard final-answer inference.

Figures

Figures reproduced from arXiv: 2607.19088 by Caiyue Xu, Dahai Hu, Ming Fan, Ting Liu, Xicheng Zhang, Yu Wang, Zhihu Wang, Zhiyong Li.

Figure 1
Figure 1. Figure 1: Comparison between conventional supervised targets and [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Overview of DAIS data construction. Teacher CoT rationales are decomposed into subtasks, normalized into dataset-level schemas, and converted into stage-level QA records. Later intermediate records receive serialized previous subtask outputs as context, whereas the final-answer record remains the original final-task record and does not receive constructed intermediate states. Uesato et al., 2022; Lightman … view at source ↗
Figure 3
Figure 3. Figure 3: Examples of task-conditioned subtask schemas. whose rationale is internally consistent: its sub￾stantive reasoning steps should be grounded in the input, context, answer options, or preceding steps, and should not support a different conclusion. If no candidate passes the filter, the instance is dis￾carded. The retained rationales are treated as silver supervision for subsequent subtask decomposition and d… view at source ↗
Figure 4
Figure 4. Figure 4: Sample efficiency on GDPR and MedQA with [PITH_FULL_IMAGE:figures/full_fig_p008_4.png] view at source ↗

discussion (0)

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

Reference graph

Works this paper leans on

68 extracted references · 15 linked inside Pith

  1. [2]

    Maciej Besta, Nils Blach, Ales Kubicek, Robert Gerstenberger, Michal Podstawski, Lukas Gianinazzi, Joanna Gajda, Tomasz Lehmann, Hubert Niewiadomski, Piotr Nyczyk, and 1 others. 2024. Graph of thoughts: Solving elaborate problems with large language models. In Proceedings of the AAAI conference on artificial intelligence, volume 38, pages 17682--17690

    2024

  2. [4]

    DeepSeek-AI. 2026. Deepseek-v4: Towards highly efficient million-token context intelligence

    2026

  3. [5]

    Jay DeYoung, Sarthak Jain, Nazneen Fatema Rajani, Eric Lehman, Caiming Xiong, Richard Socher, and Byron C Wallace. 2020. Eraser: A benchmark to evaluate rationalized nlp models. In Proceedings of the 58th annual meeting of the association for computational linguistics, pages 4443--4458

    2020

  4. [6]

    Luyu Gao, Aman Madaan, Shuyan Zhou, Uri Alon, Pengfei Liu, Yiming Yang, Jamie Callan, and Graham Neubig. 2023. Pal: Program-aided language models. In International conference on machine learning, pages 10764--10799. PMLR

    2023

  5. [8]

    Cheng-Yu Hsieh, Chun-Liang Li, Chih-Kuan Yeh, Hootan Nakhost, Yasuhisa Fujii, Alex Ratner, Ranjay Krishna, Chen-Yu Lee, and Tomas Pfister. 2023. Distilling step-by-step! outperforming larger language models with less training data and smaller model sizes. In Findings of the Association for Computational Linguistics: ACL 2023, pages 8003--8017

    2023

  6. [9]

    Wenbin Hu, Haoran Li, Huihao Jing, Qi Hu, Ziqian Zeng, Sirui Han, Xu Heli, Tianshu Chu, Peizhao Hu, and Yangqiu Song. 2025. Context reasoner: Incentivizing reasoning capability for contextualized privacy and safety compliance via reinforcement learning. In Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing, pages 865--883

    2025

  7. [10]

    Sarthak Jain and Byron C Wallace. 2019. Attention is not explanation. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 3543--3556

    2019

  8. [13]

    Takeshi Kojima, Shixiang Shane Gu, Machel Reid, Yutaka Matsuo, and Yusuke Iwasawa. 2022. Large language models are zero-shot reasoners. Advances in neural information processing systems, 35:22199--22213

    2022

  9. [14]

    Gonzalez, Hao Zhang, and Ion Stoica

    Woosuk Kwon, Zhuohan Li, Siyuan Zhuang, Ying Sheng, Lianmin Zheng, Cody Hao Yu, Joseph E. Gonzalez, Hao Zhang, and Ion Stoica. 2023. Efficient memory management for large language model serving with pagedattention. In Proceedings of the ACM SIGOPS 29th Symposium on Operating Systems Principles

    2023

  10. [16]

    Tao Lei, Regina Barzilay, and Tommi Jaakkola. 2016. Rationalizing neural predictions. In Proceedings of the 2016 conference on empirical methods in natural language processing, pages 107--117

    2016

  11. [17]

    Haoran Li, Wenbin Hu, Huihao Jing, Yulin Chen, Qi Hu, Sirui Han, Tianshu Chu, Peizhao Hu, and Yangqiu Song. 2025. Privaci-bench: Evaluating privacy with contextual integrity and legal compliance. In Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 10544--10559

    2025

  12. [18]

    Hunter Lightman, Vineet Kosaraju, Yuri Burda, Harrison Edwards, Bowen Baker, Teddy Lee, Jan Leike, John Schulman, Ilya Sutskever, and Karl Cobbe. 2024. Let's verify step by step. In International Conference on Learning Representations, volume 2024, pages 39578--39601

    2024

  13. [19]

    Lucie Charlotte Magister, Jonathan Mallinson, Jakub Adamek, Eric Malmi, and Aliaksei Severyn. 2023. Teaching small language models to reason. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers), pages 1773--1781

    2023

  14. [20]

    MiniMax-AI . 2026. MiniMax-M2.7 . https://github.com/MiniMax-AI/MiniMax-M2.7. GitHub repository

    2026

  15. [21]

    Maxwell Nye, Anders Johan Andreassen, Guy Gur-Ari, Henryk Michalewski, Jacob Austin, David Bieber, David Dohan, Aitor Lewkowycz, Maarten Bosma, David Luan, and 1 others. 2021. Show your work: Scratchpads for intermediate computation with language models

    2021

  16. [22]

    Ofir Press, Muru Zhang, Sewon Min, Ludwig Schmidt, Noah A Smith, and Mike Lewis. 2023. Measuring and narrowing the compositionality gap in language models. In Findings of the Association for Computational Linguistics: EMNLP 2023, pages 5687--5711

    2023

  17. [23]

    Qwen, :, An Yang, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chengyuan Li, Dayiheng Liu, Fei Huang, Haoran Wei, Huan Lin, Jian Yang, Jianhong Tu, Jianwei Zhang, Jianxin Yang, Jiaxi Yang, Jingren Zhou, and 25 others. 2025. https://arxiv.org/abs/2412.15115 Qwen2.5 technical report . Preprint, arXiv:2412.15115

    Pith/arXiv arXiv 2025

  18. [24]

    GLM Team, Aohan Zeng, Xin Lv, Qinkai Zheng, Zhenyu Hou, Bin Chen, Chengxing Xie, Cunxiang Wang, Da Yin, Hao Zeng, Jiajie Zhang, Kedong Wang, Lucen Zhong, Mingdao Liu, Rui Lu, Shulin Cao, Xiaohan Zhang, Xuancheng Huang, Yao Wei, and 152 others. 2025. https://arxiv.org/abs/2508.06471 Glm-4.5: Agentic, reasoning, and coding (arc) foundation models . Preprint...

    Pith/arXiv arXiv 2025

  19. [25]

    Miles Turpin, Julian Michael, Ethan Perez, and Samuel Bowman. 2023. Language models don't always say what they think: Unfaithful explanations in chain-of-thought prompting. Advances in Neural Information Processing Systems, 36:74952--74965

    2023

  20. [28]

    Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, and 1 others. 2022. Chain-of-thought prompting elicits reasoning in large language models. Advances in neural information processing systems, 35:24824--24837

    2022

  21. [29]

    An Yang, Anfeng Li, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Gao, Chengen Huang, Chenxu Lv, Chujie Zheng, Dayiheng Liu, Fan Zhou, Fei Huang, Feng Hu, Hao Ge, Haoran Wei, Huan Lin, Jialong Tang, and 41 others. 2025. https://arxiv.org/abs/2505.09388 Qwen3 technical report . Preprint, arXiv:2505.09388

    Pith/arXiv arXiv 2025

  22. [30]

    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

  23. [32]

    Eric Zelikman, Yuhuai Wu, Jesse Mu, and Noah Goodman. 2022. Star: Bootstrapping reasoning with reasoning. Advances in Neural Information Processing Systems, 35:15476--15488

    2022

  24. [34]

    Proceedings of the 24th international conference on Machine learning , pages=

    Scalable training of l 1-regularized log-linear models , author=. Proceedings of the 24th international conference on Machine learning , pages=

  25. [35]

    Dan Gusfield , title =. 1997

    1997

  26. [36]

    Tetreault , title =

    Mohammad Sadegh Rasooli and Joel R. Tetreault , title =. Computing Research Repository , volume =. 2015 , url =

    2015

  27. [37]

    A Framework for Learning Predictive Structures from Multiple Tasks and Unlabeled Data , Volume =

    Ando, Rie Kubota and Zhang, Tong , Issn =. A Framework for Learning Predictive Structures from Multiple Tasks and Unlabeled Data , Volume =. Journal of Machine Learning Research , Month = dec, Numpages =

  28. [38]

    Advances in neural information processing systems , volume=

    Chain-of-thought prompting elicits reasoning in large language models , author=. Advances in neural information processing systems , volume=

  29. [39]

    , author=

    Lora: Low-rank adaptation of large language models. , author=. Iclr , volume=

  30. [40]

    arXiv preprint arXiv:2009.13081 , year=

    What Disease does this Patient Have? A Large-scale Open Domain Question Answering Dataset from Medical Exams , author=. arXiv preprint arXiv:2009.13081 , year=

    Pith/arXiv arXiv 2009

  31. [41]

    Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=

    Privaci-bench: Evaluating privacy with contextual integrity and legal compliance , author=. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=

  32. [42]

    Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing , pages=

    Context reasoner: Incentivizing reasoning capability for contextualized privacy and safety compliance via reinforcement learning , author=. Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing , pages=

    2025

  33. [43]

    arXiv preprint arXiv:2203.11171 , year=

    Self-consistency improves chain of thought reasoning in language models , author=. arXiv preprint arXiv:2203.11171 , year=

    Pith/arXiv arXiv

  34. [44]

    URL https://arxiv

    Least-to-most prompting enables complex reasoning in large language models, 2023 , author=. URL https://arxiv. org/abs/2205.10625 , year=

    Pith/arXiv arXiv 2023

  35. [45]

    Findings of the Association for Computational Linguistics: EMNLP 2023 , pages=

    Measuring and narrowing the compositionality gap in language models , author=. Findings of the Association for Computational Linguistics: EMNLP 2023 , pages=

    2023

  36. [46]

    Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers) , pages=

    Teaching small language models to reason , author=. Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers) , pages=

  37. [47]

    URL https://arxiv

    Training verifiers to solve math word problems, 2021 , author=. URL https://arxiv. org/abs/2110.14168 , volume=

    Pith/arXiv arXiv 2021

  38. [48]

    arXiv preprint arXiv:2103.03874 , year=

    Measuring mathematical problem solving with the math dataset , author=. arXiv preprint arXiv:2103.03874 , year=

    Pith/arXiv arXiv

  39. [49]

    Proceedings of the 2018 conference on empirical methods in natural language processing , pages=

    HotpotQA: A dataset for diverse, explainable multi-hop question answering , author=. Proceedings of the 2018 conference on empirical methods in natural language processing , pages=

    2018

  40. [50]

    Transactions of the Association for Computational Linguistics , volume=

    Did aristotle use a laptop? a question answering benchmark with implicit reasoning strategies , author=. Transactions of the Association for Computational Linguistics , volume=. 2021 , publisher=

    2021

  41. [51]

    Findings of the Association for Computational Linguistics: ACL 2023 , pages=

    Challenging big-bench tasks and whether chain-of-thought can solve them , author=. Findings of the Association for Computational Linguistics: ACL 2023 , pages=

    2023

  42. [52]

    arXiv preprint arXiv:2211.14275 , year=

    Solving math word problems with process-and outcome-based feedback , author=. arXiv preprint arXiv:2211.14275 , year=

    Pith/arXiv arXiv

  43. [53]

    Advances in Neural Information Processing Systems , volume=

    Language models don't always say what they think: Unfaithful explanations in chain-of-thought prompting , author=. Advances in Neural Information Processing Systems , volume=

  44. [54]

    Proceedings of the 57th annual meeting of the association for computational linguistics , pages=

    Multi-hop reading comprehension through question decomposition and rescoring , author=. Proceedings of the 57th annual meeting of the association for computational linguistics , pages=

  45. [55]

    Proceedings of the 2021 conference on empirical methods in natural language processing , pages=

    Explaining answers with entailment trees , author=. Proceedings of the 2021 conference on empirical methods in natural language processing , pages=

    2021

  46. [56]

    Show your work: Scratchpads for intermediate computation with language models , author=

  47. [57]

    Advances in neural information processing systems , volume=

    Large language models are zero-shot reasoners , author=. Advances in neural information processing systems , volume=

  48. [58]

    Advances in Neural Information Processing Systems , volume=

    Star: Bootstrapping reasoning with reasoning , author=. Advances in Neural Information Processing Systems , volume=

  49. [59]

    Findings of the Association for Computational Linguistics: ACL 2023 , pages=

    Distilling step-by-step! outperforming larger language models with less training data and smaller model sizes , author=. Findings of the Association for Computational Linguistics: ACL 2023 , pages=

    2023

  50. [60]

    International Conference on Learning Representations , volume=

    Let's verify step by step , author=. International Conference on Learning Representations , volume=

  51. [61]

    arXiv preprint arXiv:2210.02406 , year=

    Decomposed prompting: A modular approach for solving complex tasks , author=. arXiv preprint arXiv:2210.02406 , year=

    Pith/arXiv arXiv

  52. [62]

    arXiv preprint arXiv:2205.09712 , year=

    Selection-inference: Exploiting large language models for interpretable logical reasoning , author=. arXiv preprint arXiv:2205.09712 , year=

    Pith/arXiv arXiv

  53. [63]

    arXiv preprint arXiv:2210.03629 , year=

    React: Synergizing reasoning and acting in language models , author=. arXiv preprint arXiv:2210.03629 , year=

    Pith/arXiv arXiv

  54. [64]

    International conference on machine learning , pages=

    Pal: Program-aided language models , author=. International conference on machine learning , pages=. 2023 , organization=

    2023

  55. [65]

    Advances in neural information processing systems , volume=

    Tree of thoughts: Deliberate problem solving with large language models , author=. Advances in neural information processing systems , volume=

  56. [66]

    Proceedings of the AAAI conference on artificial intelligence , volume=

    Graph of thoughts: Solving elaborate problems with large language models , author=. Proceedings of the AAAI conference on artificial intelligence , volume=

  57. [67]

    Proceedings of the 2016 conference on empirical methods in natural language processing , pages=

    Rationalizing neural predictions , author=. Proceedings of the 2016 conference on empirical methods in natural language processing , pages=

    2016

  58. [68]

    Proceedings of the 58th annual meeting of the association for computational linguistics , pages=

    ERASER: A benchmark to evaluate rationalized NLP models , author=. Proceedings of the 58th annual meeting of the association for computational linguistics , pages=

  59. [69]

    Attention is not explanation , author=. Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers) , pages=

    2019

  60. [70]

    arXiv preprint arXiv:2307.13702 , year=

    Measuring faithfulness in chain-of-thought reasoning , author=. arXiv preprint arXiv:2307.13702 , year=

    Pith/arXiv arXiv

  61. [71]

    arXiv preprint arXiv:2209.00840 , url =

    FOLIO: Natural Language Reasoning with First-Order Logic , author =. arXiv preprint arXiv:2209.00840 , url =

    Pith/arXiv arXiv

  62. [72]

    Proceedings of the ACM SIGOPS 29th Symposium on Operating Systems Principles , year=

    Efficient Memory Management for Large Language Model Serving with PagedAttention , author=. Proceedings of the ACM SIGOPS 29th Symposium on Operating Systems Principles , year=

  63. [73]

    2025 , eprint=

    Qwen2.5 Technical Report , author=. 2025 , eprint=

    2025

  64. [74]

    2025 , eprint=

    Qwen3 Technical Report , author=. 2025 , eprint=

    2025

  65. [75]

    2025 , eprint=

    GLM-4.5: Agentic, Reasoning, and Coding (ARC) Foundation Models , author=. 2025 , eprint=

    2025

  66. [76]

    arXiv preprint arXiv:2508.10925 , year=

    gpt-oss-120b & gpt-oss-20b model card , author=. arXiv preprint arXiv:2508.10925 , year=

    Pith/arXiv arXiv

  67. [77]

    DeepSeek-V4: Towards Highly Efficient Million-Token Context Intelligence , author=

  68. [78]

    2026 , howpublished =

    2026

This paper was first reviewed by deepseek-v4-flash on August 1, 2026.