pith. sign in

arxiv: 2605.21917 · v1 · pith:MOTN2NH4new · submitted 2026-05-21 · 💻 cs.CV · cs.AI

MAVEN: A Multi-stage Agentic Annotation Pipeline for Video Reasoning Tasks

Han Zhang , Wanting Jiang , Tomasz Kornuta , Tian Zheng , Vidya Murali This is my paper

Pith reviewed 2026-05-22 07:31 UTC · model grok-4.3

classification 💻 cs.CV cs.AI
keywords video annotationagentic pipelinevision language modelsevent reasoningchain of thoughtmulti-stage annotationdomain adaptationtraffic video analysis
0
0 comments X

The pith

MAVEN uses a multi-stage agentic pipeline to turn raw videos into structured CoT annotations that let a fine-tuned 8B VLM surpass Gemini models on traffic event reasoning tasks.

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

The paper introduces MAVEN as a way to generate large-scale training data for vision-language models that reason about video events, including when, where, why, and with what consequence. It builds a multi-scale spatio-temporal event description from three caption levels and uses this as the foundation for creating questions and answers in multiple formats with reasoning traces. The pipeline lets an agent adapt prompts and structure to new video domains on its own and runs a loop that classifies errors, traces them to specific stages, and rewrites parts of the process. When applied to more than 5,300 traffic videos, the resulting data produces fine-tuned models that show large gains on CCTV and accident benchmarks, outperforming or matching stronger baselines even when trained only on one camera type.

Core claim

MAVEN synthesizes a Multi-Scale Spatio-Temporal Event Description from three complementary caption levels as the sole input to downstream multi-task Q&A generation, while supporting agent-driven domain adaptation that redesigns prompts top-down and a hierarchical refinement loop that classifies errors against a taxonomy, traces root causes, and applies targeted prompt or pipeline edits to improve quality iteratively.

What carries the argument

The Multi-Scale Spatio-Temporal Event Description (MSTED) synthesized from three caption levels, which serves as the explicit intermediate representation for all subsequent question-answer generation and reasoning trace creation.

If this is right

  • Fine-tuning Cosmos-Reason2-8B on the MAVEN-labeled traffic videos yields a +38.8 point MCQ accuracy gain over zero-shot and surpasses both Gemini 2.5 Pro and 3.1 Flash on the private CCTV evaluation set.
  • CCTV-only training on the annotated data lifts the same model by +10.7 MCQ points on AccidentBench and matches Gemini 2.5 Pro performance without any dashcam examples.
  • Adding agent-adapted dashcam annotations narrows the remaining gap to Gemini 3.1 Flash, while subsequent RL post-training exceeds both Gemini baselines.
  • The same agentic workflow adapts the pipeline structure and prompts to warehouse surveillance and public safety videos with minimal manual intervention.

Where Pith is reading between the lines

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

  • The approach could lower the barrier to creating specialized video reasoning datasets in other narrow domains such as retail monitoring or medical procedure analysis.
  • Hierarchical error tracing might generalize to automated quality control loops in other data-generation pipelines beyond video.
  • If the MSTED intermediate proves reusable, it could support transfer of reasoning capabilities across different camera types or video sources without full re-annotation.

Load-bearing premise

The agent-driven multi-stage pipeline with hierarchical error classification produces annotations of high enough quality to support the reported performance gains without introducing systematic biases that would explain the improvements.

What would settle it

A detailed manual review of a sample of the generated annotations that finds frequent factual errors, missing temporal details, or biased framing, combined with the fine-tuned model failing to achieve the stated accuracy lifts on the private CCTV set or AccidentBench.

Figures

Figures reproduced from arXiv: 2605.21917 by Han Zhang, Tian Zheng, Tomasz Kornuta, Vidya Murali, Wanting Jiang.

Figure 1
Figure 1. Figure 1: The MAVEN pipeline, organized into three components: agent-assisted top-down configuration design (left), annotation pipeline execution (center), and hierarchical pipeline refinement with human feedback (right). it for completeness and accuracy before Q&A generation, preventing error propagation into the training data. Second, it is the sole input to Stage 3: Q&A generators never see the raw video or origi… view at source ↗
Figure 2
Figure 2. Figure 2: Qualitative MAVEN outputs on two generalized domains. Top (Public Safety): a complex altercation in a pedestrian underpass, where an ambush turns into a defensive counter-attack and the initial aggressor is knocked unconscious by the intended victim. Bottom (Warehouse Surveillance): a suspicious-criminal case where an individual systematically searches two workstations, conceals a white object under their … view at source ↗
read the original abstract

Training Vision Language Models (VLMs) for video event reasoning requires high-quality structured annotations capturing not only what happened, but when, where, why, and with what consequence, at a scale manual labelling cannot support. We present MAVEN (Multi-stage Agentic Video Event aNnotation), a multi-stage agentic pipeline that turns raw videos into multi-task training data with Chain-of-Thought (CoT) reasoning traces, organized around a designated Event of Focus. At its core, MAVEN synthesizes a Multi-Scale Spatio-Temporal Event Description (MSTED) from three complementary caption levels; this explicit intermediate serves as the sole input to downstream Q&A generation across multiple task formats. Crucially, MAVEN supports agent-driven domain adaptation: given a new video dataset and target question examples, the agent redesigns all prompts top-down without manual re-engineering. A hierarchical refinement loop further classifies annotation errors against a taxonomy, traces root causes to the originating pipeline stage, and applies targeted edits that rewrite prompts or modify the pipeline structure itself, iteratively improving data quality. We apply MAVEN to label over 5,300 traffic videos and fine-tune Cosmos-Reason2-8B on the resulting data. On a private CCTV evaluation set, fine-tuning surpasses both Gemini 2.5 Pro and 3.1 Flash, including a $+38.8$-point gain in MCQ accuracy over zero-shot. On AccidentBench, CCTV-only training lifts Cosmos-Reason2 by $+10.7$ MCQ points and matches Gemini 2.5 Pro despite seeing no dashcam videos; adding agent-adapted dashcam annotations narrows the gap to Gemini 3.1 Flash, and RL post-training pushes overall performance past both Gemini baselines. Qualitative results on warehouse surveillance and public safety videos further show the agentic workflow readily adapts the pipeline to new domains.

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

3 major / 2 minor

Summary. The paper introduces MAVEN, a multi-stage agentic pipeline for generating structured annotations for video event reasoning in VLMs. It produces Multi-Scale Spatio-Temporal Event Descriptions (MSTED) from three caption levels as an intermediate representation, followed by multi-task Q&A generation with CoT traces centered on an Event of Focus. The pipeline incorporates agent-driven domain adaptation for new datasets and a hierarchical refinement loop that classifies errors by taxonomy, traces them to pipeline stages, and applies targeted prompt or structural edits. Applied to over 5,300 traffic videos, the resulting data is used to fine-tune Cosmos-Reason2-8B, yielding a +38.8 MCQ accuracy gain over zero-shot on a private CCTV set (surpassing Gemini 2.5 Pro and 3.1 Flash) and a +10.7 point lift on AccidentBench that matches Gemini 2.5 Pro with CCTV-only training.

Significance. If the annotation quality claims are substantiated, MAVEN could meaningfully advance scalable data creation for video reasoning tasks, addressing the bottleneck of manual labeling for detailed spatio-temporal and causal descriptions. The agentic adaptation mechanism offers a practical route to domain transfer without per-dataset re-engineering, which is valuable for surveillance and safety applications. The reported downstream gains on both in-domain and out-of-domain benchmarks suggest potential for improving VLM reasoning capabilities when high-quality structured traces are available.

major comments (3)
  1. [Experiments / Results] The central performance claims (e.g., +38.8 MCQ gain on private CCTV and +10.7 on AccidentBench) rest on the assumption that MAVEN annotations are sufficiently accurate and unbiased. However, the manuscript provides no quantitative validation of annotation quality, such as human agreement rates on MSTED completeness, factual error rates for Event-of-Focus localization, or comparison of generated Q&A against expert labels. This omission is load-bearing because downstream fine-tuning results alone cannot distinguish genuine reasoning improvement from exploitation of consistent pipeline artifacts.
  2. [MAVEN Pipeline Description] The hierarchical refinement loop is presented as iteratively improving data quality via error taxonomy and targeted edits, yet no metrics are reported on iteration-wise changes in annotation accuracy, error reduction rates, or inter-iteration consistency. Without these, it is unclear whether the loop delivers measurable gains or merely describes a process whose output quality remains unverified.
  3. [Evaluation Setup] The private CCTV evaluation set is used to demonstrate surpassing Gemini baselines, but the manuscript does not detail its size, composition, overlap with training videos, or annotation protocol. This information is necessary to assess whether the reported gains reflect improved generalization or evaluation-set characteristics.
minor comments (2)
  1. [Introduction] The acronym MSTED is expanded on first use in the abstract but could benefit from a brief reminder of its components (Multi-Scale Spatio-Temporal Event Description) when re-introduced in the main text.
  2. [Figures] Figure captions for pipeline diagrams should explicitly label each stage (caption generation, MSTED synthesis, Q&A generation, refinement) to improve readability for readers unfamiliar with the workflow.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for their constructive and detailed feedback on our manuscript. We address each major comment below and have incorporated revisions to strengthen the presentation of our results and methodology.

read point-by-point responses

  1. Referee: [Experiments / Results] The central performance claims (e.g., +38.8 MCQ gain on private CCTV and +10.7 on AccidentBench) rest on the assumption that MAVEN annotations are sufficiently accurate and unbiased. However, the manuscript provides no quantitative validation of annotation quality, such as human agreement rates on MSTED completeness, factual error rates for Event-of-Focus localization, or comparison of generated Q&A against expert labels. This omission is load-bearing because downstream fine-tuning results alone cannot distinguish genuine reasoning improvement from exploitation of consistent pipeline artifacts.

    Authors: We agree that quantitative validation of annotation quality is important to substantiate the performance claims and to rule out pipeline artifacts. The original manuscript emphasized the pipeline design and downstream task results but did not include direct human evaluation metrics. In the revised version, we will add a new subsection reporting human agreement rates on MSTED completeness and factual accuracy, error rates for Event-of-Focus localization, and comparisons of generated Q&A against expert labels on a sampled subset of the data. revision: yes

  2. Referee: [MAVEN Pipeline Description] The hierarchical refinement loop is presented as iteratively improving data quality via error taxonomy and targeted edits, yet no metrics are reported on iteration-wise changes in annotation accuracy, error reduction rates, or inter-iteration consistency. Without these, it is unclear whether the loop delivers measurable gains or merely describes a process whose output quality remains unverified.

    Authors: We acknowledge that the manuscript describes the hierarchical refinement loop but does not provide quantitative metrics tracking its impact across iterations. In the revision, we will add results from our internal experiments, including error reduction rates and inter-iteration consistency measures, presented in a table or figure to demonstrate measurable improvements delivered by the loop. revision: yes

  3. Referee: [Evaluation Setup] The private CCTV evaluation set is used to demonstrate surpassing Gemini baselines, but the manuscript does not detail its size, composition, overlap with training videos, or annotation protocol. This information is necessary to assess whether the reported gains reflect improved generalization or evaluation-set characteristics.

    Authors: We agree that additional details on the private CCTV evaluation set are required for a complete assessment. In the revised manuscript, we will expand the evaluation setup section to specify the set size (number of videos and questions), its composition by event types, confirmation of no overlap with training videos, and the ground-truth annotation protocol used for evaluation. revision: yes

Circularity Check

0 steps flagged

No circularity: empirical results rest on external benchmarks

full rationale

The paper presents an agentic annotation pipeline (MAVEN) that generates MSTED descriptions and CoT traces from raw videos, applies it to 5300 traffic videos, fine-tunes Cosmos-Reason2-8B, and reports accuracy gains on private CCTV and AccidentBench evaluation sets against Gemini baselines. These are direct empirical measurements on held-out data, not derivations that reduce by construction to fitted inputs or self-citations. The hierarchical refinement loop is described as an iterative editing process but is not used to define or validate the final performance numbers; no equations, uniqueness theorems, or ansatzes are invoked that loop back to the pipeline outputs themselves. The work is therefore self-contained against external benchmarks.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The approach rests on the domain assumption that current LLM agents can generate accurate multi-scale video descriptions and CoT traces at sufficient quality for downstream training; no free parameters or invented entities are explicitly introduced in the abstract.

axioms (1)
  • domain assumption LLM-based agents can reliably synthesize accurate Multi-Scale Spatio-Temporal Event Descriptions and downstream Q&A from raw video without human oversight at each step.
    This underpins the entire pipeline and the claim that agent-driven adaptation and refinement produce usable training data.

pith-pipeline@v0.9.0 · 5883 in / 1558 out tokens · 64625 ms · 2026-05-22T07:31:07.992453+00:00 · methodology

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

28 extracted references · 28 canonical work pages · 9 internal anchors

  1. [1]

    Agent skills

    Anthropic. Agent skills. https : / / platform . claude . com / docs / en / agents - and - tools / agent-skills/overview, 2025. 2, 4

    work page 2025

  2. [2]

    Introducing claude opus 4.6

    Anthropic. Introducing claude opus 4.6. https://www. anthropic.com/news/claude-opus-4-6 , 2026. 4

    work page 2026

  3. [3]

    Cosmos-Reason1: From Physical Common Sense To Embodied Reasoning

    Alisson Azzolini, Junjie Bai, Hannah Brandon, Jiaxin Cao, Prithvijit Chattopadhyay, Huayu Chen, Jinju Chu, Yin Cui, Jenna Diamond, Yifan Ding, et al. Cosmos-reason1: From physical common sense to embodied reasoning.arXiv preprint arXiv:2503.15558, 2025. 1, 2, 5

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  4. [4]

    Qwen3-VL Technical Report

    Shuai Bai, Yuxuan Cai, Ruizhe Chen, Keqin Chen, Xionghui Chen, Zesen Cheng, Lianghao Deng, Wei Ding, Chang Gao, Chunjiang Ge, et al. Qwen3-vl technical report.arXiv preprint arXiv:2511.21631, 2025. 2, 8

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  5. [5]

    QVAD: A Question-Centric Agentic Framework for Efficient and Training-Free Video Anomaly Detection

    Lokman Bekit, Hamza Karim, Nghia T Nguyen, and Yasin Yilmaz. Qvad: A question-centric agentic framework for efficient and training-free video anomaly detection.arXiv preprint arXiv:2604.03040, 2026. 3

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  6. [6]

    VLM-AutoDrive: Post-Training Vision-Language Models for Safety-Critical Autonomous Driving Events

    Mohammad Qazim Bhat, Yufan Huang, Niket Agarwal, Hao Wang, Michael Woods, John Kenyon, Tsung-Yi Lin, Xi- aodong Yang, Ming-Yu Liu, and Kevin Xie. Vlm-autodrive: Post-training vision-language models for safety-critical au- tonomous driving events.arXiv preprint arXiv:2603.18178,

    work page internal anchor Pith review Pith/arXiv arXiv

  7. [7]

    Scaling rl to long videos.arXiv preprint arXiv:2507.07966, 2025

    Yukang Chen, Wei Huang, Baifeng Shi, Qinghao Hu, Han- rong Ye, Ligeng Zhu, Zhijian Liu, Pavlo Molchanov, Jan Kautz, Xiaojuan Qi, Sifei Liu, Hongxu Yin, Yao Lu, and Song Han. Scaling rl to long videos.arXiv preprint arXiv:2507.07966, 2025. 2, 5

    work page arXiv 2025

  8. [8]

    Gemini 2.5: Pushing the Frontier with Advanced Reasoning, Multimodality, Long Context, and Next Generation Agentic Capabilities

    Gheorghe Comanici, Eric Bieber, Mike Schaekermann, Ice Pasupat, Noveen Sachdeva, Inderjit Dhillon, Marcel Blistein, Ori Ram, Dan Zhang, Evan Rosen, et al. Gemini 2.5: Pushing the frontier with advanced reasoning, multimodality, long con- text, and next generation agentic capabilities.arXiv preprint arXiv:2507.06261, 2025. 2, 5

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  9. [9]

    DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning

    DeepSeek-AI. Deepseek-r1: Incentivizing reasoning capa- bility in llms via reinforcement learning.arXiv preprint arXiv:2501.12948, 2025. 5

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  10. [10]

    Nvidia nemotron nano v2 vl.arXiv preprint arXiv:2511.03929, 2025

    Amala Sanjay Deshmukh, Kateryna Chumachenko, Tuo- mas Rintamaki, Matthieu Le, Tyler Poon, Danial Mohseni Taheri, Ilia Karmanov, Guilin Liu, Jarno Seppanen, Guo Chen, et al. Nvidia nemotron nano v2 vl.arXiv preprint arXiv:2511.03929, 2025. 8

    work page arXiv 2025

  11. [11]

    Accidentbench: Benchmarking multimodal understanding and reasoning in vehicle accidents and beyond.arXiv preprint arXiv:2509.26636, 2025

    Shangding Gu, Xiaohan Wang, Donghao Ying, Haoyu Zhao, Runing Yang, Ming Jin, Boyi Li, Marco Pavone, Serena Yeung-Levy, Jun Wang, et al. Accidentbench: Benchmarking multimodal understanding and reasoning in vehicle accidents and beyond.arXiv preprint arXiv:2509.26636, 2025. 2, 4, 5

    work page arXiv 2025

  12. [12]

    Colon- bench: An agentic workflow for scalable dense lesion anno- tation in full-procedure colonoscopy videos.arXiv preprint arXiv:2603.25645, 2026

    Abdullah Hamdi, Changchun Yang, and Xin Gao. Colon- bench: An agentic workflow for scalable dense lesion anno- tation in full-procedure colonoscopy videos.arXiv preprint arXiv:2603.25645, 2026. 3

    work page arXiv 2026

  13. [13]

    Vad-r1: Towards video anomaly reasoning via perception-to-cognition chain- of-thought.arXiv preprint arXiv:2505.19877, 2025

    Chao Huang, Benfeng Wang, Jie Wen, Chengliang Liu, Wei Wang, Li Shen, and Xiaochun Cao. Vad-r1: Towards video anomaly reasoning via perception-to-cognition chain- of-thought.arXiv preprint arXiv:2505.19877, 2025. 1, 2, 5

    work page arXiv 2025

  14. [14]

    Joshi, Hanna Moazam, Heather Miller, Matei Zaharia, and Christopher Potts

    Omar Khattab, Arnav Singhvi, Paridhi Maheshwari, Zhiyuan Zhang, Keshav Santhanam, Sri Vardhamanan, Saiful Haq, Ashutosh Sharma, Thomas T. Joshi, Hanna Moazam, Heather Miller, Matei Zaharia, and Christopher Potts. Dspy: Com- piling declarative language model calls into state-of-the-art pipelines. InInternational Conference on Learning Represen- tations (IC...

    work page 2024

  15. [15]

    When single-agent with skills replace multi-agent systems and when they fail.arXiv preprint arXiv:2601.04748, 2026

    Xiaoxiao Li. When single-agent with skills replace multi-agent systems and when they fail.arXiv preprint arXiv:2601.04748, 2026. 4

    work page arXiv 2026

  16. [16]

    Surveillancevqa-589k: A benchmark for comprehensive surveillance video-language understanding with large models

    Bo Liu, Pengfei Qiao, Minhan Ma, Xuange Zhang, Yi- nan Tang, Peng Xu, Kun Liu, and Tongtong Yuan. Surveillancevqa-589k: A benchmark for comprehensive surveillance video-language understanding with large models. arXiv preprint arXiv:2505.12589, 2025. 2

    work page arXiv 2025

  17. [17]

    Nexar dashcam collision prediction dataset and challenge

    Daniel Moura, Shizhan Zhu, and Orly Zvitia. Nexar dashcam collision prediction dataset and challenge. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2583–2591, 2025. 5

    work page 2025

  18. [18]

    Cosmos-reason2-8b

    NVIDIA. Cosmos-reason2-8b. https://huggingface. co/nvidia/Cosmos-Reason2-8B, 2025. 2, 5

    work page 2025

  19. [19]

    Ryan, Josh Purtell, David Bro- man, Christopher Potts, Matei Zaharia, and Omar Khattab

    Krista Opsahl-Ong, Michael J. Ryan, Josh Purtell, David Bro- man, Christopher Potts, Matei Zaharia, and Omar Khattab. Optimizing instructions and demonstrations for multi-stage language model programs. InConference on Empirical Meth- ods in Natural Language Processing (EMNLP), 2024. 3

    work page 2024

  20. [20]

    Cadp: A novel dataset for cctv traffic camera based accident analysis

    Ankit Parag Shah, Jean-Bapstite Lamare, Tuan Nguyen-Anh, and Alexander Hauptmann. Cadp: A novel dataset for cctv traffic camera based accident analysis. In2018 15th IEEE In- ternational Conference on Advanced Video and Signal Based Surveillance (AVSS), pages 1–9, 2018. 2

    work page 2018

  21. [21]

    Real-world anomaly detection in surveillance videos

    Waqas Sultani, Chen Chen, and Mubarak Shah. Real-world anomaly detection in surveillance videos. InProceedings of the IEEE conference on computer vision and pattern recogni- tion, pages 6479–6488, 2018. 2

    work page 2018

  22. [22]

    InternVL3.5: Advancing Open-Source Multimodal Models in Versatility, Reasoning, and Efficiency

    Weiyun Wang, Zhangwei Gao, Lixin Gu, Hengjun Pu, Long Cui, Xingguang Wei, Zhaoyang Liu, Linglin Jing, Shenglong Ye, Jie Shao, et al. Internvl3. 5: Advancing open-source multimodal models in versatility, reasoning, and efficiency. arXiv preprint arXiv:2508.18265, 2025. 2

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  23. [23]

    Alpamayo-R1: Bridging Reasoning and Action Prediction for Generalizable Autonomous Driving in the Long Tail

    Yan Wang, Wenjie Luo, Junjie Bai, Yulong Cao, Tong Che, Ke Chen, Yuxiao Chen, Jenna Diamond, Yifan Ding, Wenhao Ding, et al. Alpamayo-r1: Bridging reasoning and action prediction for generalizable autonomous driving in the long tail.arXiv preprint arXiv:2511.00088, 2025. 1, 2, 5

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  24. [24]

    Le, Denny Zhou, and Xinyun Chen

    Chengrun Yang, Xuezhi Wang, Yifeng Lu, Hanxiao Liu, Quoc V . Le, Denny Zhou, and Xinyun Chen. Large lan- guage models as optimizers. InInternational Conference on Learning Representations (ICLR), 2024. 3

    work page 2024

  25. [25]

    Follow the rules: reasoning for video anomaly detection with large language models

    Yuchen Yang, Kwonjoon Lee, Behzad Dariush, Yinzhi Cao, and Shao-Yuan Lo. Follow the rules: reasoning for video anomaly detection with large language models. InEuropean Conference on Computer Vision, pages 304–322. Springer,

    work page

  26. [26]

    Panda: Towards generalist video anomaly detection via agentic ai engineer.arXiv preprint arXiv:2509.26386, 2025

    Zhiwei Yang, Chen Gao, and Mike Zheng Shou. Panda: Towards generalist video anomaly detection via agentic ai engineer.arXiv preprint arXiv:2509.26386, 2025. 3

    work page arXiv 2025

  27. [27]

    Vera: Explainable video anomaly detection via verbalized learning of vision- language models

    Muchao Ye, Weiyang Liu, and Pan He. Vera: Explainable video anomaly detection via verbalized learning of vision- language models. InProceedings of the Computer Vision and Pattern Recognition Conference, pages 8679–8688, 2025. 3

    work page 2025

  28. [28]

    DAPO: An Open-Source LLM Reinforcement Learning System at Scale

    Qiying Yu, Zheng Zhang, Ruofei Zhu, Yufeng Yuan, Xi- aochen Zuo, Yu Yue, Weinan Dai, Tiantian Fan, Gaohong Liu, Lingjun Liu, et al. Dapo: An open-source llm reinforcement learning system at scale.arXiv preprint arXiv:2503.14476,

    work page internal anchor Pith review Pith/arXiv arXiv