Pith. sign in

REVIEW 4 major objections 6 minor 36 references

From Information to Delegation: Mapping Human-AI Financial Decision Making

T0 review · 4 major / 6 minor · reviewed 2026-08-04 · deepseek-v4-flash

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

Pith's one-line read Consumers use AI to shape financial decisions, not execute them

desk verdict Useful decision-authority lens on 1.5M real financial chats, but the headline delegation-is-rare numbers rest on a synthetic-trained classifier with no validation on real execution conversations. read the letter →

arxiv 2608.02100 v1 pith:XJXWFH2B submitted 2026-08-03 cs.HC cs.LG

Iman Munire Bilal , Yingcan Carol Wang , Ajan Raj , Filippo Giovagnini , Pranav Tewari , Yuwei Zhang , Mei-Chen Zoe Liou , Qamar Zaman This is my paper
classification cs.HCcs.LG
keywords decisionauthoritybehavioralintentfinanceconversationalAILLMusagedelegationconsumeragentic
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 introduces a behavioral measurement framework that classifies human-AI financial conversations by what the user is trying to accomplish (intent) and how much decision-making authority they grant the AI (inform, shape, or act). Applied to 1.5 million real ChatGPT and Gemini interactions from 6,304 US and Indian users, the framework finds that financial services are already a substantial use case, with roughly half of users engaging in finance chats. However, delegation of financial execution is rare: Level 3 'Act' interactions account for 0.3% of US and 0.1% of Indian finance subchats, and fully delegated decision execution rounds to 0.0% in both markets. The paper argues this establishes a behavioral baseline for tracking the transition to increasingly agentic AI.

What carries the argument

The key machinery is the behavioral intent taxonomy mapped onto a decision authority ladder. Twelve intents, from Simple Retrieval to Delegated Financial Decision Execution, are each assigned a decision authority level (Inform, Shape, Act). Subchats—topically segmented conversation units—are classified using fine-tuned long-context transformer models: first a binary finance filter, then a multi-label intent classifier trained on 2.8k human-annotated chats plus 600 synthetic examples for rare high-authority intents. This two-axis coding lets the paper measure not just what topics are discussed but how much judgement is delegated to AI.

What would settle it

Find a random sample of real finance subchats, have independent human annotators label them for delegated execution intent, and compare those labels to the classifier's output; if the classifier misses a material number of true execution requests or flags many false ones, the near-zero delegation estimate is an artifact of measurement rather than a fact about behavior.

Watch

Extended reading notes

Core claim

The central discovery is a measurement framework, not a single number. By combining a 12-category behavioral intent taxonomy with a three-level decision authority scale (Level 1 Inform, Level 2 Shape, Level 3 Act), the authors characterize how much authority users offload to AI in financial matters. Applying this to real-world chat histories, they show that most finance conversations fall into Level 1 and Level 2, with users seeking information, comparisons, problem resolution, and personal analysis. Level 3 execution is negligible, and delegated financial decision execution—where the user grants the AI discretion to complete a transaction—is effectively zero. The authors conclude that curre

Load-bearing premise

The central claim depends on the intent classifier measuring rare Level 3 intents accurately, but 82–96% of the training examples for those intents are synthetic, and no validation on real delegated-execution conversations is reported.

Editorial extensions

If this is right

  • Financial services are already a mainstream use of conversational AI: about half of users in both the US and India had at least one finance-related chat in the three-month window.
  • Current LLMs act as decision-support tools: Level 1 (Inform) and Level 2 (Shape) interactions dominate, while Level 3 (Act) execution is 0.3% of US and 0.1% of Indian finance subchats.
  • Delegated financial decision execution rounds to 0.0% in both markets; the few Level 3 cases are instruction-led budgeting and tracking tasks, not autonomous decisions.
  • For financial institutions, near-term value lies in AI that augments judgement—optimization, problem resolution, personal analysis—rather than fully autonomous execution.
  • For regulators, the framework implies that the shaping of decisions, not just execution, warrants consumer-protection attention.

Reading between the lines

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

  • The near-zero Level 3 rate may reflect measurement conditions as much as user behavior: because real examples of execution intent are nearly absent in the training data, the classifier was not validated against genuine delegated-execution conversations, so the true rate could be higher or lower than reported.
  • The framework is domain-agnostic; applying the same intent-plus-authority coding to health, legal, or other high-stakes consumer decisions could reveal whether the 'inform-shape-act' distribution is specific to finance or a general pattern of early LLM adoption.
  • A testable extension is to track users over time as platforms introduce native execution features (e.g., in-chat payments); the framework predicts a gradual migration from Level 2 to Level 3 as friction drops.
  • The authors' claim that delegation is 'rare' is a statement about August–October 2025; the main long-term value of the paper may be the measurement instrument itself, since later snapshots can be scored with the same taxonomy to detect when agentic adoption accelerates.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 6 minor

Summary. The paper introduces a measurement framework that classifies human-AI financial conversations along two dimensions: behavioral intent (what users seek to accomplish) and delegated decision authority (how much authority users assign to the AI). The framework is applied to roughly 1.5 million ChatGPT and Gemini user prompts from 6,304 US and Indian users, collected over three months in 2025. The central empirical claim is that financial services are a major AI use case, but consumers overwhelmingly use AI for information retrieval and decision shaping (Levels 1 and 2), while delegation of financial execution (Level 3) is extremely rare: Level 3 subchats are reported as 0.3% in the US and 0.1% in India, with Delegated Financial Decision Execution at 0.0% in both markets (Tables 7 and 8). The authors position the framework as a behavioral baseline for tracking the transition to more agentic AI.

Significance. If the empirical estimates are trustworthy, this is a useful and timely contribution: it moves beyond topic-based analyses of human-AI conversations to measure the allocation of decision authority, and it provides a concrete baseline for a question—how fast are consumers delegating financial decisions to AI?—that is currently the subject of much speculation. The paper is also transparent in reporting the heavy use of synthetic data for rare intents in Table 4 and footnote 6, and it evaluates its classifiers on held-out sets. However, the central claim that delegation is essentially nonexistent depends critically on the validity of the intent classifier for the very intents that are rarest in real data, and the current validation does not establish that validity. As a result, the headline prevalence figures should be treated as provisional lower bounds rather than point estimates.

major comments (4)
  1. [§4.2.3, Tables 4 and 7] The headline finding—that Delegated Financial Decision Execution is 0.0% and Level 3 (Act) is 0.3%/0.1%—rests on an intent classifier for which 96.01% of the training examples for Delegated Financial Decision Execution and 82.03% for Instruction-led Financial Execution are synthetic GPT-5.5 generations (Table 4, footnote 6). The held-out test set is drawn from the same mostly-synthetic distribution, and label-specific thresholds were tuned on that test set, so the reported F1 micro of 70.6 does not measure performance on real user language. No validation on authentic Level 3 conversations is reported. Since the claim is precisely the near-absence of a behavior, under-detection of real execution requests would spuriously produce 0.0%. The footnote-6 observation that fewer than 1% of conversations had these intents at collection time is a data description, not a validation of the classifie
  2. [§4.2.3, Table 3] The mapping from each intent to a single decision-authority level is asserted rather than validated. For example, 'Financial Problem Resolution' is placed at Level 2 (Shape), but a request such as 'Cancel my subscription' or 'Stop this payment' is arguably execution-oriented; indeed, an almost identical phrasing ('Cancel my Ocado subscription') is listed as an example of Instruction-led Financial Execution at Level 3. Because Table 7 aggregates all Level 3 estimates from this fixed mapping, any ambiguity in the intent-to-DA assignment directly affects the central prevalence claim. The paper should report inter-annotator agreement on the DA mapping and preferably allow the DA level to be annotated independently of the intent label, rather than treating it as a deterministic consequence.
  3. [§5.2, Table 7] Reporting '0.0%' without raw counts or confidence intervals is inadequate for a prevalence claim about a rare event. If the true count is, say, 3 subchats out of 291.9k US finance subchats, the estimate is statistically indistinguishable from 0.0% and should be presented with a confidence interval or at least the raw numerator. Conversely, if the classifier produced zero positive predictions, that is a strong reason to suspect under-detection given the synthetic training data. Please disclose the raw counts, the number of true positives on the full dataset, and the precision/recall trade-off at the chosen threshold.
  4. [§3.2] The sample is acknowledged to deviate substantially from the general population: the Indian sample is 76.5% male and 89.4% aged 18–34, and the US sample skews younger, female, and lower-income. The paper correctly notes that active AI users differ from the general population, but it does not quantify how sensitive the delegation estimates are to sample composition. Since delegation of financial execution may correlate with age, income, and financial sophistication, the near-zero estimates could in part reflect sample selection. A simple sensitivity analysis—or at least a comparison of the Level 3 prevalence across demographic subgroups—would strengthen the external validity claim. This is not a fatal flaw, but it is relevant to the strength of the policy conclusions in Section 6.
minor comments (6)
  1. [§4.1.2] The chat-segmentation heuristics were 'manually evaluated on a small test set' but no details are given on the size of that set, the agreement, or the error rate. Given that all downstream analyses operate on subchats, a brief quantitative evaluation would improve confidence.
  2. [§4.2.2] The entity-extraction and keyword-mapping validation was performed by one team member. Please report inter-annotator agreement or a second review round, especially because the resulting keyword dictionary drives the deterministic taxonomy matching.
  3. [§4.2.3] The paper does not report the prompt template used to generate the 600 synthetic examples with GPT-5.5, nor does it state whether the synthetic examples were reviewed or filtered for realism. For reproducibility and external scrutiny, please include the prompt and sample generated examples in an appendix.
  4. [§4.2.4] The Gini-index thresholding used to remove '9 highly skewed topics' is not described in enough detail for replication. Which threshold was used, and how sensitive are the topic-prevalence results to that choice?
  5. [§5.2] Because the setup is multi-label, Table 7 percentages sum to more than 100%. The text does not explicitly clarify that a single subchat can be counted under multiple DA levels; a brief note would prevent misinterpretation.
  6. [References] Reference [21] is incomplete ('... Pak. 2026') and would need full author and title information before publication.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the paper's claim is an empirical measurement, not a derivation; synthetic training data is a validity concern, not circularity.

full rationale

The paper introduces a behavioural taxonomy and uses a fine-tuned intent classifier to measure prevalence. There is no mathematical derivation that reduces to its inputs. The intent classifier is trained on a dataset that includes synthetic examples for rare Level 3 intents (Table 4: 96.01% synthetic for Delegated Financial Decision Execution, 82.03% for Instruction-led Financial Execution), and the held-out test set is drawn from the same annotated distribution, so the reported F1 (micro 70.6, macro 72.3) does not validate real-world sensitivity to Level 3 language. However, the reported prevalence (0.0-0.3%) is not equal to the training distribution (e.g., Delegated = 10.34% of training labels), so the measurement is not forced by construction. The paper contains no load-bearing self-citations, no imported uniqueness theorems, no ansatz smuggled via citation, and no renamed known result. The synthetic-data issue is a correctness/external-validity risk, not circularity, and should be handled in a validity review.

Assumptions & free parameters 4 free parameters · 5 assumptions · 0 invented entities

The central claim leans on five unverified inputs: sample representativeness, synthetic training data for rare intents, the intent-to-DA mapping, translation fidelity, and the completeness of the finance taxonomy. The most consequential is the synthetic training data, because it directly underwrites the headline 'delegation is rare' number.

free parameters (4)
  • Label-specific probability thresholds (intent classifier) = Not reported; tuned to increase F1 micro from 68.0 to 70.6
    Tuned on the test set to compensate for intent imbalance; these thresholds directly determine which subchats are assigned to each intent, including the rare Level 3 intents.
  • Synthetic augmentation ratio per intent = 96.01% for Delegated Financial Decision Execution; 82.03% for Instruction-led Fin. Execution; 82.76% for Financial Autom
    Chosen by hand to compensate for under 1% real occurrences of higher-agentic intents; the final classifier's ability to detect these intents is largely a function of synthetic data.
  • Chat segmentation thresholds = Session length >10 prompts to segment; short prompt <3 words; low-frequency term overlap
    Hand-set heuristics that change subchat granularity and therefore the denominator of all prevalence percentages.
  • Gini index threshold for topic removal = 9 topics removed
    Hand-chosen threshold to remove highly skewed topics; not justified with a formal criterion.
assumptions (5)
  • domain assumption The MeasureProtocol opt-in reimbursed sample is representative enough of active conversational AI users to support population-level prevalence claims.
    Section 3.2 benchmarks against Census only to note deviations; active AI users are assumed demographically distinct, but the sample of 6,304 users is not drawn from a defined sampling frame.
  • ad hoc to paper Synthetic GPT-5.5-generated subchats are a valid substitute for real-world examples of high-delegation intents.
    Section 4.2.3 and footnote 6: the training set for Level 3 intents is 82-96% synthetic; no evidence is provided that synthetic text matches the style, phrasing, or context of real user delegation attempts.
  • ad hoc to paper Each intent maps unambiguously to one decision authority level.
    Table 3 assigns DA levels by fiat; e.g., 'Product & Strategy Optimisation' is Level 2 even though asking 'Best mortgage for me?' could be interpreted as information retrieval. No inter-rater validation of the DA mapping is reported.
  • domain assumption Google Translate preserves enough semantic and behavioural signal for classification after translation of 144k non-English messages.
    Section 4.1.1: translation is applied before all downstream classification; translation errors would propagate into finance detection and intent tagging.
  • domain assumption The MECE finance taxonomy and keyword mapping cover all finance-relevant content, with unmatched keywords being uninformative.
    Section 4.2.2: validation of the taxonomy is by one team member reviewing frequent keywords only; 163k unique unmatched keywords are dismissed without systematic review.

how reviews work

0 comments
Cite this review

Pith. "Pith review of From Information to Delegation: Mapping Human-AI Financial Decision Making." pith.science (2026). https://pith.science/paper/XJXWFH2B

@misc{pith2026260802100,
  author       = {Pith},
  title        = {Pith review of: From Information to Delegation: Mapping Human-AI Financial Decision Making},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/XJXWFH2B}},
  note         = {Machine review of arXiv:2608.02100}
}
read the original abstract

As AI increasingly participates in human decision making, understanding how decision-making authority is distributed between humans and AI has become a fundamental behavioural question. We introduce a behavioural measurement framework combining intent and delegated decision authority to quantify what consumers seek from AI and how much decision-making authority they assign to it. Applied to 1.5 million real-world ChatGPT and Gemini interactions from 6,304 users in the United States and India, we find that financial services are already a substantial AI use case. Consumers overwhelmingly use AI to retrieve information and shape financial judgement, while delegation of financial execution remains rare. By shifting attention from conversation topics to delegated decision authority, this work establishes a behavioural baseline for measuring the transition to increasingly agentic AI.

Figures

Figures reproduced from arXiv: 2608.02100 by the authors.

Figure 1
Figure 1. Example of conversation classified by our behav [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

36 extracted references · 3 canonical work pages

  1. [1]

    Peters, and Arman Cohan

    Iz Beltagy, Matthew E. Peters, and Arman Cohan. 2020. Longformer: The Long- Document Transformer.arXiv:2004.05150(2020)

    arXiv 2020

  2. [2]

    Tadeusz Caliński and Jerzy Harabasz. 1974. A Dendrite Method for Cluster Analy- sis.Communications in Statistics3, 1 (1974), 1–27. doi:10.1080/03610927408827101

    doi:10.1080/03610927408827101 1974

  3. [3]

    Yingxia Cao, Fengmei Gong, and Tong Zeng. 2020. Antecedents and Conse- quences of Using Social Media for Personal Finance.Journal of Financial Coun- seling and Planning31 (03 2020), JFCP–18. doi:10.1891/JFCP-18-00049

    work page doi:10.1891/jfcp-18-00049 2020

  4. [4]

    Iñigo Casanueva, Pavel Temčinas, Daniela Gerz, Matthew Henderson, and Ivan Vulić. 2020. Efficient Intent Detection with Dual Sentence Encoders. InProceedings of the 2nd Workshop on NLP for Conversational AI. 38–45

    2020

  5. [5]

    Deming, Zoë Hitzig, Christopher Ong, Carl Yan Shan, and Kevin Wadman

    Aaron Chatterji, Tom Cunningham, David J. Deming, Zoë Hitzig, Christopher Ong, Carl Yan Shan, and Kevin Wadman. 2025.How People Use ChatGPT. NBER Working Paper 34255. National Bureau of Economic Research. https://www. nber.org/papers/w34255

    2025

  6. [6]

    Taha Choukhmane et al. 2025. AI Financial Advice: Supply, Demand, and Life- Cycle Effects. (2025). Working paper

    2025

  7. [7]

    Beatriz Costa-Gomes, Sophia Chen, Connie Hsueh, Deborah Morgan, Philipp Schoenegger, Yash Shah, Samuel Way, Yuki Zhu, Timothé Adeline, Michael Bhaskar, Mustafa Suleyman, and Seth Spielman. 2025. It’s About Time: The Temporal and Modal Dynamics of Copilot Usage.arXiv preprint arXiv:2512.11879 (2025). https://arxiv.org/abs/2512.11879

    arXiv 2025

  8. [8]

    Nour, Seth Spielman, Samuel F

    Beatriz Costa-Gomes, Pavel Tolmachev, Eloise Taysom, Viknesh Sounderajah, Hannah Richardson, Philipp Schoenegger, Xiaoxuan Liu, Matthew M. Nour, Seth Spielman, Samuel F. Way, Yash Shah, Michael Bhaskar, Harsha Nori, Christopher Kelly, Peter Hames, Bay Gross, Mustafa Suleyman, and Dominic King. 2026. Public use of a generalist LLM chatbot for health querie...

    doi:10.1038/s44360-026-00117-x 2026

Show all 36 references
  1. [9]

    Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human...

    2019

  2. [10]

    Google. [n. d.]. AI & Economy Research Program. https://ai.google/economy/. Accessed: 2026-07-27

    2026

  3. [11]

    Maarten Grootendorst. 2022. BERTopic: Neural topic modeling with a class-based TF-IDF procedure.arXiv preprint arXiv:2203.05794(2022)

    2022 arXiv

  4. [12]

    Anmol Handa et al. 2025. The Anthropic Economic Index. Anthropic Research Report. https://www.anthropic.com/news/the-anthropic-economic-index

    2025

  5. [13]

    Kai He, Rui Mao, Qika Lin, Yucheng Ruan, Xiang Lan, Mengling Feng, and Erik Cambria. 2023. A Survey of Large Language Models for Healthcare: From Data, Technology, and Applications to Accountability and Ethics.arXiv preprint arXiv:2310.05694(2023). arXiv:2310.05694 [cs.CL]

    2023 arXiv

  6. [14]

    Pratik Joshi, Sebastin Santy, Amar Budhiraja, Kalika Bali, and Monojit Choudhury

  7. [15]

    Ivar Krumpal. 2013. Determinants of Social Desirability Bias in Sensitive Surveys: A Literature Review.Quality & Quantity47, 4 (2013), 2025–2047. doi:10.1007/ s11135-011-9640-9

    2013

  8. [16]

    Lloyds Banking Group. 2025. Consumer Digital Index 2025. https: //www.lloydsbankinggroup.com/assets/pdfs/media/consumer-digital- index/2025/2025-consumer-digital-index.pdf

    2025

  9. [17]

    Yi-Te Lu and Yintong Huo. 2025. Financial Named Entity Recognition: How Far Can LLM Go?. InProceedings of the Joint Workshop of the 9th Financial Technology and Natural Language Processing (FinNLP), the 6th Financial Narrative Processing (FNP), and the 1st Workshop on Large La...

    2025

  10. [18]

    Niklas Muennighoff, Nouamane Tazi, Loic Magne, and Nils Reimers. 2023. MTEB: Massive Text Embedding Benchmark. InProceedings of the 17th Conference of the European Chapter of the Association for Computational Linguistics, Andreas Vlachos and Isabelle Augenstein (Eds.). Associa...

    2023 doi

  11. [19]

    Wei Nie, Yuhang Li, Jiyao Wang, Yixuan Wang, Zhen Zhang, and Shuai Lu

  12. [20]

    Paweł Niszczota and Sami Abbas. 2023. GPT Has Become Financially Literate: Insights from Financial Literacy Tests of GPT and a Preliminary Test of How People Use It as a Source of Advice.arXiv preprint arXiv:2309.00649(2023)

    2023 arXiv

  13. [21]

    Pak. 2026. How Individuals Use Generative AI for Personal Financial Manage- ment.International Journal of Information Management Data Insights(2026)

    2026

  14. [22]

    How Did We Get Here?

    Albert Park, Andrea L. Hartzler, Jina Huh, Gary Hsieh, David W. McDonald, and Wanda Pratt. 2016. “How Did We Get Here?”: Topic Drift in Online Health Discussions.Journal of Medical Internet Research18, 11 (2016), e284. doi:10.2196/ jmir.6297

    2016

  15. [23]

    Paydarzarnaghi et al

    M. Paydarzarnaghi et al. 2026. What Do People Ask AI About Finance? Evidence from ChatGPT.SSRN Electronic Journal(2026). https://papers.ssrn.com/sol3/ papers.cfm?abstract_id=6999350 From Information to Delegation: Mapping Human-AI Financial Decision Making Preprint, August, 2026

    2026

  16. [24]

    Michael Röder, Andreas Both, and Alexander Hinneburg. 2015. Exploring the Space of Topic Coherence Measures. InProceedings of the Eighth ACM Interna- tional Conference on Web Search and Data Mining. 399–408. doi:10.1145/2684822. 2685324

    2015 doi

  17. [25]

    Rousseeuw

    Peter J. Rousseeuw. 1987. Silhouettes: A Graphical Aid to the Interpretation and Validation of Cluster Analysis.J. Comput. Appl. Math.20 (1987), 53–65. doi:10.1016/0377-0427(87)90125-7

    1987 doi

  18. [26]

    G. V. Roy Chowdhury and Kiran Garimella. 2026. How People Use ChatGPT: Conversation-Level Evidence from India, Nigeria, Brazil and Pakistan. (2026). https://gvrkiran.github.io/content/How_people_use_ChatGPT.pdf Preprint

    2026

  19. [27]

    Chirag Shah, Ryen White, Reid Andersen, Georg Buscher, Scott Counts, Sarkar Das, Ali Montazer, Sathish Manivannan, Jennifer Neville, Nagu Rangan, Tara Safavi, Siddharth Suri, Mengting Wan, Leijie Wang, and Longqi Yang. 2025. Using Large Language Models to Generate, Validate, a...

    2025 doi

  20. [28]

    Renee Shelby, Fernando Diaz, and Vinodkumar Prabhakaran. 2025. Taxonomy of User Needs and Actions. arXiv:2510.06124 [cs.HC] https://arxiv.org/abs/2510. 06124

    2025

  21. [29]

    Akhil Theerthala. 2025. Synthesizing Behaviorally-Grounded Reasoning Chains: A Data-Generation Framework for Personal Finance LLMs. InProceedings of The 10th Workshop on Financial Technology and Natural Language Processing, Chung-Chi Chen, Genta Indra Winata, Stephen Rawls, An...

    2025 doi

  22. [30]

    Gomez, Łukasz Kaiser, and Illia Polosukhin

    Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Łukasz Kaiser, and Illia Polosukhin. 2017. Attention is all you need(NIPS’17). Curran Associates Inc., Red Hook, NY, USA, 6000–6010

    2017

  23. [31]

    Sonja Warkulat and Matthias Pelster. 2024. Social media attention and retail in- vestor behavior: Evidence from r/wallstreetbets.International Review of Financial Analysis96 (2024), 103721. doi:10.1016/j.irfa.2024.103721

    2024

  24. [32]

    Manzil Zaheer, Guru Guruganesh, Avinava Dubey, Joshua Ainslie, Chris Alberti, Santiago Ontanon, Philip Pham, Anirudh Ravula, Qifan Wang, Li Yang, and Amr Ahmed. 2020. Big bird: transformers for longer sequences(NIPS ’20). Curran Associates Inc., Red Hook, NY, USA, Article 1450...

    2020

  25. [33]

    Yanzhao Zhang, Mingxin Li, Dingkun Long, Xin Zhang, Huan Lin, Baosong Yang, Pengjun Xie, An Yang, Dayiheng Liu, Junyang Lin, Fei Huang, and Jingren Zhou

  26. [2020]

    InProceedings of the 58th Annual Meeting of the Association for Computational Linguistics

    The State and Fate of Linguistic Diversity and Inclusion in the NLP World. InProceedings of the 58th Annual Meeting of the Association for Computational Linguistics. Association for Computational Linguistics, 6282–6293

  27. [2024]

    arXiv:2406.11903 [cs.CL]

    A Survey of Large Language Models for Financial Applications: Progress, Challenges, and Future Directions.arXiv preprint arXiv:2406.11903(2024). arXiv:2406.11903 [cs.CL]

    2024 arXiv

  28. [2025]

    Qwen3 Embedding: Advancing Text Embedding and Reranking Through Foundation Models.arXiv preprint arXiv:2506.05176(2025)

    2025 arXiv

Pith tools

Reviewed August 4, 2026 · model on record in the stance chip above.