Pith. sign in

REVIEW 4 major objections 6 minor 29 references

A prompting framework that forces a conclusion-first, professionally structured explanation outperforms chain-of-thought across high-stakes classification tasks.

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 ↗

T0 review · deepseek-v4-flash

2026-08-03 11:07 UTC pith:T5XCJU4X

load-bearing objection Useful framing, but the headline accuracy comparison is unverifiable as reported because answer extraction is never specified. the 4 major comments →

arxiv 2601.07233 v1 pith:T5XCJU4X submitted 2026-01-12 cs.AI

From "Thinking" to "Justifying": Aligning High-Stakes Explainability with Professional Communication Standards

Chen Qian , Yimeng Wang , Yu Chen , Lingfei Wu , Andreas Stathopoulos This is my paper
classification cs.AI
keywords explainable AIchain-of-thoughtstructured justificationconclusion-first promptingCREAC/BLUFheuristic evaluation metricshigh-stakes NLPLLM prompting
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.

This paper argues that explanations from large language models in high-stakes settings should mirror professional communication practice: give the conclusion first, then a structured defense, rather than a free-form reasoning trace. The authors introduce a prompting framework that enforces a four-section format—commit, ground, defend, conclude—and six rule-based metrics that score how closely an explanation follows that format. Across four yes/no classification tasks in law, medicine, and finance, the framework outperforms chain-of-thought prompting by 5.3 points, and every metric correlates with correctness, suggesting that structure itself is a signal of reliability. If the claim holds, it means verifiable explanations and accurate answers can be pursued with the same intervention, not traded off against each other.

Core claim

Under the 'Result → Justify' paradigm, SEF constrains the model's output, not its hidden reasoning: the model must open with an explicit answer, list key facts, analyze how each fact supports the answer using domain terms, and close with a clear restatement. The paper claims this conclusion-first scaffold is why SEF reaches 83.9% average accuracy across four binary tasks and three high-stakes domains, beating chain-of-thought (78.6%) and even direct answer-only prompting (80.9%). The central discovery is the association between output structure and correctness: all six metrics computed on 90,608 outputs from all methods and models correlate with accuracy (r = 0.20–0.42, p < 0.001), and ablat

What carries the argument

SEF (Structured Explainability Framework): a prompting specification plus a matching suite of six rule-based metrics—Answer First/Last, Answer Clarity, Conclusion Isolation, Domain Terminology Consistency, Conclusion–Evidence Alignment, and Fact Specificity—that together operationalize the CREAC and BLUF professional-writing conventions as a four-section output template (Commit, Ground, Defend, Conclude). The metrics double as both the evaluation instrument and the design of the prompt; they are the mechanism that connects 'conclusion-first format' to measurable accuracy and that lets the paper test which part of the scaffold carries the gain.

Load-bearing premise

The measured accuracy depends on how the model's final Yes/No answer is parsed out of its free-form response, and the paper never specifies that extraction rule; if parsing keys on the 'My answer is' line that SEF's template forces, then ablations that remove that line (and methods that don't produce it) could be scored wrong even when the model's underlying answer is right.

What would settle it

Re-run the same four tasks and baselines with a parser that either scans the entire output for any clear Yes/No statement or has humans annotate the model's intended answer, then compare SEF's accuracy against CoT's under that fair extraction. If the +5.3 gap shrinks below significance, the conclusion-first advantage is an extraction artifact rather than a reasoning benefit.

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

If this is right

  • Structured justification can serve as a cheap, deterministic, first-pass filter for explanation quality in legal, medical, and financial NLP, without needing a model-based judge.
  • Conclusion-first prompting is a drop-in, training-free intervention that improves average accuracy over chain-of-thought across the tested tasks, so practitioners in high-stakes settings can adopt it immediately.
  • Because plausibility constraints (answer placement, clarity, isolation) drive most of the accuracy gain, future prompting designs should treat output organization as a first-class variable, not a cosmetic concern.
  • Human review becomes easier: even when the answer is wrong, a structured justification makes the error more visible to a reader, shifting the failure mode from hidden drift to checkable claims.
  • The Direct > CoT result implies that free-form reasoning traces can actively hurt reliability; constraining the output rather than eliciting more of it may be the safer default.

Where Pith is reading between the lines

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

  • The metrics' correlation with correctness may be partly a parsing artifact: SEF's template forces a literal 'My answer is: [Yes/No]' line, and the paper never describes how answers were extracted from free-form text. If extraction used that marker, non-SEF methods could be under-scored mechanically, inflating the reported +5.3 advantage.
  • The six metrics, being regex-based and cheap, could be repurposed as reward signals for reinforcement learning or as preference labels for fine-tuning, not just as zero-shot prompt constraints—a direction the paper leaves implicit.
  • The binary Yes/No design likely inflates the practical benefit of answer-first formatting; in open-ended generation, where there is no single answer to place first, the framework would need an analog of 'bottom line' that may not transfer cleanly.
  • A natural stress test: apply SEF to math or symbolic reasoning benchmarks where chain-of-thought is strong; if conclusion-first formatting still wins, structure is the driving factor, and if it loses, the benefit may be domain-specific to argumentative, evidence-grounded tasks.

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

4 major / 6 minor

Summary. The paper proposes SEF (Structured Explainability Framework), a conclusion-first ('Result → Justify') prompting scheme for LLM explanations in high-stakes domains, coupled with six rule-based metrics (AFL, AC, CI, DTC, CEA, FS) intended to quantify structure and grounding. Across four Yes/No tasks in legal, medical, and financial domains, the authors report that SEF achieves 83.9% average accuracy, outperforming CoT by 5.3 points, and that all six metrics correlate significantly with correctness (r = 0.20–0.42, p < 0.001). Ablations removing plausibility components (especially CI) yield large accuracy drops, which the paper interprets as evidence that structured presentation is the main driver of reliability.

Significance. If correct, the paper offers a practical, deterministic, and cheaply computable way to evaluate explanation quality, and it delivers a clear, falsifiable claim that conclusion-first structured justifications improve LLM accuracy on high-stakes binary tasks. The use of open-weight models, public datasets, and rule-based (non-learned) metrics is a genuine strength for reproducibility. However, the central empirical claims currently rest on two unverified assumptions: that the final Yes/No answer is extracted from free-form text in a method-agnostic way, and that the reported sample size for the correlation analysis is accurate. Both are load-bearing and can be addressed with additional disclosure and re-analysis.

major comments (4)
  1. [§4.1 / Appendix A.3 / Table 5] The paper never specifies how the model's final Yes/No answer is extracted from the generated text. The SEF prompt (Table 5) mandates the output to end with 'My answer is: [Yes/No]', while the CoT prompt (A.3) asks for open-ended step-by-step reasoning with no required final marker. If the extraction uses a regex on 'My answer is' or the last Yes/No token, SEF outputs are trivially parseable whereas CoT outputs may be mis-scored, making the reported +5.3-point advantage potentially an artifact of the parsing rule. The same issue affects the w/o CI ablation, whose output lacks the mandated marker. Please provide the exact extraction code/rule per method and report extraction success rates; without this, the accuracy comparison in Table 2 and the metric-correctness correlations in Table 1 are not independently verifiable.
  2. [§4.2 / Table 1] The stated correlation sample size n=90,608 is inconsistent with the paper's own setup: 1,618 test samples × 4 models × 7 methods equals 45,304 outputs, not 90,608. The discrepancy is exactly a factor of two, suggesting a counting error (e.g., double-counting or an unstated extra condition). Since the p-values and correlation estimates depend directly on n, this must be reconciled. Additionally, the paper acknowledges in Limitations that samples are not independent due to shared prompts, models, and datasets; the reported p<0.001 values should be accompanied by cluster-robust or mixed-effects estimates to support the claimed significance.
  3. [§4.4 / Table 3] The w/o CI ablation removes the 'CONCLUSION' section, which is the section that contains the 'End with: My answer is: [Yes/No]' instruction. The observed −22.5-point drop may therefore be a mechanical consequence of losing the parseable answer marker, not a demonstration that conclusion-first structure drives accuracy. The same confound applies to w/o Pres (−32.9), which removes all plausibility sections and likely the marker as well. To support the paper's central interpretation, please provide a control that holds answer extraction constant, e.g., parse the last Yes/No token in the output regardless of section, or add the answer marker post-hoc for all variants, and report the resulting ablation deltas.
  4. [§3 / Table 5 / Table 4] The six metrics are defined to reward exactly the formatting elements the SEF prompt enforces: AFL rewards an answer in the first/last 200 characters, which the prompt requires in 'ANSWER PREVIEW'; CI rewards 'CONCLUSION' headers; AC rewards 'My answer is'; DTC/FS/CEA reward the terminological and evidence-linking language requested in the 'KEY FACTS' and 'ANALYSIS' sections. Consequently, metric scores on SEF outputs partly measure prompt compliance, and the reported metric-accuracy correlations across all methods may be confounded by method identity. The claim that the six metrics provide a general signal of explanation quality would be substantially strengthened by reporting within-method correlations (e.g., within CoT or within SEF) or partial correlations controlling for method, to demonstrate that structure predicts correctness beyond the SEF prompt's formatting instructions.
minor comments (6)
  1. [Table 2 and Table 3] No confidence intervals or per-model variance are reported. With only four models, the 5.3-point average difference could be driven by one model; a per-model breakdown (or at least min/max across models) would help.
  2. [Figure 1] The right-side PubMedQA example is dense and the font is very small; the mapping between the four SEF steps and the six metrics is hard to read. Consider enlarging the in-figure text or splitting the example into a separate table.
  3. [§3 / Table 4] The metric thresholds (e.g., AFL's 200-character window, AC's 0.7/0.3 weights, DTC's term-count cutoffs, CEA/FS thresholds) appear arbitrary. A sensitivity analysis (e.g., varying the window to 150/250 or changing DTC cutoffs) would strengthen confidence that the metrics are not brittle to these choices.
  4. [Appendix A.4] Hearsay contains only 94 test samples, so the 54.5% accuracy there is based on very few examples; this should be acknowledged in the main text as a low-power result.
  5. [§4.3 / Table 2] The claim that Direct (80.9%) outperforms CoT (78.6%) is interesting but is stated without a significance test or variance estimate. Given the small number of models and task-dependent variation, a paired test across the four models would be more convincing.
  6. [Appendix A.3] The 'Tree-of-Thought' implementation is a three-stage prompting heuristic, not the full tree search described in Yao et al. (2023). This is acceptable but should be explicitly described as a prompting-only approximation to avoid overclaiming the baseline's fidelity.

Circularity Check

0 steps flagged

No significant circularity: the headline accuracy claim rests on external labels, and no prediction reduces to the SEF metrics by construction.

full rationale

The paper's central accuracy result (SEF 83.9% vs. CoT 78.6%) is evaluated against external ground-truth labels from standard benchmarks (PubMedQA, ConsumerQA, Hearsay, FPB), so it is not derived from the SEF metrics or from the prompting format itself. The six metrics in A.1 are rule-based regex/lexicon heuristics, not parameters fitted to accuracy, and the claimed metric–accuracy correlations are empirical associations, not definitional equivalences. The closest concern is that the SEF prompt template in Table 5 mandates the exact section headers and final-answer marker that the metrics reward (e.g., 'CONCLUSION:' for CI, 'My answer is: [Yes/No]' for AFL/AC), so metric scores for SEF outputs partly reflect prompt compliance. However, the correlation is computed across all methods (§4.2), and the paper explicitly frames the metrics as structural proxies rather than semantic truth in the Limitations section. The w/o CI ablation (−22.5 in Table 3) removes the mandated final-answer marker, which could be a serious confound if answer extraction depended on that marker, but the paper does not specify its extraction rule, so this remains an unexplained methodological risk rather than a demonstrated circular reduction. No load-bearing self-citations, imported uniqueness theorems, or ansatz-smuggling via citation are present. Therefore, no specific circular step can be exhibited under the stated standards.

Axiom & Free-Parameter Ledger

6 free parameters · 5 axioms · 0 invented entities

The paper does not postulate any unobserved physical entities; SEF and its six metrics are directly implemented and measured. The free parameters are the hand-chosen thresholds in the heuristic scoring rules, which shape the correlation and ablation results. The main axioms are the normative value of CREAC/BLUF, the sufficiency of binary accuracy, and the implicit answer-extraction assumption.

free parameters (6)
  • AFL window size = 200 characters
    Hand-chosen window for 'answer in first/last 200 chars'; changing it changes AFL scores and correlations.
  • AC regex thresholds = 1.0 / 0.7 / 0.3
    Hand-set scores for strong, medium, and weak answer-clarity patterns.
  • CI weights = 1.0 / 0.6
    Hand-set scores for explicit conclusion header vs. discourse markers.
  • DTC thresholds = ≥5→1.0, ≥3→0.8, ≥1→0.5, else→0.2
    Hand-set term-count thresholds; also DTC=0.5 default when no lexicon is defined.
  • CEA thresholds = ≥3 link & ≥1 analysis→1.0; ≥2 link→0.8; ≥1 link→0.5; ≥1 analysis→0.3; else 0
    Hand-set thresholds for evidence-linking and analysis cues.
  • FS thresholds = ≥4 spec & ≤1 vague→1.0; ≥3→0.8; ≥2→0.6; ≥1→0.4; else 0.2
    Hand-set thresholds for specificity and vagueness indicators.
axioms (5)
  • domain assumption Conclusion-first, structured justification (CREAC/BLUF) is a desirable norm for machine explanations.
    Invoked in §1 and §3 to justify the framework; no empirical evidence for human verifiability is provided.
  • domain assumption Binary Yes/No accuracy is a sufficient performance metric for the claimed reliability improvement.
    Stated in §4.1: 'The binary format makes accuracy a sufficient metric'; limits generality to binary tasks.
  • domain assumption Six heuristic structural scores measure justification quality (plausibility/faithfulness).
    Defined in §3 and Appendix A; the Limitations admit they measure structural compliance, not semantic correctness.
  • ad hoc to paper The generated outputs contain a uniquely extractable binary answer.
    All accuracy and ablation results depend on an unstated answer-extraction rule; the w/o CI drop suggests the rule relies on the final answer marker.
  • domain assumption Pearson p-values are meaningful under non-independent sampling.
    The Limitations acknowledge samples may be non-independent but the paper reports p<0.001 as if independent.

pith-pipeline@v1.3.0-alltime-deepseek · 9938 in / 13484 out tokens · 130285 ms · 2026-08-03T11:07:42.337137+00:00 · methodology

0 comments
read the original abstract

Explainable AI (XAI) in high-stakes domains should help stakeholders trust and verify system outputs. Yet Chain-of-Thought methods reason before concluding, and logical gaps or hallucinations can yield conclusions that do not reliably align with their rationale. Thus, we propose "Result -> Justify", which constrains the output communication to present a conclusion before its structured justification. We introduce SEF (Structured Explainability Framework), operationalizing professional conventions (e.g., CREAC, BLUF) via six metrics for structure and grounding. Experiments across four tasks in three domains validate this approach: all six metrics correlate with correctness (r=0.20-0.42; p<0.001), and SEF achieves 83.9% accuracy (+5.3 over CoT). These results suggest structured justification can improve verifiability and may also improve reliability.

Figures

Figures reproduced from arXiv: 2601.07233 by Andreas Stathopoulos, Chen Qian, Lingfei Wu, Yimeng Wang, Yu Chen.

Figure 1
Figure 1. Figure 1: SEF: Structured Explainability via CREAC/BLUF Conventions. Left: Paradigm comparison showing CoT (“Reasoning → Result”) vs. SEF (“Result → Justify”). CoT traces can be hard to verify (Direct 80.9% > CoT 78.6%); SEF achieves 83.9%. Right: PubMedQA example demonstrating the 4-step prompt (Commit, Ground, Defend, Conclude) with metric mappings. scaffold is the main driver of accuracy: remov￾ing plausibility c… 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

29 extracted references · 2 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]

    Martin Aigner and G\" u nter M. Ziegler. 2010. https://doi.org/10.1007/978-3-642-00856-6 Proofs from THE BOOK . Springer Berlin Heidelberg

    work page doi:10.1007/978-3-642-00856-6 2010

  4. [4]

    Akari Asai, Zeqiu Wu, Yizhong Wang, Avirup Sil, and Hannaneh Hajishirzi. 2023. https://doi.org/10.48550/ARXIV.2310.11511 Self-rag: Learning to retrieve, generate, and critique through self-reflection . arXiv preprint

    Pith/arXiv arXiv doi:10.48550/arxiv.2310.11511 2023

  5. [5]

    Jinze Bai, Shuai Bai, Yunfei Chu, Zeyu Cui, Kai Dang, Xiaodong Deng, Yang Fan, Wenbin Ge, Yu Han, Fei Huang, Binyuan Hui, Luo Ji, Mei Li, Junyang Lin, Runji Lin, Dayiheng Liu, Gao Liu, Chengqiang Lu, Keming Lu, and 29 others. 2023. https://doi.org/10.48550/ARXIV.2309.16609 Qwen technical report . arXiv preprint

    Pith/arXiv arXiv doi:10.48550/arxiv.2309.16609 2023

  6. [6]

    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. 2021. https://doi.org/10.48550/ARXI...

    Pith/arXiv arXiv doi:10.48550/arxiv.2108.07258 2021

  7. [7]

    Steven J Burton. 2007. https://books.google.com/books?id=h1BsEAAAQBAJ An introduction to law and legal reasoning . Wolters Kluwer Law & Business

    2007

  8. [8]

    DeepSeek-AI , Daya Guo, Dejian Yang, Haowei Zhang, Junxiao Song, Ruoyu Zhang, Runxin Xu, Qihao Zhu, Shirong Ma, Peiyi Wang, Xiao Bi, Xiaokang Zhang, Xingkai Yu, Yu Wu, Z. F. Wu, Zhibin Gou, Zhihong Shao, Zhuoshu Li, Ziyi Gao, and 181 others. 2025. https://doi.org/10.48550/ARXIV.2501.12948 Deepseek-r1: Incentivizing reasoning capability in llms via reinfor...

    Pith/arXiv arXiv doi:10.48550/arxiv.2501.12948 2025

  9. [9]

    Shehzaad Dhuliawala, Mojtaba Komeili, Jing Xu, Roberta Raileanu, Xian Li, Asli Celikyilmaz, and Jason Weston. 2024. https://doi.org/10.18653/v1/2024.findings-acl.212 Chain-of-verification reduces hallucination in large language models . In Findings of the Association for Computational Linguistics ACL 2024, page 3563–3578. Association for Computational Linguistics

    doi:10.18653/v1/2024.findings-acl.212 2024

  10. [10]

    Gemma Team , Thomas Mesnard, Cassidy Hardin, Robert Dadashi, Surya Bhupatiraju, Shreya Pathak, Laurent Sifre, Morgane Rivière, Mihir Sanjay Kale, Juliette Love, Pouya Tafti, Léonard Hussenot, Pier Giuseppe Sessa, Aakanksha Chowdhery, Adam Roberts, Aditya Barua, Alex Botev, Alex Castro-Ros, Ambrose Slone, and 89 others. 2024. https://doi.org/10.48550/ARXIV...

    Pith/arXiv arXiv doi:10.48550/arxiv.2403.08295 2024

  11. [11]

    Ho, Christopher R\' e , Adam Chilton, Aditya Narayana, Alex Chohlas-Wood, Austin Peters, Brandon Waldon, Daniel N

    Neel Guha, Julian Nyarko, Daniel E. Ho, Christopher R\' e , Adam Chilton, Aditya Narayana, Alex Chohlas-Wood, Austin Peters, Brandon Waldon, Daniel N. Rockmore, Diego Zambrano, Dmitry Talisman, Enam Hoque, Faiz Surani, Frank Fagan, Galit Sarfaty, Gregory M. Dickinson, Haggai Porat, Jason Hegland, and 21 others. 2023. https://dl.acm.org/doi/10.5555/3666122...

    arXiv 2023

  12. [12]

    Alon Jacovi and Yoav Goldberg. 2020. https://doi.org/10.48550/ARXIV.2004.03685 Towards faithfully interpretable nlp systems: How should we define and evaluate faithfulness? arXiv preprint

    Pith/arXiv arXiv doi:10.48550/arxiv.2004.03685 2020

  13. [13]

    Qiao Jin, Bhuwan Dhingra, Zhengping Liu, William Cohen, and Xinghua Lu. 2019. https://doi.org/10.18653/v1/d19-1259 Pubmedqa: A dataset for biomedical research question answering . In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP...

    doi:10.18653/v1/d19-1259 2019

  14. [14]

    Woosuk Kwon, Zhuohan Li, Siyuan Zhuang, Ying Sheng, Lianmin Zheng, Cody Hao Yu, Joseph Gonzalez, Hao Zhang, and Ion Stoica. 2023. https://doi.org/10.1145/3600006.3613165 Efficient memory management for large language model serving with pagedattention . In Proceedings of the 29th Symposium on Operating Systems Principles, SOSP ’23, page 611–626. ACM

    arXiv 2023

  15. [15]

    Leslie Lamport. 2012. https://doi.org/10.1007/s11784-012-0071-6 How to write a 21st century proof . Journal of Fixed Point Theory and Applications, 11(1):43–63

    work page doi:10.1007/s11784-012-0071-6 2012

  16. [16]

    Tamera Lanham, Anna Chen, Ansh Radhakrishnan, Benoit Steiner, Carson Denison, Danny Hernandez, Dustin Li, Esin Durmus, Evan Hubinger, Jackson Kernion, Kamilė Lukošiūtė, Karina Nguyen, Newton Cheng, Nicholas Joseph, Nicholas Schiefer, Oliver Rausch, Robin Larson, Sam McCandlish, Sandipan Kundu, and 11 others. 2023. https://doi.org/10.48550/ARXIV.2307.13702...

    Pith/arXiv arXiv doi:10.48550/arxiv.2307.13702 2023

  17. [17]

    u ttler, Mike Lewis, Wen-tau Yih, Tim Rockt\

    Patrick Lewis, Ethan Perez, Aleksandra Piktus, Fabio Petroni, Vladimir Karpukhin, Naman Goyal, Heinrich K\" u ttler, Mike Lewis, Wen-tau Yih, Tim Rockt\" a schel, Sebastian Riedel, and Douwe Kiela. 2020. https://dl.acm.org/doi/abs/10.5555/3495724.3496517 Retrieval-augmented generation for knowledge-intensive nlp tasks . In Proceedings of the 34th Internat...

    arXiv 2020

  18. [18]

    Lundberg and Su-In Lee

    Scott M. Lundberg and Su-In Lee. 2017. https://dl.acm.org/doi/10.5555/3295222.3295230 A unified approach to interpreting model predictions . In Proceedings of the 31st International Conference on Neural Information Processing Systems, NIPS'17, page 4768–4777, Red Hook, NY, USA. Curran Associates Inc

    arXiv 2017

  19. [19]

    Qing Lyu, Shreya Havaldar, Adam Stein, Li Zhang, Delip Rao, Eric Wong, Marianna Apidianaki, and Chris Callison-Burch. 2023. https://doi.org/10.18653/v1/2023.ijcnlp-main.20 Faithful chain-of-thought reasoning . In Proceedings of the 13th International Joint Conference on Natural Language Processing and the 3rd Conference of the Asia-Pacific Chapter of the ...

    doi:10.18653/v1/2023.ijcnlp-main.20 2023

  20. [20]

    Pekka Malo, Ankur Sinha, Pekka Korhonen, Jyrki Wallenius, and Pyry Takala. 2013. https://doi.org/10.1002/asi.23062 Good debt or bad debt: Detecting semantic orientations in economic texts . Journal of the Association for Information Science and Technology, 65(4):782–796

    doi:10.1002/asi.23062 2013

  21. [21]

    Jean Mangan, Brittany Blanchard, Gabrielle Gravel, Chase Lyndale, and Connely Doizé. 2022. https://human.libretexts.org/Bookshelves/Composition/Specialized_Composition/Legal_Writing_Manual_(Mangan)/01 In Legal Writing Manual. University System of Georgia via Affordable Learning Georgia. LibreTexts

    2022

  22. [22]

    Mistral AI . 2025. https://mistral.ai/news/mistral-3 Introducing mistral 3 . Mistral AI Technical Blog

    2025

  23. [23]

    why should i trust you?

    Marco Tulio Ribeiro, Sameer Singh, and Carlos Guestrin. 2016. https://doi.org/10.1145/2939672.2939778 "why should i trust you?": Explaining the predictions of any classifier . In Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, KDD '16, page 1135–1144, New York, NY, USA. Association for Computing Machinery

    arXiv 2016

  24. [24]

    Kabir Sehgal. 2016. https://hbr.org/2016/11/how-to-write-email-with-military-precision How to write email with military precision . Harvard Business Review

    2016

  25. [25]

    Miles Turpin, Julian Michael, Ethan Perez, and Samuel R. Bowman. 2023. https://dl.acm.org/doi/10.5555/3666122.3669397 Language models don't always say what they think: unfaithful explanations in chain-of-thought prompting . In Proceedings of the 37th International Conference on Neural Information Processing Systems, NIPS '23, Red Hook, NY, USA. Curran Ass...

    arXiv 2023

  26. [26]

    Xuezhi Wang, Jason Wei, Dale Schuurmans, Quoc Le, Ed Chi, Sharan Narang, Aakanksha Chowdhery, and Denny Zhou. 2022. https://doi.org/10.48550/ARXIV.2203.11171 Self-consistency improves chain of thought reasoning in language models . arXiv preprint

    Pith/arXiv arXiv doi:10.48550/arxiv.2203.11171 2022

  27. [27]

    Chi, Quoc V

    Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Brian Ichter, Fei Xia, Ed H. Chi, Quoc V. Le, and Denny Zhou. 2022. https://dl.acm.org/doi/10.5555/3600270.3602070 Chain-of-thought prompting elicits reasoning in large language models . In Proceedings of the 36th International Conference on Neural Information Processing Systems, NIPS '22, Red Hook, ...

    arXiv 2022

  28. [28]

    Sarah Wiegreffe and Ana Marasović. 2021. https://doi.org/10.48550/ARXIV.2102.12060 Teach me to explain: A review of datasets for explainable natural language processing . arXiv preprint

    Pith/arXiv arXiv doi:10.48550/arxiv.2102.12060 2021

  29. [29]

    Griffiths, Yuan Cao, and Karthik Narasimhan

    Shunyu Yao, Dian Yu, Jeffrey Zhao, Izhak Shafran, Thomas L. Griffiths, Yuan Cao, and Karthik Narasimhan. 2023. https://dl.acm.org/doi/abs/10.5555/3666122.3666639 Tree of thoughts: deliberate problem solving with large language models . In Proceedings of the 37th International Conference on Neural Information Processing Systems, NIPS '23, Red Hook, NY, USA...

    arXiv 2023