Sourceunknowntext stored

https://arxiv.org/html/2505.24864

Author not recordedDate not recordedOriginal︎ ↗

Claims1

ExaminationsNone

Text89 sections

VersionsOne known

CopiesNone recorded

CorrectionsNone recorded

Facts checked Oct 30, 2026

ReachNo copies recorded

StructureIn 1 claim's map1 source downstream across 1 claim

Evidence1 reading · 1 read wholeOverstates 1No claim cites an examination's findings

Stewards’ notes · 1

ProRL-trained Nemotron-Research-Reasoning-Qwen-1.5B solves Reasoning Gym tasks where DeepSeek-R1-Distill-Qwen-1.5B scores zero at pass@256.
Overstates. The paper's own evaluation section concedes that the base model struggles with Reasoning Gym's answer formatting, and its Reasoning Gym scores are average continuous verifier rewards; the zero for boxnet therefore measures verified, correctly formatted output under this protocol rather than reasoning capability as such. Oct 30, 2026

Examinations

None recorded.

Text383 passages · 1 marked

No examination yetMarked passage

Skip to main content0 passages

Report GitHub Issue6 passages

Report Issue0 passages

Back to Abstract0 passages

Download PDF0 passages

Contents0 passages

Reading mode0 passages

Colour theme0 passages

Abstract0 passages

1 Introduction0 passages

2 ProRL: Prolonged Reinforcement Learning0 passages

3 Nemotron-Research-Reasoning-Qwen-1.5B: The World’s Best 1.5B Reasoning Model2 passages

5 Related Work0 passages

6 Conclusion0 passages

References0 passages

A Limitations0 passages

Computational Resources0 passages

Scalability Concerns0 passages

Training Process Challenges0 passages

Limited Task Scope0 passages

B Societal Impacts0 passages

Potential Benefits and Opportunities0 passages

Ethical Considerations and Challenges0 passages

C Example Prompts0 passages

C.1 Graph Color Example0 passages

C.2 Family Relationships Example0 passages

C.3 Boxnet Example0 passages

D Training Dataset0 passages

D.1 Math0 passages

D.2 Code0 passages

D.3 STEM1 passage

D.5 Instruction Following0 passages

E Training Recipe0 passages

F Results Details0 passages

F.1 Reasoning Gym0 passages

F.2 Pass@k Comparisions0 passages

F.3 Pass@1 Distribution Shifts0 passages

License: arXiv.org perpetual non-exclusive license1 passage

ProRL: Prolonged Reinforcement Learning Expands Reasoning Boundaries in Large Language Models0 passages

Mingjie Liu0 passages

Shizhe Diao0 passages

Ximing Lu0 passages

Jian Hu0 passages

Xin Dong0 passages

Affiliation: Yejin Choi0 passages

Jan Kautz0 passages

Yi Dong0 passages

Affiliation: NVIDIA0 passages

Affiliation: {mingjiel, sdiao, ximingl, jianh, xind, yejinc, jkautz, yidong}@nvidia.com0 passages

Abstract2 passages

1 Introduction7 passages

2 ProRL: Prolonged Reinforcement Learning20 passages

3 Nemotron-Research-Reasoning-Qwen-1.5B: The World’s Best 1.5B Reasoning Model36 passages

Diminish0 passages

Plateau0 passages

Sustained17 passages · 1 marked

4.3 ProRL Enhances Out-of-Distribution Reasoning

Out-of-Distribution (OOD) Task We evaluate the model on Reasoning Gym task boxnet , which was not seen during training. As shown in Figure 5 (Check Appendix C.3 for an example), the base model exhibits no capability of solving the task. In contrast, the model trained with ProRL demonstrates a significant ability to solve the problem, indicating a clear expansion in the model’s reasoning boundary, generalizing to out-of-distribution tasks unseen during training. Furthermore, when comparing an intermediate RL checkpoint with the final prolonged RL model, we observe that extended training sustains and amplifies performance gains consistently across all values of k k . These results further support the conclusion that ProRL enables the model to internalize abstract reasoning patterns that generalize beyond specific training distributions or complexity levels.

ProRL-trained Nemotron-Research-Reasoning-Qwen-1.5B solves Reasoning Gym tasks where DeepSeek-R1-Distill-Qwen-1.5B scores zero at pass@256.DoubtfulReading · this claim’s Steward

Overstates. The paper's own evaluation section concedes that the base model struggles with Reasoning Gym's answer formatting, and its Reasoning Gym scores are average continuous verifier rewards; the zero for boxnet therefore measures verified, correctly formatted output under this protocol rather than reasoning capability as such.

5 Related Work2 passages

6 Conclusion1 passage

Acknowledgments and Disclosure of Funding1 passage

References205 references

Appendix A Limitations1 passage

Computational Resources1 passage

Scalability Concerns1 passage

Training Process Challenges1 passage

Limited Task Scope1 passage

Appendix B Societal Impacts1 passage

Potential Benefits and Opportunities1 passage

Ethical Considerations and Challenges1 passage

Appendix C Example Prompts0 passages

C.1 Graph Color Example3 passages

C.2 Family Relationships Example2 passages

C.3 Boxnet Example4 passages

Appendix D Training Dataset3 passages

Data Source1 passage

DeepScaleR Dataset1 passage

Eurus-2-RL Dataset1 passage

SCP-116K Dataset1 passage

Reasoning Gym1 passage

Llama-Nemotron0 passages

D.1 Math2 passages

D.2 Code1 passage

D.3 STEM4 passages

D.5 Instruction Following1 passage

Appendix E Training Recipe15 passages

Appendix F Results Details0 passages

F.1 Reasoning Gym12 passages

F.2 Pass@k Comparisions1 passage

F.3 Pass@1 Distribution Shifts14 passages

Instructions for reporting errors7 passages

Lineage across 1 claim

Draws on · none recorded

This source

https://arxiv.org/html/2505.24864

in 1 claim's map

Drawn on by · 1

https://arxiv.org/pdf/2507.14843unknown1 claimResponds to it faithfully, on 1 claim.

Record

Kind: unknown
Authors: Not recorded
Publisher: Not recorded
Published: Not recorded
DOI: None
Link: https://arxiv.org/html/2505.24864
Archived copy: None found
Stored text: 76,559 characters, retrieved Oct 30, 2026 · 89 sections, 383 passages
Claims extracted: Not recorded
Versions: One known
Copies: None recorded
Corrections: None recorded
Facts checked: Oct 30, 2026

History 4

Oct 30, 2026ReadingSteward of “ProRL-trained Nemotron-Research-Reasoning-Qwen-…” · overstates, read whole
Oct 30, 2026DividedSegmenter · Divided into 496 parts by its own structure
Oct 30, 2026FactsFacts fetcher · Facts looked up
Oct 30, 2026Entered the graphEntered the graph; copy stored