Pith. sign in

REVIEW 3 major objections 5 minor 2 cited by

Agentar-Fin-R1: Enhancing Financial Intelligence through Domain Expertise, Training Efficiency, and Advanced Reasoning

T0 review · 3 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read A 32B financial model achieves the study's best FinEval, FinanceIQ, and Finova scores while keeping its general reasoning ability.

desk verdict Solid training recipe and a promising new benchmark, but the missing decontamination numbers keep the SOTA claims unverifiable. read the letter →

arxiv 2507.16802 v4 pith:NRUG3JLF submitted 2025-07-22 cs.CL cs.LG

Yanjun Zheng , Xiyang Du , Longfei Liao , Xiaoke Zhao , Zhaowen Zhou , Jingze Song , Bo Zhang , Jiawei Liu
show 5 more authors
Xiang Qi Zhe Li Zhiqiang Zhang Wei Wang Peng Zhang
This is my paper
classification cs.CLcs.LG
keywords financiallargelanguagemodelsreasoningtrustworthinesstasklabelsystemdifficulty-awareweightingtwo-stagetrainingGRPOFinovabenchmark
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

The paper introduces Agentar-Fin-R1, an 8B and 32B family of financial large language models built on the Qwen3 Instruct base. The authors claim that a fine-grained financial task label system, trustworthy multi-agent data synthesis, and difficulty-aware weighted two-stage training can give a 32B model the best reported scores on financial benchmarks (FinEval 1.0, FinanceIQ, and the new Finova suite) while preserving general reasoning on MATH-500 and GPQA-diamond. They also introduce Finova, a 1,350-question benchmark covering agent capabilities, complex financial reasoning, and safety and compliance, as a deployment-oriented test. If the claims hold, domain-specialized training can be both data-efficient and suitable for high-stakes financial use without sacrificing general competence.

What carries the argument

The load-bearing mechanism is the label system $L$, whose composite labels $l_i=(c_i,a_i)$ pair a scene category with a task attribute; the paper treats this as the spine of data construction, training, and error analysis. Difficulty weights are estimated before training via $\mathrm{pass@k}$: for each task label, $k$ responses are sampled from the current model and from $m$ reference models, and the raw weight combines inverse current-model pass@k, a penalty when reference models outperform the current one, and a base weight, with exponential smoothing and normalization across labels. The weighted loss is used in a supervised fine-tuning (SFT) stage for financial knowledge injection, followed by a second stage combining GRPO, a reinforcement-learning update for reasoning models, with targeted SFT for challenge tasks; a pass@1-based attribution loop then allocates extra data and reverts regressions. Finova operationalizes the deployment target by scoring 1,350 real-business queries across agent capabilities, complex reasoning, and safety and compliance.

What would settle it

Re-train Agentar-Fin-R1-32B after removing from the training set every instance that fuzzy-matches, shares named entities and numbers with, or is semantically paraphrased from the released Finova questions; if Finova accuracy drops substantially, the headline result is contamination, and if it stays flat, the generalization claim is supported.

Watch

Extended reading notes

Core claim

The paper's central claim is that a financial LLM can be made simultaneously more specialized and more trustworthy by organizing every pipeline step around a fine-grained two-dimensional task label system. Each label pairs a financial scene (banking, securities, insurance, trusts, mutual funds) with a task attribute (entity recognition, intent classification, slot filling, entity disambiguation, consultation QA), and that label drives data synthesis, stratified sampling, per-task difficulty weighting, and error attribution. Combined with a two-stage SFT-then-GRPO training recipe, the claim is that this yields a 32B model with 87.70 on FinEval 1.0, 86.79 on FinanceIQ, and 69.93 on the new Finova benchmark—the best scores in the paper's comparison—while holding MATH-500 at 93.80 and GPQA-diamond at 68.18, and that an 8B version also outperforms larger specialized baselines on financial tasks. The authors read this as evidence that domain optimization can be achieved without catastrophic forgetting and that Finova's agent, reasoning, and compliance scores reflect deployment readiness.

Load-bearing premise

The load-bearing premise is that Finova's real-world evaluation queries are not drawn from or heavily paraphrased from the same business-context data used to synthesize the training triplets, so the benchmark scores measure capability rather than memorization.

Editorial extensions

If this is right

  • If the FinEval, FinanceIQ, and Finova results hold under clean evaluation, a 32B specialized model can outperform much larger general reasoning systems on financial tasks, making specialized financial deployment cheaper.
  • The 50%-data configuration beating full-data vanilla SFT (75.95 vs 74.82 average) implies that label-guided difficulty weighting can cut training data requirements roughly in half at equal or better accuracy.
  • The two-stage recipe (SFT, then GRPO with targeted SFT) adds about 0.75 average points over SFT-only while keeping MATH-500 near the base model's level, supporting the paper's no-catastrophic-forgetting claim.
  • Finova gives the field a deployment-oriented benchmark in which the 32B model's 87.00 safety-and-compliance score is its strongest dimension, supporting the trustworthiness narrative.
  • The 8B variant's financial scores also beat larger specialized models in the comparison, so the pipeline's efficiency advantage is not exclusive to the 32B scale.

Reading between the lines

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

  • The cleanest test the paper leaves open is a quantitative decontamination audit of Finova against the proprietary financial corpus and the synthesized triplets used in training; the paper describes decontamination only generically, so this audit would settle whether the Finova gains are generalization.
  • The same label-plus-difficulty-weighting recipe should transfer to other regulated domains (legal, medical, engineering) that can define a scene-by-task taxonomy and have a pass@k oracle for grading.
  • The attribution loop could be run prospectively, before training, to decide which labels need more data, turning it from a debugging tool into a data-budget planner.
  • Finova's expression-generation labels are effectively a hallucination-detection signal; wired into a confidence or citation check, they could serve as a runtime guardrail rather than only an evaluation metric.
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

3 major / 5 minor

Summary. The paper introduces Agentar-Fin-R1, an 8B and 32B financial LLM family built on Qwen3-Instruct, combining a financial task label system, multi-agent data synthesis, verification/cleansing, difficulty-weighted SFT, GRPO, and targeted SFT in two stages. It also introduces Finova, a 1,350-sample benchmark with three subdomains: agent capabilities, complex reasoning, and safety/compliance. The principal claim is that Agentar-Fin-R1-32B achieves state-of-the-art results on FinEval 1.0 (87.70), FinanceIQ (86.79), and Finova (69.93) while retaining competitive general reasoning (MATH-500 93.80, GPQA-diamond 68.18), and that the label-guided weighted pipeline reaches these results with strong data efficiency. The paper includes ablation tables on data-budget settings and on the two-stage pipeline, and it releases the Finova benchmark on GitHub.

Significance. If the evaluation evidence holds, the contribution is practically valuable: it demonstrates a reproducible recipe for domain specialization with strong results at 8B and 32B scale, introduces a benchmark that targets agent-level and compliance-oriented financial skills, and provides ablations that separate the label-system and weighting contributions. The explicit comparison against Qwen3 base models and larger specialist baselines is useful, and the release of Finova is a concrete artifact. The weaker link is not the training methodology itself but the evaluation hygiene: the absence of quantified decontamination, the same-team provenance of the new benchmark, and the single-run point estimates all prevent the SOTA claim from being accepted at face value. If the missing contamination analysis is added and confirms disjointness, the central claims are credible and the paper would be a solid systems contribution.

major comments (3)
  1. [Section 2.3.3 and Table 2] The paper's main SOTA claim rests on FinEval 1.0, FinanceIQ, and Finova, but the only safeguard against training/evaluation overlap is the single bullet 'Decontamination: Identifying and removing training data instances that overlap with evaluation benchmarks' in Section 2.3.3. No overlap statistics, no matching procedure, and no benchmark scores before and after decontamination are reported. Because the training mixture includes public general corpora (Ring-Lite, Llama-Nemotron, OpenThoughts) plus proprietary data from the same team's business context, leakage into any of the three evaluation sets could inflate the observed gains over Qwen3-32B and other baselines. Please report the number of matched instances, the overlap-detection method (exact match, near-duplicate embedding, etc.), and the post-decontamination evaluation numbers for all three financial benchmarks.
  2. [Section 4.1, Table 1, and Section 4.2] Finova is introduced, labeled, and scored by the same authors who designed the training data, and it is constructed from 'real-world queries accumulated from actual business environments' while the training set includes the authors' proprietary Agentar-DeepFinance data. The paper provides no contamination analysis for Finova and no inter-annotator agreement statistics for its 1,350 labels. As a consequence, the 69.93 Finova score cannot currently be read as independent evidence that the proposed method, rather than same-source overlap between synthetic training triplets and evaluation queries, drives the improvement. Please provide evidence of source disjointness, third-party annotation of a held-out subset, or an independently constructed external benchmark for the Finova claim.
  3. [Section 4.4, Tables 2 and 3] All reported accuracies are single-run point estimates with no error bars, no cross-seed variance, and no significance tests. The claimed advantage over strong baselines is often only a few percentage points (e.g., FinEval 1.0 87.70 vs DeepSeek-R1 84.93; FinanceIQ 86.79 vs 83.98), and the paper's own comparison on MATH-500 shows a small decrease relative to Qwen3-32B (93.80 vs 95.40) without any variance information. Since the benchmark sizes (notably Finova, with 1,350 total samples and small subtask counts such as 100-150 per subtask) are moderate, the reported margins may be within run-to-run noise. Please report multiple seeds, bootstrap confidence intervals, or an explicit statement of the evaluation protocol's variance.
minor comments (5)
  1. [Abstract and Introduction] The abstract contains the typo 'tow-stage training pipeline' and the name 'Fineva'; the correct name used elsewhere is 'Finova'. Please reconcile these.
  2. [Section 4.5.1] The key-findings text states that the 10% data configuration 'achieves competitive performance (76.68 average)', but Table 4 lists the 10% configuration's 'All Datasets' average as 74.39. The text and table need to be aligned.
  3. [Section 4.2] The text refers to 'our proprietary Agentar-DeepFinance-100K[31]', but reference [31] is titled 'Agentar-DeepFinance-300K: A large-scale financial dataset...'. The dataset name/size in the text and the reference's title are inconsistent; please clarify which dataset was used.
  4. [Section 4.4 and 4.5] The paper does not state whether the trained model weights, the training recipe hyperparameters (alpha, beta, gamma, rho, tau, and the target increment in P_target), or the synthetic data generation code will be released. Releasing the model weights and the hyperparameter configuration would substantially improve reproducibility.
  5. [Section 4.2] The training-detail section gives GPU counts and precision but does not report the total number of SFT examples, the GRPO rollout budget, or the number of optimization steps; these details are needed for the claimed training-efficiency comparisons to be independently assessed.

Circularity Check

2 steps flagged · score 6.0 of 10

Finova, the paper's headline real-world validation instrument, is an author-constructed benchmark from the same Ant Group business context as the proprietary Agentar-DeepFinance-100K training data, and the only stated leakage safeguard is a generic decontamination bullet with no overlap statistics; the Finova-based SOTA claim is therefore partially circular, though the external FinEval 1.0 and…

  1. self definitional [Section 4.1 (Finova benchmark); Section 4.2 (training data); Section 2.3.3 (decontamination); Table 2]
    "The dataset includes real-world queries accumulated from actual business environments, ensuring that the model is tested on high-value, realistic scenarios... Beyond the synthetically generated data derived from our data synthesis framework detailed in Section 2, our dataset incorporates financial reasoning data from our proprietary Agentar-DeepFinance-100K[31]... Decontamination: Identifying and removing training data instances that overlap with evaluation benchmarks to prevent data leakage and ensure fair, unbiased model assessment."

    Finova, the benchmark used for the paper's 'real-world deployment' SOTA claim, is assembled by the same Ant Group authors from the same business environment that produced the proprietary Agentar-DeepFinance-100K training data. The only stated safeguard against overlap is a generic bullet listing 'Decontamination' with no overlap counts, no concrete procedure, and no re-evaluation after removing matched instances. The Finova results (e.g., 69.93 for the 32B model) are therefore not independent evidence of generalization: if Finova queries resemble the synthesized triplets from the same source, the measured advantage is an artifact of training/evaluation overlap.

  2. self citation load bearing [Section 4.2, reference [31]; Section 2.3.2 (synthesis)]
    "Beyond the synthetically generated data derived from our data synthesis framework detailed in Section 2, our dataset incorporates financial reasoning data from our proprietary Agentar-DeepFinance-100K[31], general-purpose training corpora[22], as well as datasets sourced from Llama-Nemotron[1] and openthoughts[9]."

    The core financial reasoning training signal is sourced from an overlapping-author dataset ([31] shares more than ten authors with the present paper). Using a self-created dataset for training is not circular by itself, but it becomes load-bearing here because the same author-controlled pipeline defines both the training distribution and the Finova evaluation distribution. The paper's central claim of a 'trustworthy solution for high-stakes financial applications' depends on combining these two same-context, self-cited artifacts without external validation of Finova or any reported overlap analysis. Thus the strongest deployment claim is supported by a self-citation chain rather than by an independent benchmark.

full rationale

The paper's internal methodology claims - label-guided synthesis, difficulty-aware weighted training, and the two-stage pipeline - are not circular in themselves: the ablations in Tables 4 and 5 compare training variants on fixed evaluation sets, and the weighting scheme (pass@k-based, Eq. 12-16) is fitted to model performance on training labels, not to the evaluation benchmarks. FinEval 1.0 and FinanceIQ are external, publicly available benchmarks, and the strong scores there provide genuine independent evidence that the trained models know financial facts, which prevents a higher score. However, the paper's headline real-world deployment evidence is Finova, a benchmark introduced, labeled, and scored by the same authors from Ant Group business environments, while the training mix includes Agentar-DeepFinance-100K [31], an overlapping-author proprietary dataset from the same business context. The only safeguard described is the one-line decontamination bullet in Section 2.3.3; the paper reports no overlap statistics, no concrete decontamination protocol, no inter-annotator agreement for Finova labels, and no re-evaluation after removing matched instances. Without quantitative separation, the Finova scores (69.93 for Agentar-Fin-R1-32B vs. 61.28 for DeepSeek-R1) could reflect training/evaluation overlap rather than the proposed method, making the 'trustworthy deployment' claim partially circular. This is not an assertion of actual leakage, but a structural circularity in the evidence chain: the validation instrument and the training data are both author-controlled and same-context, and the paper supplies no independent check to rule out the reduction.

Assumptions & free parameters 5 free parameters · 4 assumptions · 1 invented entities

The central empirical claims depend on many unreported hyperparameters and on the independence of the authors' own benchmark and proprietary data.

free parameters (5)
  • weighting hyperparameters alpha, beta, gamma = not reported
    Eq. (12) defines raw difficulty weight as alpha(1-pass@k) + beta*max(0, pass@k_ref - pass@k_current) + gamma; the reported SOTA models depend on these values, but they are never given.
  • exponential smoothing coefficient rho = not reported
    Eq. (14) mixes previous and raw task weights; rho is not reported, so the training dynamics are not reproducible.
  • sampling and generation counts n, m, k = not reported
    Algorithm 1 draws n instances per label, uses m reference models, and k responses per instance; none of these counts are specified.
  • rating-model quality threshold tau = not reported
    Eq. (11) keeps only data with score(d_i) > tau; tau is determined through rigorous empirical validation but never quantified.
  • target performance increment = SOTA + 5 or 10
    Section 3.3 sets P_target as current SOTA plus a fixed increment such as 5 or 10, a hand-chosen stopping target.
assumptions (4)
  • domain assumption Synthetic (query, thinking, answer) triplets generated by LLM agents and filtered by a rating model are accurate enough to serve as training signal.
    Sections 2.3.2-2.3.3; no external audit of the generated data beyond small human samples.
  • ad hoc to paper Decontamination removes all overlap between training data and FinEval, FinanceIQ, and Finova.
    Section 2.3.3 asserts decontamination but supplies no method, no overlap statistics, and no leakage tests.
  • domain assumption Finova ground-truth labels are correct and representative of real deployment.
    Section 4.1 describes Finova tasks but not the annotation protocol, labeler qualifications, or agreement metrics.
  • domain assumption Baseline scores in Tables 2 and 3 were obtained under the same prompts, decoding settings, and evaluation harness as the Agentar models.
    Section 4.4 reports baseline numbers without stating the evaluation code, prompt templates, or whether baselines were rerun by the authors.
invented entities (1)
  • Finova benchmark
    purpose: Evaluate financial agent capabilities (intent, slotting, tool planning, expression), complex reasoning, and safety and compliance for real-world deployment.
    Introduced in this paper by the same team that trained the model; no external use or validation yet; only example items shown.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Agentar-Fin-R1: Enhancing Financial Intelligence through Domain Expertise, Training Efficiency, and Advanced Reasoning." pith.science (2026). https://pith.science/paper/NRUG3JLF

@misc{pith2026250716802,
  author       = {Pith},
  title        = {Pith review of: Agentar-Fin-R1: Enhancing Financial Intelligence through Domain Expertise, Training Efficiency, and Advanced Reasoning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/NRUG3JLF}},
  note         = {Machine review of arXiv:2507.16802}
}
read the original abstract

Large Language Models (LLMs) exhibit considerable promise in financial applications; however, prevailing models frequently demonstrate limitations when confronted with scenarios that necessitate sophisticated reasoning capabilities, stringent trustworthiness criteria, and efficient adaptation to domain-specific requirements. We introduce the Agentar-Fin-R1 series of financial large language models (8B and 32B parameters), specifically engineered based on the Qwen3 foundation model to enhance reasoning capabilities, reliability, and domain specialization for financial applications. Our optimization approach integrates a high-quality, systematic financial task label system with a comprehensive multi-layered trustworthiness assurance framework. This framework encompasses high-quality trustworthy knowledge engineering, multi-agent trustworthy data synthesis, and rigorous data validation governance. Through label-guided automated difficulty-aware optimization, tow-stage training pipeline, and dynamic attribution systems, we achieve substantial improvements in training efficiency. Our models undergo comprehensive evaluation on mainstream financial benchmarks including Fineva, FinEval, and FinanceIQ, as well as general reasoning datasets such as MATH-500 and GPQA-diamond. To thoroughly assess real-world deployment capabilities, we innovatively propose the Finova evaluation benchmark, which focuses on agent-level financial reasoning and compliance verification. Experimental results demonstrate that Agentar-Fin-R1 not only achieves state-of-the-art performance on financial tasks but also exhibits exceptional general reasoning capabilities, validating its effectiveness as a trustworthy solution for high-stakes financial applications. The Finova bench is available at https://github.com/antgroup/Finova.

Figures

Figures reproduced from arXiv: 2507.16802 by the authors.

Figure 1
Figure 1. Agentar-Fin-R1-32B performance: significantly outperforms on financial benchmarks [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Overview of the Agentar-Fin-R1 development pipeline. [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Overview of the dual-track data synthesis pipeline, incorporating both task-oriented [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figures from the paper (6 more)
Figure 4
Figure 4. Figure 4: A comprehensive overview diagram of the Finova benchmark, consisting of three [PITH_FULL_IMAGE:figures/full_fig_p015_4.png]
Figure 5
Figure 5. Figure 5: Performance comparison between Agentar-Fin-R1 and Qwen3 models (8B and 32B [PITH_FULL_IMAGE:figures/full_fig_p019_5.png]
Figure 6
Figure 6. Figure 6: An example for Financial Intent Detection of Finova. [PITH_FULL_IMAGE:figures/full_fig_p029_6.png]
Figure 7
Figure 7. Figure 7: An example for Financial Slot Recognition of Finova. [PITH_FULL_IMAGE:figures/full_fig_p029_7.png]
Figure 8
Figure 8. Figure 8: An example for Financial Tool Planning of Finova.. [PITH_FULL_IMAGE:figures/full_fig_p030_8.png]
Figure 9
Figure 9. Figure 9: An example for Financial Expression Generation of Finova. [PITH_FULL_IMAGE:figures/full_fig_p030_9.png]

Discussion (0). Sign in to comment.

Forward citations

Cited by 2 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

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

    cs.CL 2026-07 conditional novelty 6.0 of 10

    A new benchmark shows LLMs' financial calculation accuracy collapses without explicit formulas and degrades further when they must generate multi-metric tables.

  2. SCOPE and SCION: A Benchmark and an Auditable Reference Pipeline for Schema Induction and Fusion from Text

    cs.AI 2026-05 conditional novelty 6.0 of 10

    A 24-dataset benchmark for inducing schema graphs from raw text, plus an auditable LLM-based pipeline that reports the highest scores on the benchmark's four schema-similarity metrics.

Reference graph

Works this paper leans on

32 extracted references · 13 canonical work pages · cited by 2 Pith papers

  1. [1]

    Llama-nemotron: Efficient reasoning models, 2025

    Akhiad Bercovich, Itay Levy, Izik Golan, Mohammad Dabbah, Ran El-Yaniv, Omri Puny, Ido Galil, Zach Moshe, Tomer Ronen, Najeeb Nabwani, Ido Shahaf, Oren Tropp, Ehud Karpas, Ran Zilberstein, Jiaqi Zeng, Soumye Singhal, et al. Llama-nemotron: Efficient reasoning models, 2025. URL https://arxiv.org/abs/2505.00949

    arXiv 2025

  2. [2]

    Le, Christopher Ré, and Azalia Mirhoseini

    Bradley Brown, Jordan Juravsky, Ryan Ehrlich, Ronald Clark, Quoc V . Le, Christopher Ré, and Azalia Mirhoseini. Large language monkeys: Scaling inference compute with repeated sampling, 2024. URL https://arxiv.org/abs/2407.21787

    arXiv 2024

  3. [3]

    Disc-finllm: A chinese financial large language model based on multiple experts fine-tuning, 2023

    Wei Chen, Qiushi Wang, Zefei Long, Xianyin Zhang, Zhongtian Lu, Bingxuan Li, Siyuan Wang, Jiarong Xu, Xiang Bai, Xuanjing Huang, and Zhongyu Wei. Disc-finllm: A chinese financial large language model based on multiple experts fine-tuning, 2023. URL https: //arxiv.org/abs/2310.15205

    arXiv 2023

  4. [4]

    Fnspid: A comprehensive financial news dataset in time series

    Zihan Dong, Xinyu Fan, and Zhiyuan Peng. Fnspid: A comprehensive financial news dataset in time series. In Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, pages 4918–4927, 2024

    work page 2024

  5. [5]

    FinanceIQ, 2023

    Duxiaoman DI Team. FinanceIQ, 2023. URL https://github.com/Duxiaoman-DI/Xu anYuan/tree/main/FinanceIQ. Accessed: 2024-03-18

    work page 2023

  6. [6]

    XuanYuan3–70b-chat, 2024

    Duxiaoman DI Team. XuanYuan3–70b-chat, 2024. URL https://github.com/Duxiaom an-DI/XuanYuan. Accessed: 2024-03-18

    work page 2024

  7. [7]

    XuanYuan-finx1-preview, 2024

    Duxiaoman DI Team. XuanYuan-finx1-preview, 2024. URL https://github.com/Dux iaoman-DI/XuanYuan. Accessed: 2024-03-18

    work page 2024

  8. [8]

    Can large language models beat wall street? unveiling the potential of ai in stock selection

    Georgios Fatouros, Konstantinos Metaxas, John Soldatos, and Dimosthenis Kyriazis. Can large language models beat wall street? unveiling the potential of ai in stock selection. arXiv preprint arXiv:2401.03737, 2024

    arXiv 2024

Show all 32 references
  1. [9]

    Openthoughts: Data recipes for reasoning models, 2025

    Etash Guha, Ryan Marten, Sedrick Keh, Negin Raoof, Georgios Smyrnis, Hritik Bansal, Marianna Nezhurina, Jean Mercat, Trung Vu, Zayne Sprague, Ashima Suvarna, Benjamin Feuer, Liangyu Chen, Zaid Khan, Eric Frankel, Sachin Grover, et al. Openthoughts: Data recipes for reasoning m...

    2025 arXiv

  2. [11]

    Fineval: A chinese finan- cial domain knowledge evaluation benchmark for large language models, 2024

    Xin Guo, Haotian Xia, Zhaowei Liu, Hanyang Cao, Zhi Yang, Zhiqiang Liu, Sizhe Wang, Jinyi Niu, Chuqi Wang, Yanhui Wang, Xiaolong Liang, Xiaoming Huang, Bing Zhu, Zhongyu Wei, Yun Chen, Weining Shen, and Liwen Zhang. Fineval: A chinese finan- cial domain knowledge evaluation be...

    2024 arXiv

  3. [12]

    Measuring mathematical problem solving with the MATH dataset

    Dan Hendrycks, Collin Burns, Saurav Kadavath, Akul Arora, Steven Basart, Eric Tang, Dawn Song, and Jacob Steinhardt. Measuring mathematical problem solving with the MATH dataset. In Proceedings of NeurIPS, 2021

    2021

  4. [13]

    Alphafin: Benchmarking financial analysis with retrieval-augmented stock-chain framework

    Xiang Li, Zhenyu Li, Chen Shi, Yong Xu, Qing Du, Mingkui Tan, Jun Huang, and Wei Lin. Alphafin: Benchmarking financial analysis with retrieval-augmented stock-chain framework. arXiv preprint arXiv:2403.12582, 2024. 26

    2024 arXiv

  5. [14]

    Deepseek-v3 technical report

    Aixin Liu, Bei Feng, Bing Xue, Bingxuan Wang, Bochao Wu, et al. Deepseek-v3 technical report. CoRR, abs/2412.19437, 2024. URL https://arxiv.org/abs/2412.19437

    2024 arXiv

  6. [15]

    Fin-r1: A large language model for financial reasoning through reinforcement learning

    Zhaowei Liu, Xin Guo, Fangqi Lou, Lingfeng Zeng, Jinyi Niu, Zixuan Wang, Jiajie Xu, Weige Cai, Ziwei Yang, Xueqian Zhao, Chao Li, Sheng Xu, Dezhi Chen, Yun Chen, Zuo Bai, and Liwen Zhang. Fin-r1: A large language model for financial reasoning through reinforcement learning. Co...

    2025

  7. [16]

    Gpt-4o technical report

    OpenAI. Gpt-4o technical report. https://openai.com/research/gpt-4o, 2024

    2024

  8. [17]

    Learning to reason with llms

    OpenAI. Learning to reason with llms. https://openai.com/index/learning-to-r eason-with-llms/, 2024

    2024

  9. [18]

    Fino1: On the transferability of reasoning enhanced llms to finance

    Lingfei Qian, Weipeng Zhou, Yan Wang, Xueqing Peng, Han Yi, Jimin Huang, Qianqian Xie, and Jianyun Nie. Fino1: On the transferability of reasoning enhanced llms to finance. CoRR, abs/2502.08127, 2025. URL https://arxiv.org/abs/2502.08127

    2025 arXiv

  10. [19]

    QwQ: Reflect Deeply on the Boundaries of the Unknown

    Qwen. QwQ: Reflect Deeply on the Boundaries of the Unknown. https://github.com /QwenLM/QwQ, 2024

    2024

  11. [20]

    David Rein, Betty Li Hou, Asa Cooper Stickland, Jackson Petty, Richard Yuanzhe Pang, Julien Dirani, Julian Michael, and Samuel R. Bowman. GPQA: A graduate-level google- proof q&a benchmark. In Proceedings of COLM, 2024

    2024

  12. [21]

    ByteDance Seed, Jiaze Chen, Tiantian Fan, Xin Liu, Lingjun Liu, Zhiqi Lin, Mingxuan Wang, Chengyi Wang, Xiangpeng Wei, Wenyuan Xu, et al. Seed1. 5-thinking: Advancing superb reasoning models with reinforcement learning. arXiv e-prints, pages arXiv–2504, 2025

    2025

  13. [22]

    Ring-lite: Scalable reasoning via c3po-stabilized reinforcement learning for llms, 2025

    Ling Team, Bin Hu, Cai Chen, Deng Zhao, Ding Liu, Dingnan Jin, Feng Zhu, Hao Dai, Hongzhi Luan, Jia Guo, Jiaming Liu, Jiewei Wu, Jun Mei, Jun Zhou, Junbo Zhao, Junwu Xiong, Kaihong Zhang, Kuan Xu, Lei Liang, Liang Jiang, Liangcheng Fu, Longfei Zheng, Qiang Gao, Qing Cui, Quan ...

    2025 arXiv

  14. [23]

    Ploutos: Towards interpretable stock movement prediction with financial large language model

    Hanshuang Tong, Jun Li, Ning Wu, Ming Gong, Dongmei Zhang, and Qi Zhang. Ploutos: Towards interpretable stock movement prediction with financial large language model. arXiv preprint arXiv:2403.00782, 2024

    2024 arXiv

  15. [24]

    Quantagent: Seeking holy grail in trading by self-improving large language model

    Saizhuo Wang, Hang Yuan, Lionel M Ni, and Jian Guo. Quantagent: Seeking holy grail in trading by self-improving large language model. arXiv preprint arXiv:2402.03755, 2024

    2024 arXiv

  16. [25]

    Pixiu: A large language model, instruction data and evaluation benchmark for finance, 2023

    Qianqian Xie, Weiguang Han, Xiao Zhang, Yanzhao Lai, Min Peng, Alejandro Lopez- Lira, and Jimin Huang. Pixiu: A large language model, instruction data and evaluation benchmark for finance, 2023. URL https://arxiv.org/abs/2306.05443

    2023 arXiv

  17. [26]

    Finnlp-agentscen-2024 shared task: Financial challenges in large language models-finllms

    Qianqian Xie, Jimin Huang, Dong Li, Zhengyu Chen, Ruoyu Xiang, Mengxi Xiao, Yangyang Yu, Vijayasai Somasundaram, Kailai Yang, Chenhan Yuan, et al. Finnlp-agentscen-2024 shared task: Financial challenges in large language models-finllms. In Proceedings of the Eighth Financial T...

    2024

  18. [27]

    Qwen2.5 technical report

    An Yang, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, et al. Qwen2.5 technical report. CoRR, abs/2412.15115, 2024. URL https://arxiv.org/abs/2412.15115

    2024 arXiv

  19. [28]

    Qwen3 technical report, 2025

    An Yang, Anfeng Li, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, et al. Qwen3 technical report, 2025. URL https://arxiv.org/abs/2505.09388

    2025 arXiv

  20. [29]

    Does reinforcement learning really incentivize reasoning capacity in llms beyond the base model?, 2025

    Yang Yue, Zhiqi Chen, Rui Lu, Andrew Zhao, Zhaokai Wang, Yang Yue, Shiji Song, and Gao Huang. Does reinforcement learning really incentivize reasoning capacity in llms beyond the base model?, 2025. URL https://arxiv.org/abs/2504.13837

    2025 arXiv

  21. [30]

    Baichuan4-finance technical report, 2025

    Hanyu Zhang, Boyu Qiu, Yuhao Feng, Shuqi Li, Qian Ma, Xiyuan Zhang, Qiang Ju, Dong Yan, and Jian Xie. Baichuan4-finance technical report, 2025. URL https://arxiv.org/ abs/2412.15270

    2025 arXiv

  22. [31]

    Agentar-deepfinance-300k: A large-scale financial dataset via systematic chain-of-thought synthesis optimization, 2025

    Xiaoke Zhao, Zhaowen Zhou, Lin Chen, Lihong Wang, Zhiyi Huang, Kaiyuan Zheng, Yanjun Zheng, Xiyang Du, Longfei Liao, Jiawei Liu, Xiang Qi, Bo Zhang, Peng Zhang, Zhe Li, and Wei Wang. Agentar-deepfinance-300k: A large-scale financial dataset via systematic chain-of-thought synt...

    2025

  23. [32]

    "基金名称"",

    Jie Zhu, Qian Chen, Huaixia Dou, Junhui Li, Lifan Guo, Feng Chen, and Chi Zhang. Dianjin- r1: Evaluating and enhancing financial reasoning in large language models, 2025. URL https://arxiv.org/abs/2504.15716. 28 Appendix Example for Financial Intent DetectionQuestion：你是一个金融意图识...

    2025 arXiv

  24. [2025]

    URL https://arxiv.org/abs/2501.12948

    arXiv

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.