Pith. sign in

REVIEW 4 major objections 4 minor 1 cited by

Reasoning or Overthinking: Evaluating Large Language Models on Financial Sentiment Analysis

T0 review · 4 major / 4 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read Prompting LLMs to reason step-by-step lowers their agreement with human financial sentiment labels; the best zero-shot configuration is GPT-4o with no chain of thought.

desk verdict Useful negative result on CoT for financial sentiment, but the hidden UNKNOWN-to-label mapping and the false monotonicity claim keep me from trusting the headline numbers. read the letter →

arxiv 2506.04574 v1 pith:DQ3LTO6K submitted 2025-06-05 cs.CL cs.AI

Dimitris Vamvourellis , Dhagash Mehta This is my paper
classification cs.CLcs.AI
keywords financialsentimentanalysiszero-shotpromptingchain-of-thoughtoverthinkingmacro-F1PhraseBankdual-processtheorypromptstructure
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

This paper asks whether prompting large language models to reason step by step improves zero-shot financial sentiment classification or makes it worse. Testing three proprietary LLMs and four prompting styles on the expert-labeled Financial PhraseBank corpus, it finds that explicit reasoning consistently lowers macro-F1 alignment with human sentiment labels, and the best zero-shot configuration is GPT-4o answering directly with no chain of thought. Reasoning-first prompts hurt most on easy, unambiguous sentences, while label-first explanation (LIRA) beats reasoning-first chain-of-thought. The authors interpret the negative correlation between completion length and accuracy as evidence of overthinking: simulated deliberation injects uncertainty and biases the model rather than sharpening its judgment. If correct, this challenges the default assumption that more reasoning is always better and gives a practical rule for financial sentiment pipelines.

What carries the argument

The argument runs on four prompting paradigms mapped to dual-process cognitive styles: No-CoT (a direct label, System 1-like), CoT-Short and CoT-Long (a label conditioned on brief or elaborate reasoning, System 2-like), and LIRA (a label produced first, then explained, called System 1 with metacognition). Performance is measured as macro-F1 on the 4,845 expert-labeled Financial PhraseBank sentences, stratified by annotator-agreement level and by Flesch-Kincaid readability quartile. Completion-token counts are then binned into five quantiles per model-method combination; the monotonic drop in macro-F1 across those bins is the machinery that carries the overthinking conclusion.

What would settle it

Stratify completion-token quantiles within each readability quartile, or cap output length and prompt the same sentences at short and long budgets, and check whether macro-F1 still falls as token count rises; if the decline disappears or reverses, the overthinking mechanism is an artifact of confounded sentence difficulty.

Watch

Extended reading notes

Core claim

The paper claims that on Financial PhraseBank, zero-shot LLM sentiment classification aligns best with expert labels when the model answers directly: GPT-4o with No-CoT prompting reaches the highest macro-F1 among zero-shot LLM configurations, while every chain-of-thought variant of GPT-4o and GPT-4.1 scores lower in every agreement stratum and every readability quartile. The reasoning-tuned model o3-mini, which continues internal reasoning even under No-CoT prompts, trails both. Across all models and prompting methods, macro-F1 falls monotonically as completion-token count rises, which the authors read as evidence of overthinking: longer simulated deliberation does not reflect deeper understanding but instead drifts away from the fast, heuristic judgments human annotators apply. A secondary result is that LIRA prompting, which commits to the label before explaining it, outperforms reasoning-first chain-of-thought, showing that the position of the reasoning tokens relative to the label is itself a consequential design choice.

Load-bearing premise

The overthinking conclusion assumes that completion-token count measures reasoning effort and that extra tokens cause the errors, but the analysis does not control for sentence difficulty within token bins, so harder sentences could drive both longer outputs and lower F1.

Editorial extensions

If this is right

  • For zero-shot financial sentiment classification on similar news text, answering directly with No-CoT prompting is the default that maximizes agreement with expert labels.
  • Adding chain-of-thought, short or long, costs accuracy across all agreement levels and all readability quartiles, with the widest gap on sentences where annotators fully agree.
  • A reasoning-tuned model such as o3-mini cannot simply be switched off: it underperforms even under No-CoT prompting, roughly matching a reasoning-prompted GPT-4.1.
  • Label-first prompting (LIRA) outperforms reasoning-first chain-of-thought, so where the label sits in the generated sequence is a real design decision.
  • Longer model completions are associated with lower macro-F1 in every model-method combination, consistent with overthinking rather than deeper understanding.

Reading between the lines

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

  • A testable extension is to apply the same four prompting paradigms to other subjective annotation tasks, such as review helpfulness, hate-speech severity, or medical triage notes, and check whether direct No-CoT classification systematically beats chain-of-thought whenever human agreement is the gold standard.
  • The token-count story could be confounded by output-position effects: LIRA places the label before the explanation, so its advantage may come from autoregressive position rather than from intuition; an ablation that randomizes label position would separate the two.
  • If these findings generalize, financial NLP pipelines should treat reasoning-prompt choice as a hyperparameter validated on human-labeled data, and completion length as a diagnostic signal to monitor rather than a proxy for output quality.
  • The o3-mini tendency to over-label neutral headlines as positive suggests that reasoning-optimized training may encode a directional prior; probing the model with label-balanced prompts or calibration tests could reveal whether the bias sits in the reasoning trace or in the final-label head.
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 / 4 minor

Summary. The paper investigates zero-shot financial sentiment classification on the Financial PhraseBank dataset, comparing three proprietary LLMs (GPT-4o, GPT-4.1, o3-mini) under four prompting strategies (No-CoT, CoT-Short, CoT-Long, LIRA) and two fine-tuned FinBERT baselines. The central claim is that explicit or implicit reasoning does not improve alignment with human sentiment labels: GPT-4o with No-CoT prompting achieves the highest macro-F1 among LLMs, CoT variants consistently underperform No-CoT for GPT-4o and GPT-4.1, and o3-mini, a reasoning-optimized model, trails behind. The paper further argues that longer completion-token counts are associated with lower macro-F1, interpreting this as evidence of 'overthinking', and it analyzes performance across annotator-agreement strata, readability quartiles, and confusion matrices.

Significance. If the headline result holds, the paper provides a practically useful and counterintuitive guideline: for zero-shot financial sentiment classification, direct No-CoT prompting can outperform chain-of-thought prompting, and reasoning-optimized models may be at a disadvantage. The study uses an external, expert-annotated benchmark, explicitly disclaims the System 1/System 2 analogy as heuristic, and reports detailed confusion matrices and qualitative failure cases, which are useful contributions. However, the central ranking is currently fragile because of a missing description of how the prompts' YES/NO/UNKNOWN responses are mapped to the three sentiment labels, the absence of any statistical uncertainty quantification, and an over-reaching causal interpretation of the token-length analysis.

major comments (4)
  1. [Appendix A / §4.2] All four prompt templates instruct the model to answer 'YES', 'NO', or 'UNKNOWN', yet the paper never states how these values are converted to the Financial PhraseBank labels positive, neutral, and negative. The confusion matrices in Tables 6–11 and 13–20 contain only three predicted classes, so every 'UNKNOWN' response must have been mapped somewhere, but the mapping rule and the frequency of 'UNKNOWN' per model-method are not reported. If 'UNKNOWN' is mapped to 'neutral', the evaluation conflates genuine neutral sentiment with model hedging, and the relative frequency of 'UNKNOWN' could differ substantially across prompting strategies and models, potentially changing the No-CoT versus CoT ranking. The authors must report the mapping rule, the fraction of UNKNOWN responses per condition, and ideally the sensitivity of the headline comparison to alternative mappings.
  2. [§5, Table 5] The text states that 'Performance monotonically declines with increasing completion length across all models and prompting strategies', but Table 5 contradicts this. For example, GPT-4o LIRA has macro-F1 values 0.704, 0.725, 0.773, 0.749, 0.754 across Q1–Q5, and GPT-4.1 LIRA has 0.773, 0.721, 0.730, 0.700, 0.697; even GPT-4o CoT-Short is non-monotonic (0.757, 0.760, 0.726, 0.743, 0.631). At best the data support a broad negative association in most configurations, not monotonic decline. This overstatement weakens the 'overthinking' narrative and should be corrected, with the analysis reframed as an association rather than a monotonic trend.
  3. [§4.3 / §5, Table 5] The token-quantile analysis assumes that completion-token count is a faithful measure of reasoning effort and that additional tokens cause classification errors. This assumption is not tested: sentences that are harder or more ambiguous may independently elicit both longer model outputs and lower macro-F1, creating a spurious negative association. The paper does not control for sentence-level difficulty, annotator agreement, or readability within quantile bins, nor does it attempt to establish the direction of causality. The authors should either provide a per-sentence analysis with controls (e.g., a regression with sentence-level features and model-method fixed effects) or materially soften the causal 'overthinking' claim, treating the token-length result as a correlation that requires further evidence.
  4. [§5, Tables 3–5] All reported results come from a single API run per model-method combination, with no confidence intervals, bootstrap estimates, significance tests, or repeated sampling to account for LLM stochasticity. Many of the headline differences are small—for instance, several macro-F1 values in Tables 3–5 differ by 0.01–0.02—so the stability of the ranking (especially the GPT-4o No-CoT versus GPT-4o LIRA gap) is unknown. The paper's practical recommendation would be much better supported if the authors reported variance across repeated runs, or at least explicitly acknowledged the absence of uncertainty quantification as a limitation when drawing comparative conclusions.
minor comments (4)
  1. [Figure 2 caption] The caption reads 'Macro F1 scores across models and prompting strategies', but the figure itself plots completion-token distributions; the caption should be corrected to describe the token-count boxplot.
  2. [Table 12] There are transcription errors in the example table: 'o3-mini & CoT-Long' and 'o3-mini & CoT-Short' rows under Example 2 reproduce the RPK/ASK text from Example 1 rather than the net-sales sentence, and the LIRA row for Example 4 contains 'PositivePositive'. These should be fixed because the qualitative failure-mode analysis relies on the accuracy of these responses.
  3. [Appendix C, Tables 8 and 17] Table 8 (GPT-4.1 & No-CoT) and Table 17 (GPT-4.1 & CoT-Long) contain identical numerical entries, which is unlikely to be correct; one of the tables is probably misreported. The authors should check the underlying confusion matrices and correct the duplication.
  4. [§6.1] The sentence 'CoT-style prompts do not significantly bias GPT-4o in one direction compared to No-CoT methods' uses the word 'significantly' without any statistical test; either supply a test or use a non-statistical phrasing such as 'do not appear to bias'.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity; the paper is a self-contained empirical evaluation against an external human-labeled benchmark with no fitted parameters or self-citation chain.

full rationale

The paper's claims are empirical comparisons, not derivations. Macro-F1 scores are computed by comparing model outputs to the externally labeled Financial PhraseBank benchmark (Malo et al., 2014), and no parameter is fitted to the test set except the disclosed FinBERT-Prosus baseline, which the authors explicitly treat as a reference point rather than a zero-shot competitor. The prompting strategies are defined independently of the outcome metric, and the LIRA prompt is attributed to prior work by Lopez-Lira and Tang [24], not to the present authors. The 'overthinking' interpretation is an inference from token counts and F1 scores, not an equation that reduces to its inputs; while Table 5 shows some non-monotonic patterns that weaken the 'monotonically declines' wording, a flawed or overstated interpretation is not circular reasoning. The unstated mapping of 'UNKNOWN' responses to the three sentiment classes is a reporting and validity concern, not a circularity concern, because the evaluation still depends on external human labels rather than on the paper's own assumptions. No self-citation is load-bearing, no uniqueness theorem is imported, and no known result is renamed. The central comparison (No-CoT outperforms CoT for GPT-4o and GPT-4.1) is an externally falsifiable empirical result.

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

The paper introduces no fitted constants and no new physical or computational entities. It relies on several domain assumptions: the prompt label 'UNKNOWN' is scored as 'neutral'; token count is treated as reasoning depth; Flesch-Kincaid is used as complexity; human labels are the target. These are reasonable but largely unvalidated for this task.

assumptions (4)
  • domain assumption The prompt response 'UNKNOWN' is treated as the dataset label 'neutral'.
    Prompts in Appendix A ask for YES/NO/UNKNOWN, while Financial PhraseBank labels are positive/neutral/negative. No discussion of this mapping; confusion matrices score UNKNOWN against neutral.
  • domain assumption Completion token count is a proxy for reasoning effort.
    Table 5 stratifies by token quantiles and interprets low-F1 high-token bins as overthinking. Token count is not adjusted for input difficulty or formatting.
  • domain assumption Flesch-Kincaid readability score is a valid measure of linguistic complexity for financial sentences.
    Used to create complexity quartiles in Table 4. No validation that this readability metric tracks annotator-perceived ambiguity.
  • domain assumption Human annotator labels are the ground truth for sentiment.
    Section 4.4 explicitly scopes evaluation to human alignment, not market outcomes. This is standard for the benchmark.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Reasoning or Overthinking: Evaluating Large Language Models on Financial Sentiment Analysis." pith.science (2026). https://pith.science/paper/DQ3LTO6K

@misc{pith2026250604574,
  author       = {Pith},
  title        = {Pith review of: Reasoning or Overthinking: Evaluating Large Language Models on Financial Sentiment Analysis},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/DQ3LTO6K}},
  note         = {Machine review of arXiv:2506.04574}
}
read the original abstract

We investigate the effectiveness of large language models (LLMs), including reasoning-based and non-reasoning models, in performing zero-shot financial sentiment analysis. Using the Financial PhraseBank dataset annotated by domain experts, we evaluate how various LLMs and prompting strategies align with human-labeled sentiment in a financial context. We compare three proprietary LLMs (GPT-4o, GPT-4.1, o3-mini) under different prompting paradigms that simulate System 1 (fast and intuitive) or System 2 (slow and deliberate) thinking and benchmark them against two smaller models (FinBERT-Prosus, FinBERT-Tone) fine-tuned on financial sentiment analysis. Our findings suggest that reasoning, either through prompting or inherent model design, does not improve performance on this task. Surprisingly, the most accurate and human-aligned combination of model and method was GPT-4o without any Chain-of-Thought (CoT) prompting. We further explore how performance is impacted by linguistic complexity and annotation agreement levels, uncovering that reasoning may introduce overthinking, leading to suboptimal predictions. This suggests that for financial sentiment classification, fast, intuitive "System 1"-like thinking aligns more closely with human judgment compared to "System 2"-style slower, deliberative reasoning simulated by reasoning models or CoT prompting. Our results challenge the default assumption that more reasoning always leads to better LLM decisions, particularly in high-stakes financial applications.

Figures

Figures reproduced from arXiv: 2506.04574 by the authors.

Figure 1
Figure 1. Macro F1 scores across models and prompting strategies over the entire Financial Phrasebank dataset. [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Macro F1 scores across models and prompting strate [PITH_FULL_IMAGE:figures/full_fig_p008_2.png] view at source ↗

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. Task Complexity Matters: An Empirical Study of Reasoning in LLMs for Sentiment Analysis

    cs.CL 2026-02 conditional novelty 6.0 of 10

    Reasoning in LLMs helps only for complex 27-class emotion recognition and hurts simple binary sentiment, across seven model families and 504 configurations.

Reference graph

Works this paper leans on

46 extracted references · 13 canonical work pages · cited by 1 Pith paper

  1. [1]

    Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Floren- cia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. 2023. Gpt-4 technical report. arXiv preprint arXiv:2303.08774 (2023)

    arXiv 2023

  2. [2]

    Dogu Araci. 2019. Finbert: Financial sentiment analysis with pre-trained language models. arXiv preprint arXiv:1908.10063 (2019)

    arXiv 2019

  3. [3]

    Gagan Bhatia, El Moatez Billah Nagoudi, Hasan Cavusoglu, and Muhammad Abdul-Mageed. 2024. Fintral: A family of gpt-4 level multimodal financial large language models. arXiv preprint arXiv:2402.10986 (2024)

    arXiv 2024

  4. [4]

    Marcel Binz and Eric Schulz. 2023. Using cognitive psychology to understand GPT-3. Proceedings of the National Academy of Sciences120, 6 (2023), e2218523120

    work page 2023

  5. [5]

    Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. 2020. Language models are few-shot learners. Advances in neural information processing systems 33 (2020), 1877–1901

    2020

  6. [6]

    Xingyu Chen, Jiahao Xu, Tian Liang, Zhiwei He, Jianhui Pang, Dian Yu, Lin- feng Song, Qiuzhi Liu, Mengfei Zhou, Zhuosheng Zhang, et al . 2024. Do not think that much for 2+ 3=? on the overthinking of o1-like llms. arXiv preprint arXiv:2412.21187 (2024)

    arXiv 2024

  7. [7]

    Yanxi Chen, Xuchen Pan, Yaliang Li, Bolin Ding, and Jingren Zhou. 2024. A simple and provable scaling law for the test-time compute of large language models. arXiv preprint arXiv:2411.19477 (2024)

    arXiv 2024

  8. [8]

    George Fatouros, Kostas Metaxas, John Soldatos, and Dimosthenis Kyriazis. 2024. Can Large Language Models beat wall street? Evaluating GPT-4’s impact on financial decision-making with MarketSenseAI. Neural Computing and Applica- tions (2024), 1–26

    work page 2024

Show all 46 references
  1. [9]

    Guhao Feng, Bohang Zhang, Yuntian Gu, Haotian Ye, Di He, and Liwei Wang

  2. [10]

    John H. Flavell. 1979. Metacognition and cognitive monitoring: A new area of cognitive–developmental inquiry. American Psychologist 34, 10 (1979), 906–911. doi:10.1037/0003-066X.34.10.906

    1979 doi

  3. [11]

    Paul Glasserman and Caden Lin. 2023. Assessing look-ahead bias in stock return predictions generated by gpt sentiment analysis. arXiv preprint arXiv:2309.17322 (2023)

    2023 arXiv

  4. [12]

    Daya Guo, Dejian Yang, Haowei Zhang, Junxiao Song, Ruoyu Zhang, Runxin Xu, Qihao Zhu, Shirong Ma, Peiyi Wang, Xiao Bi, et al . 2025. Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning. arXiv preprint arXiv:2501.12948 (2025)

    2025 arXiv

  5. [13]

    Tingxu Han, Zhenting Wang, Chunrong Fang, Shiyu Zhao, Shiqing Ma, and Zhenyu Chen. 2024. Token-budget-aware llm reasoning. arXiv preprint arXiv:2412.18547 (2024)

    2024 arXiv

  6. [14]

    Allen H Huang, Hui Wang, and Yi Yang. 2023. FinBERT: A large language model for extracting information from financial text.Contemporary Accounting Research 40, 2 (2023), 806–841

    2023

  7. [15]

    Mingyu Jin, Qinkai Yu, Dong Shu, Haiyan Zhao, Wenyue Hua, Yanda Meng, Yongfeng Zhang, and Mengnan Du. 2024. The impact of reasoning step length on large language models. arXiv preprint arXiv:2401.04925 (2024)

    2024 arXiv

  8. [16]

    Daniel Kahneman. 2011. Thinking, fast and slow. Farrar, Straus and Giroux (2011)

    2011

  9. [17]

    Alex Kim, Maximilian Muhn, and Valeri Nikolaev. 2023. Bloated disclosures: can ChatGPT help investors process information? arXiv preprint arXiv:2306.10224 (2023)

    2023

  10. [18]

    Alex G Kim and Valeri V Nikolaev. 2024. Context-Based Interpretation of Finan- cial Information. Journal of Accounting Research (2024)

    2024

  11. [19]

    J Peter Kincaid, Robert P Fishburne Jr, Richard L Rogers, and Brad S Chissom

  12. [20]

    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 (2022), 22199–22213

    2022

  13. [21]

    Thanos Konstantinidis, Giorgos Iacovides, Mingxue Xu, Tony G Constantinides, and Danilo Mandic. 2024. Finllama: Financial sentiment classification for algo- rithmic trading applications. arXiv preprint arXiv:2403.12285 (2024)

    2024 arXiv

  14. [22]

    Andrew K Lampinen, Ishita Dasgupta, Jacob Hilton, Aida Nematzadeh, Joshua B Tenenbaum, Christopher Clark, and Samuel R Bowman. 2022. Can Language Models Learn from Explanations in Context? arXiv preprint arXiv:2204.02329 (2022). https://arxiv.org/abs/2204.02329

    2022 arXiv

  15. [23]

    Zhaowei Liu, Xin Guo, Fangqi Lou, Lingfeng Zeng, Jinyi Niu, Zixuan Wang, Jiajie Xu, Weige Cai, Ziwei Yang, Xueqian Zhao, et al. 2025. Fin-r1: A large language model for financial reasoning through reinforcement learning. arXiv preprint arXiv:2503.16252 (2025)

    2025

  16. [24]

    Alejandro Lopez-Lira and Yuehua Tang. 2023. Can chatgpt forecast stock price movements? return predictability and large language models. arXiv preprint arXiv:2304.07619 (2023)

    2023

  17. [25]

    Pekka Malo, Ankur Sinha, Pekka Korhonen, Jyrki Wallenius, and Pyry Takala

  18. [26]

    Niklas Muennighoff, Zitong Yang, Weijia Shi, Xiang Lisa Li, Li Fei-Fei, Hannaneh Hajishirzi, Luke Zettlemoyer, Percy Liang, Emmanuel Candès, and Tatsunori Hashimoto. 2025. s1: Simple test-time scaling. arXiv preprint arXiv:2501.19393 (2025)

    2025 arXiv

  19. [27]

    Ummara Mumtaz and Summaya Mumtaz. 2023. Potential of ChatGPT in predict- ing stock market trends based on Twitter Sentiment Analysis. arXiv preprint arXiv:2311.06273 (2023)

    2023 arXiv

  20. [28]

    Nelson and Louis Narens

    Thomas O. Nelson and Louis Narens. 1990. Metamemory: A theoretical frame- work and new findings. The Psychology of Learning and Motivation 26 (1990), 125–173. doi:10.1016/S0079-7421(08)60053-5

    1990 doi

  21. [29]

    Matthew Renze and Erhan Guven. 2024. Self-reflection in llm agents: Effects on problem-solving performance. arXiv preprint arXiv:2405.06682 (2024)

    2024 arXiv

  22. [30]

    Noah Shinn, Federico Cassano, Beck Labash, Ashwin Gopinath, Karthik Narasimhan, and Shunyu Yao. 2023. Reflexion: Language agents with verbal reinforcement learning, 2023. URL https://arxiv. org/abs/2303.11366 (2023)

    2023 arXiv

  23. [31]

    Charlie Snell, Jaehoon Lee, Kelvin Xu, and Aviral Kumar. 2024. Scaling llm test- time compute optimally can be more effective than scaling model parameters. arXiv preprint arXiv:2408.03314 (2024). Reasoning or Overthinking: Evaluating Large Language Models on Financial Sentime...

    2024 arXiv

  24. [32]

    KE Stanovich, RF West, and S Epstein. 2000. Individual differences in reasoning: Implications for the rationality debate?-Open Peer Commentary-The rationality debate from the perspective of cognitive-experiential self-theory. (2000)

    2000

  25. [33]

    Rick Steinert and Saskia Altmann. 2023. Linking microblogging sentiments to stock price movement: An application of GPT-4. arXiv preprint arXiv:2308.16771 (2023)

    2023 arXiv

  26. [34]

    Yang Sui, Yu-Neng Chuang, Guanchu Wang, Jiamu Zhang, Tianyi Zhang, Jiayi Yuan, Hongyi Liu, Andrew Wen, Shaochen Zhong, Hanjie Chen, et al. 2025. Stop overthinking: A survey on efficient reasoning for large language models. arXiv preprint arXiv:2503.16419 (2025)

    2025 arXiv

  27. [35]

    Jean-Francois Ton, Muhammad Faaiz Taufiq, and Yang Liu. 2024. Understand- ing Chain-of-Thought in LLMs through Information Theory. arXiv preprint arXiv:2411.11984 (2024)

    2024 arXiv

  28. [36]

    Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. 2023. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971 (2023)

    2023 arXiv

  29. [37]

    Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. 2022. Chain-of-thought prompting elicits reason- ing in large language models. Advances in neural information processing systems 35 (2022), 24824–24837

    2022

  30. [38]

    Shijie Wu, Ozan Irsoy, Steven Lu, Vadim Dabravolski, Mark Dredze, Sebas- tian Gehrmann, Prabhanjan Kambadur, David Rosenberg, and Gideon Mann

  31. [39]

    Yangzhen Wu, Zhiqing Sun, Shanda Li, Sean Welleck, and Yiming Yang. 2024. Scaling inference computation: Compute-optimal inference for problem-solving with language models. In The 4th Workshop on Mathematical Reasoning and AI at NeurIPS, Vol. 24

    2024

  32. [40]

    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 (2023), 11809–11822

    2023

  33. [41]

    Shengjia Zhang, Junjie Wu, Jiawei Chen, Changwang Zhang, Xingyu Lou, Wangchunshu Zhou, Sheng Zhou, Can Wang, and Jun Wang. 2025. OThink-R1: In- trinsic Fast/Slow Thinking Mode Switching for Over-Reasoning Mitigation.arXiv preprint arXiv:2506.02397 (2025). doi:10.48550/arXiv.25...

    2025 doi

  34. [42]

    arXiv preprint arXiv:2303.17564 (2023)

    Bloomberggpt: A large language model for finance. arXiv preprint arXiv:2303.17564 (2023)

    2023 arXiv

  35. [46]

    YES” if good news, “NO

    Denny Zhou, Nathanael Schärli, Le Hou, Jason Wei, Nathan Scales, Xuezhi Wang, Dale Schuurmans, Claire Cui, Olivier Bousquet, Quoc Le, et al. 2022. Least-to- most prompting enables complex reasoning in large language models. arXiv preprint arXiv:2205.10625 (2022). Vamvourellis ...

    2022 arXiv

  36. [1975]

    Derivation of new readability formulas (automated readability index, fog count and flesch reading ease formula) for navy enlisted personnel. (1975)

    1975

  37. [2014]

    Journal of the Association for Information Science and Technology 65, 4 (2014), 782–796

    Good debt or bad debt: Detecting semantic orientations in economic texts. Journal of the Association for Information Science and Technology 65, 4 (2014), 782–796

    2014

  38. [2023]

    Advances in Neural Information Processing Systems 36 (2023), 70757– 70798

    Towards revealing the mystery behind chain of thought: a theoretical perspective. Advances in Neural Information Processing Systems 36 (2023), 70757– 70798

    2023

Pith tools

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