pith. sign in

arxiv: 2508.15202 · v2 · submitted 2025-08-21 · 💻 cs.CL

Fin-PRM: A Domain-Specialized Process Reward Model for Financial Reasoning in Large Language Models

Jie Zhu , Yuanchen Zhou , Shuo Jiang , Junhui Li , Lifan Guo , Feng Chen , Chi Zhang This is my paper

Pith reviewed 2026-05-18 22:37 UTC · model grok-4.3

classification 💻 cs.CL
keywords process reward modelfinancial reasoninglarge language modelstrajectory supervisionbest-of-n inferencereinforcement learningdomain specialization
0
0 comments X

The pith

Fin-PRM improves financial reasoning in LLMs by providing step-level and trajectory-level supervision from multi-source labels.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

General process reward models trained on broad data struggle with the fact-sensitive and structured steps needed in financial reasoning. Fin-PRM trains a specialized model on 3,000 financial trajectories whose step and trajectory labels are generated automatically through Monte Carlo rollouts, LLM judgments, and explicit financial knowledge checks. It produces a single unified ranking score that combines local step correctness with global coherence. The model is tested in three practical settings: selecting trajectories for fine-tuning, guiding Best-of-N inference, and shaping rewards during reinforcement learning. On financial benchmarks including CFLUE and FinQA it outperforms both general-purpose PRMs and other strong baselines.

Core claim

Fin-PRM is a trajectory-aware process reward model that jointly models step-level correctness and trajectory-level coherence through binary supervision signals derived from Monte Carlo rollouts, LLM-based evaluation, and explicit financial knowledge verification, yielding a unified ranking score that improves performance when applied to offline trajectory selection, reward-guided Best-of-N inference, and process-aware reward shaping for reinforcement learning on financial reasoning tasks.

What carries the argument

The unified ranking score that integrates binary step-level and trajectory-level rewards generated from Monte Carlo rollouts, LLM evaluation, and financial knowledge verification.

If this is right

  • Offline selection of reasoning trajectories for supervised fine-tuning on financial tasks becomes more accurate.
  • Best-of-N inference at test time gains from process-level signals rather than final-answer rewards alone.
  • Reinforcement learning for financial reasoning can use finer-grained process supervision to shape rewards.
  • The performance gains hold across multiple financial benchmarks including CFLUE and FinQA.

Where Pith is reading between the lines

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

  • The multi-source automatic labeling approach could be adapted to create reliable supervision in other knowledge-intensive domains such as legal or medical reasoning.
  • Combining rollout-based signals with explicit domain knowledge checks may reduce reliance on costly human annotations for new specialized tasks.
  • The same trajectory-aware scoring could be tested for improving calibration and error detection in non-financial structured reasoning problems.

Load-bearing premise

Automatically derived labels from Monte Carlo rollouts, LLM evaluation, and financial knowledge verification produce reliable step and trajectory supervision without systematic bias or circularity.

What would settle it

Human experts rating a random sample of the auto-labeled trajectories and finding frequent errors in step correctness, or a replication experiment in which Fin-PRM shows no advantage or underperforms general PRMs on the same benchmarks.

Figures

Figures reproduced from arXiv: 2508.15202 by Chi Zhang, Feng Chen, Jie Zhu, Junhui Li, Lifan Guo, Shuo Jiang, Yuanchen Zhou.

Figure 1
Figure 1. Figure 1: Total process from data construction to model [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: BoN test on Cflue dataset. Fin-PRM is the [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Performance of GRPO policy optimization using [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Ablation study on the ranking score weight [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] view at source ↗
read the original abstract

Process Reward Models (PRMs) supervise intermediate reasoning steps in large language models (LLMs), but existing PRMs are mainly trained on general-domain data and struggle with the structured, symbolic, and fact-sensitive nature of financial reasoning. Financial tasks require not only correct final answers but also verifiable intermediate steps grounded in domain knowledge. In this paper, we propose Fin-PRM, a domain-specialized, trajectory-aware PRM for financial reasoning that jointly models step-level correctness and trajectory-level coherence, producing binary supervision signals for both local and global reasoning quality. To support reliable supervision, we construct a high-quality financial reasoning dataset of 3K trajectories, where step- and trajectory-level labels are automatically derived from multi-source reward signals, including Monte Carlo rollouts, LLM-based evaluation, and explicit financial knowledge verification. Fin-PRM defines a unified ranking score that integrates step- and trajectory-level rewards, enabling consistent use across multiple settings. We evaluate Fin-PRM in three scenarios: (1) offline trajectory selection for supervised fine-tuning, (2) reward-guided Best-of-$N$ inference for test-time scaling, and (3) process-aware reward shaping for reinforcement learning. Experiments on financial reasoning benchmarks, including CFLUE and FinQA, show that Fin-PRM consistently outperforms general-purpose PRMs and strong baselines. Our project resources will be available at https://github.com/aliyun/qwen-dianjin.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 1 minor

Summary. The paper introduces Fin-PRM, a domain-specialized process reward model for financial reasoning in LLMs. It constructs a 3K-trajectory dataset with step- and trajectory-level binary labels automatically derived from Monte Carlo rollouts, LLM-based evaluation, and explicit financial knowledge verification. A unified ranking score is defined, and the model is evaluated in three settings: offline trajectory selection for SFT, reward-guided Best-of-N inference, and process-aware reward shaping for RL. Experiments on CFLUE and FinQA benchmarks claim consistent outperformance over general-purpose PRMs and baselines.

Significance. If the multi-source labeling produces reliable, unbiased supervision independent of the target models, this could meaningfully advance process supervision for structured, fact-sensitive domains like finance. The three-scenario evaluation and planned resource release strengthen potential utility and reproducibility.

major comments (2)
  1. [Dataset construction] Dataset construction (abstract and §3): no quantitative validation is reported for the automatic labeling pipeline, such as human agreement rates, error rates on the 3K trajectories, or ablation results isolating the contribution of Monte Carlo rollouts versus LLM evaluation versus knowledge verification. This directly undermines confidence in the binary supervision signals that support all three evaluation scenarios.
  2. [Labeling process] Labeling process (abstract): the description of labels derived from LLM-based evaluation and Monte Carlo rollouts does not address potential circularity when the evaluator models share architecture, training data, or prompting style with the base models used in the CFLUE/FinQA experiments. This is load-bearing because the central outperformance claim rests on the independence of the supervision signal.
minor comments (1)
  1. [Abstract] The abstract states that project resources will be available at the GitHub link but does not enumerate exactly which artifacts (dataset, code, trained model) will be released.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive feedback on our work. We address each major comment below and describe the revisions planned for the next version of the manuscript.

read point-by-point responses

  1. Referee: [Dataset construction] Dataset construction (abstract and §3): no quantitative validation is reported for the automatic labeling pipeline, such as human agreement rates, error rates on the 3K trajectories, or ablation results isolating the contribution of Monte Carlo rollouts versus LLM evaluation versus knowledge verification. This directly undermines confidence in the binary supervision signals that support all three evaluation scenarios.

    Authors: We agree that the absence of quantitative validation for the automatic labeling pipeline is a limitation that reduces confidence in the reported supervision signals. In the revised manuscript we will add a human evaluation on a random subset of 200 trajectories, reporting inter-annotator agreement and estimated error rates. We will also include ablation studies that measure the incremental contribution of Monte Carlo rollouts, LLM-based evaluation, and explicit financial knowledge verification to the final Fin-PRM performance. These additions will be placed in a new subsection of §3. revision: yes

  2. Referee: [Labeling process] Labeling process (abstract): the description of labels derived from LLM-based evaluation and Monte Carlo rollouts does not address potential circularity when the evaluator models share architecture, training data, or prompting style with the base models used in the CFLUE/FinQA experiments. This is load-bearing because the central outperformance claim rests on the independence of the supervision signal.

    Authors: We acknowledge that the current description does not explicitly discuss independence between the labeling models and the downstream evaluation models. The multi-source pipeline combines Monte Carlo outcome rollouts (which depend only on final-answer correctness) with explicit financial knowledge verification that uses rule-based fact checking rather than LLM judgments. The LLM component uses a model prompted with domain-specific financial instructions that differ from the general-purpose prompting used in the CFLUE and FinQA experiments. We will expand §3 to detail the specific models, prompting strategies, and verification rules employed, thereby clarifying how the supervision signal remains independent of the target models. revision: partial

Circularity Check

0 steps flagged

No significant circularity in derivation chain

full rationale

The paper constructs a 3K-trajectory dataset with step- and trajectory-level binary labels derived from multi-source signals (Monte Carlo rollouts, LLM-based evaluation, and explicit financial knowledge verification). Fin-PRM is then trained on these labels to produce a unified ranking score, which is applied in offline selection, Best-of-N, and RL shaping. Evaluation occurs on external benchmarks (CFLUE, FinQA) against general-purpose PRMs and baselines. No equations or steps reduce the claimed outperformance to the labeling inputs by construction, no self-citation chains justify core premises, and no ansatz or uniqueness result is smuggled in. The supervision pipeline is presented as independent domain-specific construction rather than a definitional loop or fitted-input renaming.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The central claim rests on the quality of the automatically labeled 3K trajectories and the assumption that multi-source signals (Monte Carlo, LLM eval, financial knowledge verification) produce unbiased binary labels for both local and global reasoning quality.

axioms (1)
  • domain assumption Multi-source automatic labeling (Monte Carlo rollouts + LLM evaluation + explicit financial knowledge verification) produces reliable step- and trajectory-level binary labels.
    Invoked in the dataset construction paragraph of the abstract; no human validation or inter-annotator agreement is mentioned.

pith-pipeline@v0.9.0 · 5803 in / 1336 out tokens · 36329 ms · 2026-05-18T22:37:41.171384+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

What do these tags mean?
matches
The paper's claim is directly supported by a theorem in the formal canon.
supports
The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends
The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses
The paper appears to rely on the theorem as machinery.
contradicts
The paper's claim conflicts with a theorem or certificate in the canon.
unclear
Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Forward citations

Cited by 1 Pith paper

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

  1. Rewarding the Scientific Process: Process-Level Reward Modeling for Agentic Data Analysis

    cs.CL 2026-04 unverdicted novelty 7.0

    DataPRM is a new process reward model for data analysis agents that detects silent errors via environment interaction and ternary rewards, yielding 7-11% gains on benchmarks and further RL improvements.

Reference graph

Works this paper leans on

32 extracted references · 32 canonical work pages · cited by 1 Pith paper · 19 internal anchors

  1. [1]

    , " * write output.state after.block = add.period write newline

    ENTRY address archivePrefix author booktitle chapter edition editor eid eprint howpublished institution isbn journal key month note number organization pages publisher school series title type volume year label extra.label sort.label short.list INTEGERS output.state before.all mid.sentence after.sentence after.block FUNCTION init.state.consts #0 'before.a...

    work page

  2. [2]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION word.in bbl.in capitalize " " * FUNCT...

    work page

  3. [3]

    Bai, Y.; Jones, A.; Ndousse, K.; Askell, A.; Chen, A.; DasSarma, N.; Drain, D.; Fort, S.; Ganguli, D.; Henighan, T.; Joseph, N.; Kadavath, S.; Kernion, J.; Conerly, T.; El-Showk, S.; Elhage, N.; Hatfield-Dodds, Z.; Hernandez, D.; Hume, T.; Johnston, S.; Kravec, S.; Lovitt, L.; Nanda, N.; Olsson, C.; Amodei, D.; Brown, T.; Clark, J.; McCandlish, S.; Olah, ...

    work page internal anchor Pith review Pith/arXiv arXiv 2022

  4. [4]

    Cui, G.; Yuan, L.; Wang, Z.; Wang, H.; Li, W.; He, B.; Fan, Y.; Yu, T.; Xu, Q.; Chen, W.; Yuan, J.; Chen, H.; Zhang, K.; Lv, X.; Wang, S.; Yao, Y.; Han, X.; Peng, H.; Cheng, Y.; Liu, Z.; Sun, M.; Zhou, B.; and Ding, N. 2025. Process Reinforcement through Implicit Rewards. arXiv:2502.01456

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  5. [5]

    DeepSeek-AI; Guo, D.; Yang, D.; Zhang, H.; Song, J.; Zhang, R.; Xu, R.; Zhu, Q.; Ma, S.; Wang, P.; Bi, X.; Zhang, X.; Yu, X.; Wu, Y.; Wu, Z. F.; Gou, Z.; Shao, Z.; Li, Z.; Gao, Z.; Liu, A.; Xue, B.; Wang, B.; Wu, B.; Feng, B.; Lu, C.; Zhao, C.; Deng, C.; Zhang, C.; Ruan, C.; Dai, D.; Chen, D.; Ji, D.; Li, E.; Lin, F.; Dai, F.; Luo, F.; Hao, G.; Chen, G.; ...

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  6. [6]

    Gu, J.; Jiang, X.; Shi, Z.; Tan, H.; Zhai, X.; Xu, C.; Li, W.; Shen, Y.; Ma, S.; Liu, H.; Wang, S.; Zhang, K.; Wang, Y.; Gao, W.; Ni, L.; and Guo, J. 2025. A Survey on LLM-as-a-Judge. arXiv:2411.15594

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  7. [7]

    Guha, E.; Marten, R.; Keh, S.; Raoof, N.; Smyrnis, G.; Bansal, H.; Nezhurina, M.; Mercat, J.; Vu, T.; Sprague, Z.; Suvarna, A.; Feuer, B.; Chen, L.; Khan, Z.; Frankel, E.; Grover, S.; Choi, C.; Muennighoff, N.; Su, S.; Zhao, W.; Yang, J.; Pimpalgaonkar, S.; Sharma, K.; Ji, C. C.-J.; Deng, Y.; Pratt, S.; Ramanujan, V.; Saad-Falcon, J.; Li, J.; Dave, A.; Al...

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  8. [8]

    Gunasekar, S.; Zhang, Y.; Aneja, J.; Mendes, C. C. T.; Giorno, A. D.; Gopi, S.; Javaheripi, M.; Kauffmann, P.; de Rosa, G.; Saarikivi, O.; Salim, A.; Shah, S.; Behl, H. S.; Wang, X.; Bubeck, S.; Eldan, R.; Kalai, A. T.; Lee, Y. T.; and Li, Y. 2023. Textbooks Are All You Need. arXiv:2306.11644

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  9. [9]

    Y.; Zeng, L.; Wang, X.; Wang, B.; Li, Y.; Zhang, F.; Xu, J.; An, B.; Liu, Y.; and Zhou, Y

    He, J.; Wei, T.; Yan, R.; Liu, J.; Wang, C.; Gan, Y.; Tu, S.; Liu, C. Y.; Zeng, L.; Wang, X.; Wang, B.; Li, Y.; Zhang, F.; Xu, J.; An, B.; Liu, Y.; and Zhou, Y. 2024. Skywork-o1 Open Series. https://huggingface.co/Skywork

    work page 2024

  10. [10]

    Khalifa, M.; Agarwal, R.; Logeswaran, L.; Kim, J.; Peng, H.; Lee, M.; Lee, H.; and Wang, L. 2025. Process Reward Models That Think. arXiv:2504.16828

    work page arXiv 2025

  11. [12]

    Lightman, H.; Kosaraju, V.; Burda, Y.; Edwards, H.; Baker, B.; Lee, T.; Leike, J.; Schulman, J.; Sutskever, I.; and Cobbe, K. 2023 b . Let's Verify Step by Step. arXiv:2305.20050

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  12. [13]

    Liu, R.; Gao, J.; Zhao, J.; Zhang, K.; Li, X.; Qi, B.; Ouyang, W.; and Zhou, B. 2025. Can 1B LLM Surpass 405B LLM? Rethinking Compute-Optimal Test-Time Scaling. arXiv:2502.06703

    work page arXiv 2025

  13. [14]

    Luo, H.; Sun, Q.; Xu, C.; Zhao, P.; Lou, J.; Tao, C.; Geng, X.; Lin, Q.; Chen, S.; Tang, Y.; and Zhang, D. 2025. WizardMath: Empowering Mathematical Reasoning for Large Language Models via Reinforced Evol-Instruct. arXiv:2308.09583

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  14. [15]

    s1: Simple test-time scaling

    Muennighoff, N.; Yang, Z.; Shi, W.; Li, X. L.; Fei-Fei, L.; Hajishirzi, H.; Zettlemoyer, L.; Liang, P.; Candès, E.; and Hashimoto, T. 2025. s1: Simple test-time scaling. arXiv:2501.19393

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  15. [16]

    Mukherjee, S.; Mitra, A.; Jawahar, G.; Agarwal, S.; Palangi, H.; and Awadallah, A. 2023. Orca: Progressive Learning from Complex Explanation Traces of GPT-4. arXiv:2306.02707

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  16. [17]

    Setlur, A.; Nagpal, C.; Fisch, A.; Geng, X.; Eisenstein, J.; Agarwal, R.; Agarwal, A.; Berant, J.; and Kumar, A. 2024. Rewarding Progress: Scaling Automated Process Verifiers for LLM Reasoning. arXiv:2410.08146

    work page internal anchor Pith review arXiv 2024

  17. [18]

    DeepSeekMath: Pushing the Limits of Mathematical Reasoning in Open Language Models

    Shao, Z.; Wang, P.; Zhu, Q.; Xu, R.; Song, J.; Bi, X.; Zhang, H.; Zhang, M.; Li, Y. K.; Wu, Y.; and Guo, D. 2024. DeepSeekMath: Pushing the Limits of Mathematical Reasoning in Open Language Models. arXiv:2402.03300

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  18. [19]

    Snell, C.; Lee, J.; Xu, K.; and Kumar, A. 2024. Scaling LLM Test-Time Compute Optimally can be More Effective than Scaling Model Parameters. arXiv:2408.03314

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  19. [20]

    Uesato, J.; Kushman, N.; Kumar, R.; Song, F.; Siegel, N.; Wang, L.; Creswell, A.; Irving, G.; and Higgins, I. 2022. Solving math word problems with process- and outcome-based feedback. arXiv:2211.14275

    work page internal anchor Pith review Pith/arXiv arXiv 2022

  20. [21]

    Math-Shepherd: Verify and Reinforce LLMs Step-by-step without Human Annotations

    Wang, P.; Li, L.; Shao, Z.; Xu, R. X.; Dai, D.; Li, Y.; Chen, D.; Wu, Y.; and Sui, Z. 2024. Math-Shepherd: Verify and Reinforce LLMs Step-by-step without Human Annotations. arXiv:2312.08935

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  21. [22]

    Wei, J.; and Zou, K. 2019. EDA: Easy Data Augmentation Techniques for Boosting Performance on Text Classification Tasks. arXiv:1901.11196

    work page arXiv 2019

  22. [23]

    Xia, M.; Malladi, S.; Gururangan, S.; Arora, S.; and Chen, D. 2024. LESS: Selecting Influential Data for Targeted Instruction Tuning. arXiv:2402.04333

    work page arXiv 2024

  23. [24]

    Le, Tengyu Ma, and Adams Wei Yu

    Xie, S. M.; Pham, H.; Dong, X.; Du, N.; Liu, H.; Lu, Y.; Liang, P.; Le, Q. V.; Ma, T.; and Yu, A. W. 2023. DoReMi: Optimizing Data Mixtures Speeds Up Language Model Pretraining. arXiv:2305.10429

    work page arXiv 2023

  24. [25]

    Yang, H.; Liu, X.-Y.; and Wang, C. D. 2023. FinGPT: Open-Source Financial Large Language Models. arXiv:2306.06031

    work page arXiv 2023

  25. [26]

    Ye, Y.; Huang, Z.; Xiao, Y.; Chern, E.; Xia, S.; and Liu, P. 2025. LIMO: Less is More for Reasoning. arXiv:2502.03387

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  26. [27]

    MetaMath: Bootstrap Your Own Mathematical Questions for Large Language Models

    Yu, L.; Jiang, W.; Shi, H.; Yu, J.; Liu, Z.; Zhang, Y.; Kwok, J. T.; Li, Z.; Weller, A.; and Liu, W. 2024. MetaMath: Bootstrap Your Own Mathematical Questions for Large Language Models. arXiv:2309.12284

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  27. [28]

    Zhang, K.; Zhang, J.; Li, H.; Zhu, X.; Hua, E.; Lv, X.; Ding, N.; Qi, B.; and Zhou, B. 2024. OpenPRM: Building Open-domain Process-based Reward Models with Preference Trees. ICLR 2024

    work page 2024

  28. [29]

    Zhang, Z.; Zheng, C.; Wu, Y.; Zhang, B.; Lin, R.; Yu, B.; Liu, D.; Zhou, J.; and Lin, J. 2025. The Lessons of Developing Process Reward Models in Mathematical Reasoning. arXiv:2501.07301

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  29. [30]

    Judging LLM-as-a-Judge with MT-Bench and Chatbot Arena

    Zheng, L.; Chiang, W.-L.; Sheng, Y.; Zhuang, S.; Wu, Z.; Zhuang, Y.; Lin, Z.; Li, Z.; Li, D.; Xing, E. P.; Zhang, H.; Gonzalez, J. E.; and Stoica, I. 2023. Judging LLM-as-a-Judge with MT-Bench and Chatbot Arena. arXiv:2306.05685

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  30. [31]

    Zhu, J.; Chen, Q.; Dou, H.; Li, J.; Guo, L.; Chen, F.; and Zhang, C. 2025. DianJin-R1: Evaluating and Enhancing Financial Reasoning in Large Language Models. arXiv:2504.15716

    work page arXiv 2025

  31. [32]

    Zhu, J.; Li, J.; Wen, Y.; and Guo, L. 2024. Benchmarking Large Language Models on CFLUE - A C hinese Financial Language Understanding Evaluation Dataset. In Ku, L.-W.; Martins, A.; and Srikumar, V., eds., Findings of the Association for Computational Linguistics: ACL 2024, 5673--5693. Bangkok, Thailand: Association for Computational Linguistics

    work page 2024

  32. [33]

    Zou, J.; Yang, L.; Gu, J.; Qiu, J.; Shen, K.; He, J.; and Wang, M. 2025. ReasonFlux-PRM: Trajectory-Aware PRMs for Long Chain-of-Thought Reasoning in LLMs. arXiv:2506.18896

    work page arXiv 2025