Pith. sign in

REVIEW 3 major objections 4 minor 23 references

High-quality rubrics for evaluating financial deep-research reports can be generated and executed without human experts in the loop, and the resulting 2,600 consensus-derived gold rubrics separate ten systems by 36 percentage points in pass

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

2026-08-02 06:36 UTC pith:EHGBDASU

load-bearing objection Genuinely novel expert-free rubric pipeline with honest validation, but the gold rubrics may encode report style rather than quality; referee-worthy with conditions. the 3 major comments →

arxiv 2607.12252 v2 pith:EHGBDASU submitted 2026-07-14 cs.CL

FinResearchBench II: A Deep Research Benchmark with Consensus-Derived Gold Rubrics for Distinguishing Financial Report Quality

Beidi Luan , Rui Sun , Sinuo Wang , Yan Gu , Chao Li , Zhenliang Xiong , Jing Li , Zuo Bai This is my paper
classification cs.CL
keywords deep research benchmarkfinancial report evaluationrubric-based evaluationLLM-as-a-judgeconsensus filteringreport qualitybenchmark constructionquery-specific rubrics
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.

Long-form financial research reports are now produced by AI agents, but evaluating their quality usually needs human experts to write and score detailed rubrics, which does not scale. This paper tries to establish that the human-expert step can be removed: rubrics can be auto-generated from the reports themselves, screened by a three-model judge panel, and filtered down to a set that ranks systems reliably. The authors report that on rubric items where both humans and the LLM panel are unanimous, their labels agree 98.67% of the time, which they take as evidence that LLM screening can replace expert execution at scale. Applying a strict unanimity filter and a distinguishability filter to 14,450 candidate rubrics leaves 2,600 consensus-derived gold rubrics, and these produce pass rates from 58.58% down to 22.23% across ten deep research systems. If the claim holds, benchmark-scale evaluation of financial reports becomes feasible without a human expert in the final loop.

Core claim

The paper's central claim is that high-quality rubrics for evaluating financial deep-research reports do not need to be designed or executed by human experts. Starting from 104 real user queries and 1,040 reports from ten systems, the authors prompt LLMs to synthesize 14,450 query-specific binary rubric items covering comprehensiveness, insight, instruction following, and readability. They validate a three-LLM judge panel against three human analysts on 4,052 sampled rubric–report items: when both panels are unanimous, their labels agree 98.67% of the time (Cohen's κ = 0.9733), and LLM–human agreement exceeds human–human agreement. They then screen the full candidate pool with a strict consi

What carries the argument

The central mechanism is the consensus-derived gold rubric set plus the two-filter pipeline that creates it. A consensus-derived gold rubric is a query-specific binary criterion that survives two screens: a strict consistency filter requiring unanimous yes/no labels from all three LLM judges on every report under the same query, and a distinguishability filter requiring both a majority-yes and a majority-no outcome across the evaluated systems. The first filter removes rubrics that are not reliably executable; the second removes one-sided rubrics (mostly always-yes items) that are stable but uninformative. Product-level pass rate — the fraction of the 2,600 gold rubrics a system satisfies —

Load-bearing premise

The load-bearing premise is that the rubrics auto-generated from the very reports being evaluated capture genuine report quality (comprehensiveness, insight, instruction following, readability) and not superficial regularities such as length, structure, or shared writing style, because the paper's human–LLM validation only checks whether judges can apply a given rubric consistently, not whether the rubric itself is a valid quality criterion.

What would settle it

Compute each product's pass rate on the 2,600 gold rubrics and regress it on surface features of the reports — total length, number of sections, table/figure count, and word-frequency patterns — across the ten systems. If any single surface feature or small combination accounts for most of the pass-rate variance, the gold set is encoding report-pool regularities rather than independent quality dimensions.

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

If this is right

  • Benchmark construction for long-form report evaluation no longer requires expert rubric authors: generate candidates from model outputs, screen with the two filters, and rank.
  • The 36-percentage-point spread in pass rates gives evaluation-driven system development a clear target: improving a system on the gold rubrics is a measurable, query-specific objective.
  • The stability analyses (ρ=0.988 under judge removal, average ρ=0.978 across 120 product splits) imply product tier rankings, not just overall scores, are reproducible from the rubric set.
  • Because rubrics are derived per query from reports, the pipeline transfers to other professional domains without hand-curated evaluation criteria.

Where Pith is reading between the lines

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

  • The paper defines gold operationally — reproducible and discriminative — not as semantically verified quality. A natural next test is whether LLM-derived rubric pass rates track rankings from expert-written rubrics on the same reports; if they diverge, the gold set is measuring pool regularities rather than quality.
  • Because the rubric pool is generated from the very systems it ranks, the benchmark is vulnerable to style drift: if all systems adopt similar report templates, the distinguishability filter will shed rubrics and the pass-rate spread may compress. Re-generating rubrics from fresh reports each evaluation cycle would keep the signal sharp.
  • The 98.67% agreement on jointly unanimous items suggests a cheaper hybrid design the authors do not discuss: use the LLM panel to screen all candidates, then route only non-unanimous items to human experts, keeping humans in a narrow loop rather than none at all.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 4 minor

Summary. The paper proposes FinResearchBench II, a financial deep-research benchmark with an expert-free rubric-derivation pipeline. It collects 104 real user queries and 1,040 reports from 10 commercial deep-research systems, synthesizes 14,450 query-specific candidate rubrics via LLMs, validates a three-LLM judge panel against three human experts on 4,052 sampled rubric–report items, and then applies a strict consistency filter and a distinguishability filter to obtain 2,600 'consensus-derived gold rubrics.' Product-level pass rates are computed over these rubrics, producing a tiered ranking across the 10 systems. The paper is clearly written and transparent about its operational definition of 'gold,' but its central construct-validity claim—that the resulting rubrics distinguish report quality—is not directly established.

Significance. If the construct-validity gap were closed, the pipeline would be a useful contribution to scalable evaluation of long-form financial research reports. The paper has several concrete strengths: the benchmark is grounded in real user queries; the human–LLM validation is a serious attempt to measure judge executability; the stability diagnostics (within-model rollout, leave-one-judge-out, 120 product-split analyses) are thoughtful and reported in detail; and the two-stage filtering statistics are transparent. However, the current evidence supports only the narrower claims that (a) the three-LLM panel can execute given rubrics consistently and (b) the final rubric set is stable and discriminative across the tested systems. It does not support the broader claim that the rubrics measure report quality rather than surface regularities of the report pool. Because this is load-bearing for the title and abstract claims, the paper needs additional external validation before the contribution can be accepted as stated.

major comments (3)
  1. [§3.2, §4.1.4] The candidate rubrics are synthesized from the 1,040 reports of the very systems that the benchmark then ranks, and §4.1.4 explicitly defines 'gold' as reproducibility plus informativeness under the LLM-consensus pipeline, not independent semantic correctness. The human–LLM validation in §5.2 measures whether judges can execute a rubric consistently, not whether the rubric captures an external quality dimension such as comprehensiveness, insight, instruction following, or readability. Table 3's pass rates could therefore reward surface regularities shared by the evaluated systems' outputs—length, section structure, stylistic markers, or boilerplate—rather than report quality. This is the load-bearing construct-validity gap for the paper's central claim. To close it, the authors should audit rubric content against an external standard; for example, have human financial analysts rate a hel
  2. [§4.1.3, §5.3, Table 3] The distinguishability filter retains a rubric only if it assigns at least one majority-yes and one majority-no label across the evaluated reports. Thus the final rubric set is selected to produce spread, and the 58.58%–22.23% item-level pass-rate range in Table 3 is partly mechanical. The 120-split analysis in §5.5 shows the filter is stable under product subsetting, but stability does not establish that the selected dimension is quality. The filter-stage ablation in §5.3 (Spearman ρ=1.0 but a narrower spread) reinforces that distinguishability primarily widens the gap rather than changing the ordering. To support the quality claim, the authors should compare the final rubric set against a control set of non-quality or surface-feature rubrics and demonstrate that the quality-based pass rates diverge from that control. As written, the ranking may be a self-fulfilling consequence of the f
  3. [§4.1.2, §5.2] The consistency filter is applied at the rubric level: a rubric passes only if all three LLM judges agree on every one of the 10 reports under a query. The validation, however, is item-level: 4,052 sampled rubric–report items, which is roughly 2.8% of the 144,500 rubric–report combinations. The confusion matrix in Figure 3 reports item-level unanimity, not rubric-level unanimity. Consequently, the paper does not directly validate the criterion actually used for screening. Moreover, among the 3,436 LLM-unanimous items, humans are unanimous on only 83.4%, so 16.6% of items that the automatic screening would retain are not confirmed by human unanimity; the consequences for rubric-level retention are not quantified. The authors should either validate at the rubric level, or justify the item-to-rubric aggregation, and they should describe how the validation sample was selected across queries
minor comments (4)
  1. [§4.1.4] The PassRate(m) equation sums over all query-specific gold rubrics without an explicit query-weighting term. The later distinction between item-level pass rate and query-level macro pass rate is helpful, but the equation should state that it is the unweighted item-level definition; otherwise a reader could mistake it for a query-balanced quantity.
  2. [Table 3] The table lists nine published products and describes the held-out Internal System only in the text. For completeness and reproducibility, either include all 10 systems in the table or clearly mark the held-out system in a separate row.
  3. [§5.5] The leave-one-judge-out analysis (ρ=0.988) and the 120-split analysis are reassuring for stability, but stability under judge removal and product subsetting should be explicitly framed as a reliability check, not a validity check. The manuscript does not overstate this, but a one-sentence clarification would prevent misinterpretation.
  4. [Abstract] The abstract's phrase 'without human experts in the final loop' is accurate, but the earlier wording 'without human experts' in the first sentence could be misread as no human involvement at all. Suggest clarifying that human experts are used once for validation, not for final execution.

Circularity Check

2 steps flagged

The gold rubric set is defined by separability and by rubrics synthesized from the evaluated reports; the reported separation and 'report-quality' signal are therefore partly constructed, though the exact ranking is not forced.

specific steps
  1. fitted input called prediction [Abstract; §4.1.3 Distinguishability Filter; §5.4 Consensus-Derived Gold Rubrics Evaluation]
    "a distinguishability filter, which keeps a rubric only if it assigns at least one majority-yes and at least one majority-no label across the evaluated systems. ... Using this final rubric set, we obtain clearly differentiated rankings across 10 deep research systems, with item-level pass rates ranging from 58.58% to 22.23%."

    The final gold set G is defined in §4.1.3 to contain only rubrics that separate the evaluated systems (at least one majority-yes and one majority-no). The headline result that the set yields 'clearly differentiated rankings' and a 58.58%–22.23% pass-rate spread is thus a direct consequence of the selection rule, not an independent empirical discovery. The exact ordering and spread still depend on the LLM labels, so the circularity is partial, but the paper presents the filter's own criterion as evidence that the benchmark 'can actually separate products'.

  2. self definitional [§3.2 Rubrics Construction; §4.1.4 Consensus-Derived Gold Rubrics and Product-Level Pass Rate; §6 Conclusion]
    "Here, 'gold' denotes reproducibility and informativeness under the validated consensus pipeline, rather than independent proof of semantic correctness for every item. ... This shows that the resulting rubric set is not only executable, but also genuinely discriminative for report-quality comparison."

    Candidate rubrics are synthesized by prompting LLMs on the 1,040 reports of the very systems to be ranked (§3.2), and 'gold' is explicitly defined as reproducibility/informativeness under that pipeline (§4.1.4). The conclusion that the set is 'genuinely discriminative for report-quality comparison' therefore re-labels an internal, self-referential operationalization as an externally meaningful quality construct. The human–LLM validation (§5.2) checks only whether judges execute a given rubric consistently, not whether the rubrics capture externally valid quality dimensions, so no independent anchor breaks the definitional loop.

full rationale

The human–LLM validation is genuinely independent and non-circular: three experienced financial analysts were compared with a three-LLM panel on 4,052 sampled rubric–report items, yielding 98.67% label-level agreement on jointly unanimous items (κ=0.9733) and higher LLM–human than human–human agreement. The exact product ordering (Kimi > Doubao/Gemini > ChatGPT/Qwen > ... > Perplexity) is also not forced by the distinguishability filter, which only requires each retained rubric to have at least one majority-yes and one majority-no among the ten systems; the specific pass rates and rank order depend on the LLM judgments. However, two load-bearing components of the central claim do reduce to the pipeline's own definitions. First, the advertised 'clearly differentiated rankings' / separation is partially an automatic consequence of the distinguishability filter that defines the gold set; presenting the spread as a finding is partly self-confirming. Second, because the candidate rubrics are generated from the evaluated reports and 'gold' is defined as consensus reproducibility rather than semantic correctness, the conclusion that the set measures 'report quality' is a definitional re-labeling rather than an externally validated criterion. The sensitivity analyses (leave-one-judge-out; 120 held-out splits) test stability of the ranking, not the validity of rubric content, so they do not break the circularity. The only self-citation (FinResearchBench, Sun et al. 2025, with overlapping author Zuo Bai) is related-work motivation and is not load-bearing.

Axiom & Free-Parameter Ledger

4 free parameters · 5 axioms · 0 invented entities

No physical or conceptual entities are invented. The load-bearing assumptions are methodological: the semantic validity of LLM-generated rubrics, the ground-truth status of human expert labels, the executability interpretation of LLM unanimity, and the representativeness of a single-service query sample. The two filter thresholds and the validation sample size are the main hand-chosen settings; none are fitted to data in a predictive sense.

free parameters (4)
  • Consistency filter unanimity threshold = 3/3 LLM judges agree on all 10 reports under a query
    Chosen hyperparameter of the pipeline; drives retention from 14,450 to 3,687 rubrics (25.52%). Justified as precision-oriented but not derived from any optimization or external criterion.
  • Distinguishability filter threshold = at least one majority-yes and one majority-no report per rubric
    Chosen non-triviality condition; removes 1,087 consistency-passed rubrics (1,056 always-yes, 31 always-no). Sensible but arbitrary.
  • Validation sample size = 4,052 rubric–report items
    Sampled subset used to claim LLM–human alignment; roughly 2.8% of the 144,500-item full pool. Sampling protocol is not specified beyond category and dimension coverage.
  • Stability-diagnostic rollout count = 3 repeated batched runs
    Within-model consistency and flip rates are measured over three rollouts; the reported mean consistency 0.970 and the 2–1 split structure depend on this choice.
axioms (5)
  • domain assumption LLM-generated rubrics from the report pool encode semantically meaningful report-quality dimensions rather than surface artifacts.
    Load-bearing premise of §3.2/§4.1. Never externally audited; human validation in §5.2 tests rubric execution, not rubric content validity.
  • domain assumption Three human experts' judgments are ground truth for rubric execution on the validation subset.
    Used in §5.2 to anchor the LLM judge panel; assumes expert agreement is itself reliable despite human–human pairwise κ = 0.622.
  • domain assumption LLM unanimity on all 10 reports indicates the rubric is objectively executable; non-unanimity indicates it is not.
    The strict consistency filter of §4.1.2. Validation shows only 83.4% of LLM-unanimous items are human-unanimous, so the premise is only partially supported.
  • domain assumption The 104 queries sampled from one deployed service (StepFun/FinStep) are representative of financial deep-research user needs.
    Ecological-validity premise of §3.1; single-service provenance and undisclosed stratification limits generalizability.
  • domain assumption Batched presentation of ~100 rubrics per report does not bias labels beyond the measured within-model stability.
    §5.1 measures within-model consistency (mean 0.970) but not the validity of batched labels versus careful single-rubric judgment.

pith-pipeline@v1.3.0-alltime-deepseek · 10544 in / 16003 out tokens · 151512 ms · 2026-08-02T06:36:14.590046+00:00 · methodology

0 comments
read the original abstract

Deep research agents are increasingly used to produce long-form financial reports, yet large-scale evaluation remains bottlenecked by the need for human experts to define and execute high-quality rubrics. We address this problem by proposing a scalable pipeline for generating high-quality rubrics without human experts in the final loop. We build a financial deep research benchmark from 104 real-world user queries and automatically synthesize 14,450 query-specific candidate rubrics from model-generated reports. To justify removing human experts from rubric execution, we compare rubric judgments from three human experts with those from a three-LLM judge panel on a sampled subset, and show that LLM-based evaluation is sufficiently consistent with human evaluation to replace it for large-scale rubric screening, including 98.67\% label-level agreement on jointly unanimous items. We then derive consensus-derived gold rubrics through two filters: a strict consistency filter, which keeps a rubric only if the three LLM judges unanimously agree on every report under the same query, and a distinguishability filter, which keeps a rubric only if it assigns at least one majority-yes and at least one majority-no label across the evaluated systems. This process retains 3,687 consistency-passed rubrics, of which 2,600 remain distinguishable and form the final set of consensus-derived gold rubrics. Using this final rubric set, we obtain clearly differentiated rankings across 10 deep research systems, with item-level pass rates ranging from 58.58\% to 22.23\%. More broadly, because the pipeline removes human-expert execution from rubric generation and evaluation, it is naturally scalable for benchmark evaluation, automatic system comparison, and future studies of evaluation-driven system improvement.

Figures

Figures reproduced from arXiv: 2607.12252 by Beidi Luan, Chao Li, Jing Li, Rui Sun, Sinuo Wang, Yan Gu, Zhenliang Xiong, Zuo Bai.

Figure 1
Figure 1. Figure 1: Pipeline for deriving consensus-derived gold rubrics from query-specific candidate rubrics and using the [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Distribution of user queries across fine-grained categories. [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Confusion matrix between human unanimity [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗

Review history (2 revisions) →

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

23 extracted references · 14 linked inside Pith

  1. [3]

    Fineval: A chi- nese financial domain knowledge evaluation bench- mark for large language models. InProceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Lin- guistics: Human Language Technologies (Volume 1: Long Papers), pages 6258–6292. X. Guo, R. Zhang, G. Lu, et al

    2025

  2. [5]

    Deer: A comprehen- sive and reliable benchmark for deep-research expert reports.arXiv e-prints, page arXiv:2512.17776. X. Ho, A. K. D. Nguyen, S. Sugawara, et al

    Pith/arXiv arXiv

  3. [6]

    Finsearch- comp: Towards a realistic, expert-level evaluation of financial search and reasoning.arXiv preprint arXiv:2509.13160. P. Lewis, E. Perez, A. Piktus, et al

    arXiv

  4. [9]

    Bizfinbench: A business-driven real-world financial benchmark for evaluating LLMs.arXiv preprint arXiv:2505.19457. C. Lv, J. Zhou, W. Zhao, et al

    Pith/arXiv arXiv

  5. [10]

    Learning query-specific rubrics from human preferences for DeepResearch report generation.arXiv preprint arXiv:2602.03619. G. Mialon, C. Fourrier, T. Wolf, et al

    arXiv

  6. [11]

    Hellobench: Eval- uating long text generation capabilities of large lan- guage models.arXiv preprint arXiv:2409.16191. S. Schmidgall, Y . Su, Z. Wang, et al

    Pith/arXiv arXiv

  7. [12]

    InFindings of the Association for Computational Linguistics: EMNLP 2025, pages 5977–6043

    Agent laboratory: Using LLM agents as research assistants. InFindings of the Association for Computational Linguistics: EMNLP 2025, pages 5977–6043. Rui Sun, Zuo Bai, Wentao Zhang, Yuxiang Zhang, Li Zhao, Shan Sun, and Zhengwen Qiu

    2025

  8. [13]

    MedResearchBench: A multi-domain benchmark for evaluating AI research agents on clinical medical research.medRxiv, page 2026.03.30.26349749. H. Trivedi, N. Balasubramanian, T. Khot, et al

    2026

  9. [14]

    FIRE-Bench: Evaluating agents on the rediscovery of scientific insights.arXiv preprint arXiv:2602.02905. J. Wei, Z. Sun, S. Papay, et al

    Pith/arXiv arXiv

  10. [15]

    Browsecomp: A simple yet challenging benchmark for browsing agents.arXiv preprint arXiv:2504.12516. 9 S. Wu, Y . Li, X. Qu, et al

    Pith/arXiv arXiv

  11. [16]

    Longeval: A comprehensive analysis of long-text generation through a plan-based paradigm.arXiv preprint arXiv:2502.19103. Q. Xie, W. Han, Z. Chen, et al

    Pith/arXiv arXiv

  12. [17]

    A comprehensive survey of deep research: Systems, methodologies, and applica- tions.arXiv preprint arXiv:2506.12594. Z. Yang, P. Qi, S. Zhang, et al

    Pith/arXiv arXiv

  13. [19]

    arXiv preprint arXiv:2601.05111

    Agent-as-a-Judge. arXiv preprint arXiv:2601.05111. L. Zeng, F. Lou, Z. Wang, et al

    arXiv

  14. [20]

    FinGAIA: A chi- nese benchmark for AI agents in real-world financial domain.arXiv preprint arXiv:2507.17186. Q. Zhang, J. Zhou, Y . Wang, et al. 2026a. RubricBench: Aligning model-generated rubrics with human stan- dards.arXiv preprint arXiv:2603.01562. X. Zhang, H. Wu, J. Guo, et al. 2026b. FIRE: A compre- hensive benchmark for financial intelligence and...

    Pith/arXiv arXiv

  15. [21]

    InProceedings of the 2025 Conference on Empirical Methods in Natural Lan- guage Processing, pages 414–431

    Deepresearcher: Scaling deep research via reinforcement learning in real-world environments. InProceedings of the 2025 Conference on Empirical Methods in Natural Lan- guage Processing, pages 414–431. J. Zhong, H. Zhang, C. Southern, et al

    2025

  16. [22]

    DRACO: a cross-domain benchmark for deep research accu- racy, completeness, and objectivity.arXiv preprint arXiv:2602.11685. Y . Zhuang, Y . Yu, K. Wang, et al

    arXiv

  17. [23]

    Yes” •If the report does not satisfy the criterion, answer “No

    Agent-as-a- judge: Evaluate agents with agents.arXiv preprint arXiv:2410.10934. 10 A Rubrics Evaluation Prompt The following prompt template is used for LLM- based report evaluation against binary rubrics: Rubrics Evaluation Prompt You are a professional financial research report reviewer. Please evaluate whether the given research report satisfies the sp...

    Pith/arXiv arXiv

  18. [2018]

    InProceedings of the 2018 Conference on Empirical Methods in Natural Language Process- ing, pages 2369–2380

    HotpotQA: A dataset for diverse, explainable multi-hop question answering. InProceedings of the 2018 Conference on Empirical Methods in Natural Language Process- ing, pages 2369–2380. S. Yao, J. Zhao, D. Yu, et al

    2018

  19. [2020]

    Retrieval- augmented generation for knowledge-intensive NLP tasks.Advances in Neural Information Processing Systems, 33:9459–9474. D. Li, B. Jiang, L. Huang, et al. 2025a. From gener- ation to judgment: Opportunities and challenges of LLM-as-a-judge. InProceedings of the 2025 Con- ference on Empirical Methods in Natural Language Processing, pages 2757–279...

    Pith/arXiv arXiv 2025

  20. [2023]

    Agent- bench: Evaluating LLMs as agents.arXiv preprint arXiv:2308.03688. G. Lu, X. Guo, R. Zhang, et al

    Pith/arXiv arXiv

  21. [2024]

    Longwriter: Un- leashing 10,000+ word generation from long context LLMs.arXiv preprint arXiv:2408.07055. X. Deng, Y . Gu, B. Zheng, et al

    Pith/arXiv arXiv

  22. [2025]

    Deepresearch bench: A comprehensive benchmark for deep re- search agents.arXiv preprint arXiv:2506.11763. X. Guo, H. Xia, Z. Liu, et al

    Pith/arXiv arXiv

  23. [2026]

    BizFinBench v2: A unified dual-mode bilingual benchmark for expert- level financial capability alignment.arXiv preprint arXiv:2601.06401. J. Han, H. Kim, C. Lee, et al

    Pith/arXiv arXiv