Pith. sign in

REVIEW 2 major objections 4 minor 31 references

Modern LLMs do not yet reason structurally over financial statements: they depend on in-context formulas and collapse into shallow heuristics when generating multi-metric tables.

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 00:44 UTC pith:KYVVAXVM

load-bearing objection A genuinely useful financial-reasoning benchmark whose hint-collapse finding holds up, but the 'structural bottleneck' mechanism is not proven by the current ablation. the 2 major comments →

arxiv 2607.28661 v1 pith:KYVVAXVM submitted 2026-07-22 cs.CL

Are the Financial Reasoning from LLMs Credible? A Real World Test over Long-Horizon Statements

Xinke Tong , Xuanming Zhang , Tianyi Tang , An Yang , Jiatu Hu , Guojie Lin , Zhenzhen Shi , Lingfeng Zeng
show 5 more authors
Boyu Yang Bing Zhao Hu Wei Lin Qu Dayiheng Liu
This is my paper
classification cs.CL
keywords FinIndicesfinancial reasoningLLM evaluation benchmarktabular reasoningtemporal de-cumulationcaliber alignmentstructured output generationsupervised fine-tuning
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.

FinIndices is a benchmark built from uncropped Chinese financial statements (up to 32K tokens) with two task formats: single-value index computation and multi-index, multi-period table generation. The paper's central claim is that modern LLMs mostly pattern-match rather than reason: when explicit formula hints are removed, accuracy collapses across all tested models, with the strongest model falling from about 71% to 38% on table tasks. A second claim is that the act of generating a structured table itself consumes reasoning capacity: the same indicators that models compute flawlessly in isolation get corrupted inside a table, producing temporal misalignment (grabbing an adjacent column) and aggregation shortcuts (lazy arithmetic instead of accounting adjustments). A supervised fine-tuning pass on distilled reasoning trajectories recovers part of the zero-hint gap (+8.54 points on single, +3.82 on table tasks) without hurting general financial QA. The paper matters because autonomous financial agents will be trusted only if their intermediate tables are auditable, and these results say that trust is not yet warranted.

Core claim

The paper introduces a benchmark and evaluates many models. Its central discovery is two bottlenecks. Knowledge Bottleneck: even though models have memorized financial formulas in pretraining, they cannot reliably reconstruct them; removing the formula from the prompt causes systemic collapse, especially on temporal de-cumulation (e.g., deriving a standalone quarter from year-to-date figures) and stock-flow caliber alignment (matching point-in-time balances with period flows). Structural Bottleneck: generating a multi-metric, multi-period table is not just a formatting problem; the structured-output constraint drains reasoning, so models that solve the same indicator in isolation substitute

What carries the argument

FinIndices itself is the central object: an automated synthesis pipeline over 829 listed companies and 384 indices that extracts uncropped financial statements, injects adversarial traps (distractor periods and omitted critical reports), and programmatically executes canonical textbook/accounting-standard formulas to produce ground truth. The two task paradigms isolate output-structure difficulty: Single-Index returns one number; Table-Index returns a matrix (average 7.75 values) in HTML/JSON/tuple format. The four capability axes — domain understanding, caliber alignment, temporal+caliber, temporal+domain — orthogonalize the reasoning skills under test. The controlled ablation (comparing th

Load-bearing premise

The structural-bottleneck conclusion assumes that the drop from single-value to table performance is caused by cognitive load rather than by the extra output-formatting, instruction complexity, and all-or-nothing scoring in table tasks; the paper's own Appendix D ablation covers only a few hand-picked indicators, not a systematic random sample, and the limitations section notes the benchmark covers only core corporate statements.

What would settle it

A matched-pair experiment in which the same indicators are queried one-by-one and inside a table, with identical required arithmetic, a pre-filled output skeleton, and per-cell scoring that ignores formatting: if per-cell accuracy does not decline in the table condition, the 'reasoning drain' claim is refuted. A second check: compare the with-hint vs without-hint gap on adversarial 'insufficient information' items against parsing errors; if most failures are malformed output rather than wrong math, the knowledge-bottleneck interpretation weakens.

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

If this is right

  • Benchmark scores obtained with explicit formulas in the prompt materially overstate models' financial capability; removing the hint is a cheap way to test for real understanding.
  • Single-number accuracy on cropped tables is not a reliable predictor of performance on realistic, multi-value tabular workflows.
  • Autonomous financial agents should not be expected to produce audit-ready multi-metric tables from raw statements without external formula lookup, verification, or fine-tuning.
  • The 'Insufficient information' test shows that reliable rejection of unanswerable queries is a separate, partially missing skill; hallucinated numbers are a distinct failure mode.
  • SFT on structured reasoning trajectories can improve zero-hint accuracy and generalizes to other financial QA benchmarks without catastrophic forgetting.

Where Pith is reading between the lines

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

  • If the structural bottleneck is real, a practical mitigation for agent pipelines is to offload table scaffolding to deterministic code (pre-filled skeletons, programmatic formatting) and let the LLM compute only cell values.
  • The knowledge bottleneck suggests test-time retrieval of formulas from a trusted source might close much of the zero-hint gap; the paper's SFT gains establish a floor, not the ceiling.
  • The same benchmark design could transfer to other high-stakes numerical domains (medical dosages, legal damages, engineering estimates) where refusing to answer from insufficient data is the correct professional behavior.
  • A testable extension: vary the number of output cells while holding total arithmetic constant; if per-cell accuracy declines monotonically with table size, the zero-sum attention mechanism is supported.

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

2 major / 4 minor

Summary. The paper introduces FININDICES, a Chinese-language benchmark for evaluating LLM data-processing fidelity on uncropped, long-context financial statements. It contains 68,307 instances spanning 829 listed companies and 384 financial indicators, organized into Single-Index value computation and Table-Index multi-metric tabulation, with adversarial traps and canonical formula hints. Ground truth is generated programmatically and expert-validated on a 640-instance test set. The main empirical findings are: (i) a 'Knowledge Bottleneck' — removing in-context formulas causes large accuracy drops across models (e.g., Gemini-3.1-Pro from 70.70% to 38.22% on Table-Index); (ii) a 'Structural Bottleneck' — models that answer isolated indicators correctly make more errors when generating multi-period tables; and (iii) SFT on Gemini-distilled reasoning trajectories improves zero-hint performance by +8.54% (Single) and +3.82% (Table) on Qwen3.5-35B-A3B. The paper concludes that current LLMs cannot be trusted for autonomous financial data processing without external formulas or targeted fine-tuning.

Significance. The benchmark fills a real gap: it goes beyond cropped tables and single numbers, requiring cross-statement, multi-period, caliber-aligned reasoning over 32K-token contexts. The strengths are substantial: formulas sourced from canonical textbooks and accounting standards; programmatic execution with expert majority-voting validation; adversarial 'insufficient information' queries; and SFT evaluation that includes external benchmarks. If the measurements hold, the knowledge-bottleneck result is an important cautionary finding for financial AI adoption. However, the paper's second headline mechanism — that generating tables 'drains reasoning capacity' — is not established by the evidence as presented. The aggregate Single-vs-Table comparison and the Appendix D ablation both fail to isolate structural load from item difficulty, output-format brittleness, and scoring artifacts. The dataset itself and the without-hint/SFT results can support a strong paper after the mechanistic claims are either re-evidenced or appropriately softened.

major comments (2)
  1. [§3, §D, Tables 15–16, Eq. (2)] The claim that table generation 'actively drains reasoning capacity' is not supported by the current evidence. The only controlled ablation (Appendix D) compares different companies and periods: Table 15 contrasts Xinlaifu's isolated Net Asset YoY with Tongdahai's in-table value, and Table 16 contrasts ST Tongde's isolated Operating Net/Total Profit with Zhenyu Tech's in-table value. These differ in company, period, source statement, and metric difficulty; the failure could be item difficulty rather than structural load. The aggregate Single vs Table comparison in Table 3 also changes output format, instruction complexity, number of required values (mean 7.75), and the all-or-nothing scoring in Eq. (2), which zeros entire outputs for a single missing or malformed value. Appendix C.2 shows that smaller models fail on formatting/parsing, so formatting failures are a plausible alternative e
  2. [§3, Abstract; Tables 3, 11, 12] The 'Knowledge Bottleneck' interpretation is stated as 'despite memorizing formulas during pre-training, models demonstrate fragile pattern matching.' The benchmark shows that without in-context formulas performance drops, but it does not test whether the model has actually memorized the formulas; the failure could equally indicate that the model never internalized the formula, that the no-hint query is under-specified, or that instruction-following differs. The paper should either add a formula-recall or formula-identification probe (e.g., ask the model to state the formula before computing) or soften the claim to 'models do not reliably infer/apply the required formulas without explicit hints.' Since the knowledge-bottleneck result is otherwise a robust and valuable observation, this should be reworded rather than discarded.
minor comments (4)
  1. [§3, Table 3] The text says the evaluation uses '13,299 FININDICES instances from the Train split,' but Table 3 reports Train, Dev, and Test columns. Please clarify how the balanced subset, development set, and 640-instance test set relate.
  2. [Abstract and §5] Phrases such as 'fatal flaws' and 'cannot be trusted for autonomous financial data processing' are stronger than the evidence supports. The paper's own limitations section acknowledges restricted data modalities; a more measured tone would match the empirical scope.
  3. [Throughout] Minor naming inconsistencies: 'single_indice'/'table_indice' versus 'Single-Index/Table-Index', and 'FININDICES' versus 'FinIndices'. Also, Table 3's header 'DevTest' needs spacing.
  4. [§2.2] For the 171 test instances with only two-expert agreement, it would be useful to report inter-annotator agreement (e.g., Cohen's kappa) or a per-instance confidence score to support the 'ultra-reliable' characterization.

Circularity Check

0 steps flagged

No significant circularity: formulas are externally sourced, ground truth is programmatic and expert-validated, and SFT gains are checked on held-out and external benchmarks.

full rationale

The paper's derivation chain is self-contained with respect to its inputs. In Section 2.1, calculation formulas are 'curated strictly from canonical financial textbooks and formal accounting standards' and initial ground-truth answers are 'derived via programmatic execution' followed by expert majority-voting validation, so model predictions are never fitted to define the benchmark's target values. The knowledge-bottleneck result is an empirical prompt-ablation (with-hint vs without-hint) over the same tasks; the structural-bottleneck result is an empirical comparison, not a relation built by definition. Appendix D's controlled ablation compares different companies (Xinlaifu vs Tongdahai; ST Tongde vs Zhenyu Tech), which is a validity/confound concern about the causal interpretation, but it is not circular: the quoted predictions are observed outputs, not quantities derived from the benchmark's construction. The SFT improvement is measured on the held-out test split and against external benchmarks (CFinBench, FinEval, FLAME, and an internal FinMath benchmark), so it does not reduce to the training objective. The only overlapping-author citation (Fin-R1) is used as an evaluated baseline model rather than as support for the central claims; therefore any self-citation is not load-bearing. No step in the paper is equivalent by construction to its outputs.

Axiom & Free-Parameter Ledger

3 free parameters · 4 axioms · 0 invented entities

The paper is an empirical benchmark, not a theory. Its load-bearing postulates are domain assumptions about Chinese accounting standards and formula canonicity, plus an interpretive assumption that accuracy drops without hints indicate missing internalized knowledge. No physical entities are invented, and no parameters are fitted to model outputs.

free parameters (3)
  • Context length cutoff = 32768 tokens
    Chosen by the authors to emulate real-world long contexts; determines which samples enter the benchmark and strongly affects measured difficulty.
  • Annualization multipliers for turnover days = 90 / 180 / 270 / 360
    Fixed by financial convention but selected by the authors; these multipliers directly enter ground-truth values for turnover and cash-cycle metrics (Appendix A.4).
  • Adversarial trap injection rate = not reported
    The pipeline randomly inserts distractors and omits necessary reports, but the paper does not report the injection probabilities; this affects benchmark difficulty and the magnitude of the observed performance collapse.
axioms (4)
  • domain assumption Chinese Accounting Standards line-item classifications (e.g., Notes Payable as non-interest-bearing) are unambiguous and canonical.
    Used in ground-truth computations for Total Invested Capital and similar metrics (Appendix A.1, A.5); if the classification is contested, the ground truth changes.
  • domain assumption Canonical textbook formulas (Revsine, CFA/CPA curricula, CAS) are the correct and only acceptable calculation paths for the benchmark's indices.
    Section E.1 states all ground-truth answers are generated by executing these formulas; alternative valid accounting treatments would make exact-match scoring wrong.
  • ad hoc to paper The performance gap without formula hints reflects lack of internalized knowledge rather than prompt ambiguity or differences in instruction-following.
    The 'Knowledge Bottleneck' interpretation assumes models should have memorized the exact formulas from pre-training; the paper does not test whether a human expert could infer the required formula from the prompt alone.
  • standard math Standard real-number arithmetic and decimal rounding as executed programmatically.
    Every ground-truth computation assumes ordinary arithmetic and rounding; no nonstandard mathematics is used.

pith-pipeline@v1.3.0-alltime-deepseek · 23842 in / 10551 out tokens · 103372 ms · 2026-08-03T00:44:18.382856+00:00 · methodology

0 comments
read the original abstract

Do Large Language Models (LLMs) possess genuine structural reasoning, or merely rely on surface-level pattern matching? The financial domain, demanding numerical precision and multi-step logic over long contexts, is an ideal testbed. Existing benchmarks fail to capture real-world industrial complexity, predominantly relying on multiple-choice questions or single-hop QA over cropped tables while ignoring intricate cross-statement dynamics and temporal de-cumulation. To bridge this gap, we introduce FinIndices, a large-scale benchmark evaluating data-processing fidelity over uncropped financial statements (up to 32K tokens). Utilizing an automated synthesis pipeline with adversarial traps, FinIndices encompasses Single-Index computation and Table-Index tabulation to test complex domain, temporal, and caliber reasoning. Our evaluation reveals two severe LLM vulnerabilities. First, a "Knowledge Bottleneck": despite memorizing formulas during pre-training, models demonstrate fragile pattern matching. Removing explicit formula hints causes performance to collapse (e.g., Gemini-3.1-Pro drops from 70.70% to 38.22% on table tasks), exposing fatal flaws in temporal de-cumulation and stock-flow caliber mismatch. Second, a "Structural Bottleneck": the intense cognitive load of generating multi-metric, multi-period tables actively drains reasoning capacity. Under structural pressure, LLMs that flawlessly execute isolated derivations regress to shallow heuristics, such as fetching incorrect adjacent columns or substituting deep accounting adjustments with lazy literal arithmetic. Finally, Supervised Fine-Tuning (SFT) yields substantial zero-hint gains (+8.54% Single, +3.82% Table), validating that structured logic can be partially restored via data-centric alignment.

Figures

Figures reproduced from arXiv: 2607.28661 by An Yang, Bing Zhao, Boyu Yang, Dayiheng Liu, Guojie Lin, Hu Wei, Jiatu Hu, Lingfeng Zeng, Lin Qu, Tianyi Tang, Xinke Tong, Xuanming Zhang, Zhenzhen Shi.

Figure 1
Figure 1. Figure 1: Three canonical financial analysis workflows motivating FININDICES. Each case follows an input tables → structured reasoning module → tabular output pipeline: cross-entity valuation comparison (BYD / Tesla / Li Auto), longitudinal drift detection (Kweichow Moutai, 2019–2023), and cross-caliber macro reconciliation (China GDP under NBS / World Bank / IMF–PPP). Together they exemplify the three core challeng… view at source ↗
Figure 2
Figure 2. Figure 2: Overview of the FININDICES data synthesis pipeline. The workflow consists of three stages: (1) Finan￾cial Statement Extraction parses and format-randomizes core tables from raw reports; (2) Context Construction & Index Alignment aligns temporal data and injects adversarial traps (e.g., missing info); and (3) Instruction & QA Pair Generation constructs prompts using canonical formulas and programmatic execu… 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

31 extracted references · 15 linked inside Pith

  1. [1]

    arXiv preprint arXiv:2511.11562 , doi =

    Akyürek, Afra Feyza and Gosai, Advait and Zhang, Chen Bo Calvin and Gupta, Vipul and Jeong, Jaehwan and Gunjal, Anisha and Rabbani, Tahseen and Mazzone, Maria and Randolph, David and Meymand, Mohammad Mahmoudi and Chattha, Gurshaan and Rodriguez, Paula and Mares, Diego and Singh, Pavit and Liu, Michael and Chawla, Subodh and Cline, Pete and Ogaz, Lucy and...

    arXiv

  2. [2]

    Bigeard, Antoine and Nashold, Langston and Krishnan, Rayan and Wu, Shirley , year =. Finance. arXiv preprint arXiv:2508.00828 , doi =

    Pith/arXiv arXiv

  3. [3]

    Proceedings of the 2021

    Chen, Zhiyu and Chen, Wenhu and Smiley, Charese and Shah, Sameena and Borova, Iana and Langdon, Dylan and Moussa, Reema and Beane, Matt and Huang, Ting-Hao and Routledge, Bryan and Wang, William Yang , year =. Proceedings of the 2021. doi:10.18653/v1/2021.emnlp-main.300 , url =

    doi:10.18653/v1/2021.emnlp-main.300 2021

  4. [4]

    arXiv preprint arXiv:2506.21591 , doi =

    Dou, Shaoyu and Shen, Yutian and Chen, Mofan and Wang, Zixuan and Xu, Jiajie and Guo, Qi and Shao, Kailai and Chen, Chao and Hu, Haixiang and Shi, Haibo and Min, Min and Zhang, Liwen , year =. arXiv preprint arXiv:2506.21591 , doi =

    arXiv

  5. [5]

    Gema, Aryo Pradipta and Leang, Joshua Ong Jun and Hong, Giwon and Devoto, Alessio and Mancino, Alberto Carlo Maria and Saxena, Rohit and He, Xuanli and Zhao, Yu and Du, Xiaotang and Madani, Mohammad Reza Ghasemi and Barale, Claire and McHardy, Robert and Harris, Joshua and Kaddour, Jean and van Krieken, Emile and Minervini, Pasquale , year =. Are. arXiv p...

    Pith/arXiv arXiv

  6. [6]

    Guo, Jiayu and Guo, Yu and Li, Martha and Tan, Songtao , journal =

  7. [7]

    arXiv preprint arXiv:2308.09975 , doi =

    Guo, Xin and Xia, Haotian and Liu, Zhaowei and Cao, Hanyang and Yang, Zhi IQ and Liu, Zhiqiang and Wang, Sizhe and Niu, Jinyi and Wang, Chuqi and Wang, Yanhui and Liang, Xiaolong and Huang, Xiaoming and Zhu, Bing and Wei, Zhongyu and Chen, Yun and Shen, Weining and Zhang, Liwen , year =. arXiv preprint arXiv:2308.09975 , doi =

    Pith/arXiv arXiv

  8. [8]

    arXiv preprint arXiv:2509.13160 , doi =

    Hu, Liang and Jiao, Jianpeng and Liu, Jiashuo and Ren, Yanle and Wen, Zhoufutu and Zhang, Kaiyuan and Zhang, Xuanliang and Gao, Xiang and He, Tianci and Hu, Fei and Liao, Yali and Wang, Zaiyuan and Yang, Chenghao and Yang, Qianyu and Yin, Mingren and Zeng, Zhiyuan and Zhang, Ge and Zhang, Xinyi and Zhao, Xiying and Zhu, Zhenwei and Namkoong, Hongseok and ...

    arXiv

  9. [9]

    Islam, Pranab and Kannappan, Anand and Kiela, Douwe and Qian, Rebecca and Scherrer, Nino and Vidgen, Bertie , journal =

  10. [10]

    Expect the

    Kamble, Kiran and Russak, Melisa and Mozolevskyi, Dmytro and Ali, Muayad and Russak, Mateusz and AlShikh, Waseem , year =. Expect the. arXiv preprint arXiv:2502.06329 , doi =

    Pith/arXiv arXiv

  11. [11]

    arXiv preprint arXiv:2506.13066 , doi =

    Lan, Kai and Zhu, Jiayong and Li, Jiangtong and Cheng, Dawei and Chen, Guang and Jiang, Changjun , year =. arXiv preprint arXiv:2506.13066 , doi =

    Pith/arXiv arXiv

  12. [12]

    arXiv preprint arXiv:2512.24314 , doi =

    Li, Shupeng and Lu, Weipeng and Liu, Linyun and Lin, Chen and Li, Shaofei and Tan, Zhendong and Zhong, Hanjun and Zeng, Yucheng and Zhu, Chenghao and Liu, Mengyue and Dong, Daxiang and Wu, Jianmin and Xiao, Yunting and Li, Annan and Liu, Danyu and Zhang, Jingnan and Liu, Licen and Yin, Dawei and Shen, Dou , year =. arXiv preprint arXiv:2512.24314 , doi =

    arXiv

  13. [13]

    Liu, Zhaowei and Guo, Xin and Yang, Zhi and Lou, Fangqi and Zeng, Lingfeng and Li, Mengping and Qi, Qi and Liu, Zhiqiang and Han, Yiyang and Cheng, Dongpo and Chen, Ronghao and Wang, Huacan and Feng, Xingdong and Wang, Huixia Judy and Shi, Chengchun and Zhang, Liwen , year =. Fin-. arXiv preprint arXiv:2503.16252 , doi =

    arXiv

  14. [14]

    arXiv preprint arXiv:2505.19457 , doi =

    Lu, Guilong and Guo, Xuntao and Zhang, Rongjunchen and Zhu, Wenqiao and Liu, Ji , year =. arXiv preprint arXiv:2505.19457 , doi =

    Pith/arXiv arXiv

  15. [15]

    arXiv preprint arXiv:2505.24714 , doi =

    Luo, Junyu and Kou, Zhizhuo and Yang, Liming and Luo, Xiao and Huang, Jinsheng and Xiao, Zhiping and Peng, Jingshu and Liu, Chengzhong and Ji, Jiaming and Liu, Xuanzhe and Han, Sirui and Zhang, Ming and Guo, Yike , year =. arXiv preprint arXiv:2505.24714 , doi =

    Pith/arXiv arXiv

  16. [16]

    Mateega, Spencer and Georgescu, Carlos and Tang, Danny , journal =

  17. [17]

    Nie, Ying and Yan, Binwei and Guo, Tianyu and Liu, Hao and Wang, Haoyu and He, Wei and Zheng, Binfan and Wang, Weihao and Li, Qiang and Sun, Weijian and Wang, Yunhe and Tao, Dacheng , journal =

  18. [18]

    Proceedings of the 63rd

    Tang, Zichen and E, Haihong and Ma, Ziyan and He, Haoyang and Liu, Jiacheng and Yang, Zhongjun and Rong, Zihua and Li, Rongjin and Ji, Kun and Huang, Qing and Hu, Xinyang and Liu, Yang and Zheng, Qianhe , year =. Proceedings of the 63rd. doi:10.18653/v1/2025.acl-long.766 , url =

    doi:10.18653/v1/2025.acl-long.766 2025

  19. [19]

    and Du, Xinrun and Yao, Yifan and others , year =

    Team, M.A.P. and Du, Xinrun and Yao, Yifan and others , year =. arXiv preprint arXiv:2502.14739 , doi =

    Pith/arXiv arXiv

  20. [20]

    arXiv preprint arXiv:2406.01574 , doi =

    Wang, Yubo and Ma, Xueguang and Zhang, Ge and Ni, Yuansheng and Chandra, Abhranil and Guo, Shiguang and Ren, Weiming and Arulraj, Aaran and He, Xuan and Jiang, Ziyan and Li, Tianle and Ku, Max and Wang, Kai and Zhuang, Alex and Fan, Rongqi and Yue, Xiang and Chen, Wenhu , year =. arXiv preprint arXiv:2406.01574 , doi =

    Pith/arXiv arXiv

  21. [21]

    arXiv preprint arXiv:2308.01430 , doi =

    Wang, Ziao and Li, Yuhang and Wu, Junda and Soon, Jaehyeon and Zhang, Xiaofeng , year =. arXiv preprint arXiv:2308.01430 , doi =

    Pith/arXiv arXiv

  22. [22]

    arXiv preprint arXiv:2505.20650 , doi =

    Wang, Yan and Qian, Lingfei and Peng, Xueqing and Ren, Yang and Wang, Keyi and Han, Yi and Feng, Dongji and Mo, Fengran and Lin, Shengyuan and Zhang, Qinchuan and He, Kaiwen and Luo, Chenri and Chen, Jianxing and Wu, Junwei and Xu, Chen and Xu, Ziyang and Huang, Jimin and Xiong, Guojun and Liu, Xiao-Yang and Xie, Qianqian and Nie, Jian-Yun , year =. arXiv...

    Pith/arXiv arXiv

  23. [23]

    Xie, Qianqian and Han, Weiguang and Chen, Zhengyu and others , journal =

  24. [24]

    arXiv preprint arXiv:2602.22273 , doi =

    Zhang, Xiyuan and Wu, Huihang and Guo, Jiayu and Zhang, Zhenlin and Zhang, Yiwei and Huo, Liangyu and Ma, Xiaoxiao and Wan, Jiansong and Jiao, Xuewei and Jing, Yi and Xie, Jian , year =. arXiv preprint arXiv:2602.22273 , doi =

    arXiv

  25. [25]

    Proceedings of the 62nd

    Zhao, Yilun and Liu, Hongjun and Long, Yitao and Zhang, Rui and Zhao, Chen and Cohan, Arman , year =. Proceedings of the 62nd. doi:10.18653/v1/2024.acl-long.693 , url =

    doi:10.18653/v1/2024.acl-long.693 2024

  26. [26]

    Agentar-

    Zheng, Yanjun and Du, Xiyang and Liao, Longfei and Zhao, Xiaoke and Zhou, Zhaowen and Song, Jingze and Zhang, Bo and Liu, Jiawei and Qi, Xiang and Li, Zhe and Zhang, Zhiqiang and Wang, Wei and Zhang, Peng , year =. Agentar-. arXiv preprint arXiv:2507.16802 , doi =

    Pith/arXiv arXiv

  27. [27]

    Zhou, Yuanchen and Jiang, Shuo and Zhu, Jie and Li, Junhui and Guo, Lifan and Chen, Feng and Zhang, Chi , year =. Fin-. arXiv preprint arXiv:2508.15202 , doi =

    Pith/arXiv arXiv

  28. [28]

    Proceedings of the 59th

    Zhu, Fengbin and Lei, Wenqiang and Huang, Youcheng and Wang, Chao and Zhang, Shuo and Lv, Jiancheng and Feng, Fuli and Chua, Tat-Seng , year =. Proceedings of the 59th. doi:10.18653/v1/2021.acl-long.254 , url =

    doi:10.18653/v1/2021.acl-long.254 2021

  29. [29]

    arXiv preprint arXiv:2504.15716 , doi =

    Zhu, Jie and Chen, Qian and Dou, Huaixia and Li, Junhui and Guo, Lifan and Chen, Feng and Zhang, Chi , year =. arXiv preprint arXiv:2504.15716 , doi =

    Pith/arXiv arXiv

  30. [30]

    and Zhang, Rongjunchen and Dai, Yong , year =

    Gan, Ziliang and Lu, Yu and Zhang, Dong and Li, Haohan and Liu, Che and Liu, Jian ... and Zhang, Rongjunchen and Dai, Yong , year =. arXiv preprint arXiv:2411.03314 , doi =

    Pith/arXiv arXiv

  31. [31]

    Preprint , url =

    Xue, Siqiao and Li, Xiaojing and Zhou, Fan and Dai, Qingyang and Chu, Zhixuan and Mei, Hongyuan , year =. Preprint , url =