PaperlayerView exact revision on arXiv

Source-linked AI summary

VIDEOP2R: Video Understanding from Perception to Reasoning

Yifan Jiang, Yueying Wang, Rui Zhao, Toufiq Parag, Zhimin Chen, Zhenyu Liao, Jayakrishnan Unnikrishnan

arXiv:2511.11113v2cs.CVcs.AIcs.LG

TL;DR

Video RFT methods often conflate perception and reasoning, although the processes can introduce different errors. VideoP2R separates them in CoT supervision and RL rewards, and achieves SotA on six of seven benchmarks.

  • Problem

    Video RFT remains preliminary, with process-agnostic methods conflating perception and reasoning and assigning one final reward that blurs credit assignment.

    Source: §1 · p. 1

  • Method

    VideoP2R uses a three-step pipeline to create process-aware CoT data for SFT, then PA-GRPO assigns separate perception and reasoning rewards during RL.

    Source: §1 · p. 2, §3.2.2 · p. 4

  • Results

    VideoP2R achieves SotA on six of seven benchmarks, with average-accuracy gains of 1.9%–9.1% over base models across benchmarks.

    Source: §1 · p. 2

  • Takeaways & Limitations

    Ablations verify the effectiveness of process-aware modeling and PA-GRPO across the evaluated video benchmarks.

    Source: §1 · p. 2, §6 · p. 9

    Uniform frame sampling may miss question-critical frames, and descriptions beyond the training length targets [128,320] can lead the model to compress or drop critical perceptual details.

    Source: §13.2 · pp. 18–19

Abstract

from arXiv · show

Reinforcement fine-tuning (RFT), a two-stage framework consisting of supervised fine-tuning (SFT) and reinforcement learning (RL) has shown promising results on improving reasoning ability of large language models (LLMs). Yet extending RFT to large video language models (LVLMs) remains challenging. We propose VideoP2R, a novel process-aware video RFT framework that enhances video reasoning by modeling perception and reasoning as distinct processes. In the SFT stage, we develop a three-step pipeline to generate VideoP2R-CoT-162K, a high-quality, process-aware chain-of-thought (CoT) dataset for perception and reasoning. In the RL stage, we introduce a novel process-aware group relative policy optimization (PA-GRPO) algorithm that supplies separate rewards for perception and reasoning. Extensive experiments show that VideoP2R achieves state-of-the-art (SotA) performance on six out of seven video reasoning and understanding benchmarks. Ablation studies further confirm the effectiveness of our process-aware modeling and PA-GRPO and demonstrate that model's perception output is information-sufficient for downstream reasoning. Our project page is available at https://videop2r.github.io/videop2r/.

1. Introduction

VideoP2R treats perception and reasoning as distinct processes because existing video RFT methods conflate them, leaving perception and inference errors blurred under a single reward. It combines process-aware CoT data with separate perception and reasoning rewards, and reports SotA performance on six of seven benchmarks.

1. Introduction

  • Existing video RFT methods conflate perception, which extracts salient visual information, with reasoning, which organizes evidence and draws inferences, despite their distinct error types.A perception error can induce a reasoning error, while a single trajectory-level reward blurs credit assignment across the processes.Source: §1 · p. 1
  • The SFT pipeline generates perception and reasoning traces, checks whether perceptions contain sufficient evidence for the answer, and filters 260K VQA pairs into VIDEOP2R-CoT-162K.The resulting dataset warms up the model for process separation during inference and initializes subsequent RL.Source: §1 · p. 2
  • PA-GRPO assigns separate perception and reasoning rewards to their corresponding output segments instead of one reward to the whole trajectory.The perception reward is LLM-judged, while the reasoning reward checks final-answer correctness.Source: §1 · p. 2
  • VideoP2R achieves SotA on six of seven video benchmarks, with average-accuracy gains of 1.9%–9.1% over base models across benchmarks.The authors also report ablations validating process-aware modeling and PA-GRPO.Source: §1 · p. 2

2. Related work

  • Earlier approaches use frozen perception preprocessors [53] [57] [68] or rigid staged designs [13] [18] [37] [55] [58], whereas VideoP2R separates perception and reasoning in an end-to-end trainable pipeline.The paper notes that frozen preprocessors cannot be improved to curb error propagation or information loss, while rigid designs can limit generalization [50] [52].Source: §2 · pp. 2–3
  • Prior video RFT work adapts GRPO and uses varied task-specific rewards [7] [14] [26] [36] [39] [52] [56] [71], but typically scores the entire output with one final reward.VideoP2R assigns separate rewards to perception and reasoning token segments for more precise error attribution.Source: §2 · p. 3

3. VIDEOP2R RFT Framework

VideoP2R follows the two-stage RFT setup, pairing a process-aware CoT dataset for SFT with PA-GRPO for RL. The framework separates perception from reasoning in both its traces and reward signals.

3. VIDEOP2R RFT Framework

  • The SFT stage trains the base model on a three-step-generated process-aware CoT dataset, followed by PA-GRPO reinforcement learning with separate perception and reasoning rewards.The framework follows the standard two-stage RFT setup.Source: §3 · p. 3

3.1. Process-aware CoT Dataset

  • The CoT template places extracted visual evidence in <observation> and reasoning and the final answer in <think> and <answer>.The model is trained to follow this format at inference.Source: §3.1 · p. 3

3.1.1. Three-Step CoT Generation Pipeline

  • The three-step pipeline generates and verifies process-aware CoT traces, then checks whether observation segments alone contain sufficient evidence for the answer.It uses Qwen2.5-VL-72B-Instruct for initial traces and Claude 3.7 Sonnet for observation sufficiency checks.Source: §3.1.1 · pp. 3–4

3.1.2. Data Construction and Statistics

  • Applying the pipeline to 260K VQA samples yields 162K high-quality process-aware CoT examples spanning multiple question types.The resulting dataset is named VIDEOP2R-CoT-162K.Source: §3.1.2 · p. 4

3.2.1. Group Relative Policy Optimization (GRPO)

  • GRPO samples a group of responses, scores them with rule-based rewards, and normalizes rewards within the group to produce a relative advantage.Its objective uses clipped policy updates and KL regularization against a fixed reference model.Source: §3.2.1 · p. 4

3.2.2. Process-Aware Group Relative Policy Optimization (PA-GRPO)

  • PA-GRPO assigns separate rewards to perception and reasoning tokens, aligning reinforcement-learning feedback with the two processes.Figure 3 illustrates the algorithm.Source: §3.2.2 · p. 4, Figure 3 · p. 4
  • PA-GRPO provides accuracy, format, and length rewards separately for perception and reasoning, with perception judged by whether its evidence supports the answer.Reasoning accuracy uses task-specific metrics; length rewards apply only when accuracy and format rewards are nonzero and segment lengths fall within target ranges.Source: §3.2.2 · p. 5

3.2.3. Process-Aware Reward Assignment

  • PA-GRPO splits rewards into perception and reasoning groups, normalizes each separately, and applies each process-specific advantage only to its corresponding tokens.The paper notes that the processes’ reward scales and distributions are not directly comparable.Source: §3.2.3 · p. 5

4. Experiment Setup

Experiments use Qwen2.5-VL-7B-Instruct and evaluate VideoP2R on seven video reasoning and understanding benchmarks. The setup compares it with video RFT approaches and other open-source models.

Training setup

  • Training uses one SFT epoch on VIDEOP2R-CoT-162K, followed by 1K PA-GRPO updates on the original 260K visual QA data.Qwen2.5-VL-7B-Instruct is the base LVLM.Source: §4 · p. 5

Benchmarks

  • Evaluation covers four video reasoning and three video understanding benchmarks spanning spatial reasoning, knowledge-intensive QA, temporal logic, and general video understanding.The listed benchmarks are VSI-Bench, VideoMMMU, MMVU, VCR-Bench, MVBench, TempCompass, and VideoMME.Source: §4 · p. 5

Baselines

  • The comparison includes recent Qwen2.5-VL-7B-based video RFT methods and other open-source models, including several 7B-scale models.Table 1 reports benchmark comparisons in percentages; missing entries indicate unreported results.Source: §4 · pp. 5–6, Table 1 · p. 6

5. Results

The results evaluate benchmark performance, component contributions, perception support for reasoning, PA-GRPO behavior, and qualitative success and failure cases.

5.1. Main Results

  • VIDEOP2R is state of the art on six of seven benchmarks, ranks second on the remaining one, and surpasses prior state of the art by 1.3% in average accuracy.Against Qwen2.5-VL, accuracy gains range from 1.9% to 9.1% across benchmarks.Source: §5.1 · p. 6

5.2. Ablation Study

  • Ablations show a 5.8% gain from combining SFT and RL; process-aware SFT and RL improve average accuracy by 2.1% and 2.3%, while removing reward components degrades results.Removing the perception reward drops VideoMME below the SFT baseline; removing separation also underperforms PA-GRPO.Source: §5.2 · pp. 6–7, Table 2 · p. 7

5.3. Effectiveness of Perception Representations

  • With text questions alone, VIDEOP2R perception segments yield 55.5% performance, exceeding the 52.9% from raw video, and outperform Qwen’s perceptions in text and video settings.The authors interpret this as evidence that VIDEOP2R perceptions capture useful visual information for downstream reasoning.Source: §5.3 · p. 7, Figure 4 · p. 7

5.4. Advantages of PA-GRPO over GRPO

  • PA-GRPO has fewer advantage-collapse samples than GRPO and lower think–answer mismatch than the RL models Video-R1 and VideoRFT.The compared single-reward RL models show mismatch rates of at least 16%, while SFT models remain at or below 5%.Source: §5.4 · p. 8, Figure 6 · p. 8

5.5. Qualitative Results of VIDEOP2R

  • The success case reaches the correct answer by describing visual cues, while the failure case identifies relevant details but lacks domain-specific knowledge for its conclusion.The failure example concerns a question requiring the molar volume of a gas.Source: Figure 5 · p. 8, §5.5 · p. 9

6. Conclusion

The conclusion presents VIDEOP2R as a process-aware video RFT framework, with a CoT dataset for SFT and separate perception and reasoning rewards for RL. Across seven benchmarks, it reports state-of-the-art performance and ablation evidence for its components.

6. Conclusion

  • VIDEOP2R models perception and reasoning as distinct processes for video understanding.The authors frame it as a process-aware RFT framework.Source: §6 · p. 9
  • A three-step CoT generation pipeline produced VIDEOP2R-CoT-162K for fine-grained SFT supervision.Source: §6 · p. 9
  • PA-GRPO provides separate perception and reasoning rewards, and experiments across seven benchmarks report state-of-the-art performance and ablation support for process-aware modeling and PA-GRPO.Source: §6 · p. 9

Supplementary Material

The supplement lists additional material on process-aware CoT generation, evaluation and ablations, perception and RL analyses, qualitative examples, and dataset composition.

Supplementary Material

7. Details of Process-Aware CoT Generation and Data Analysis

The supplementary material documents prompt use and observation-sufficiency verification, describes the VIDEOP2R-CoT-162K dataset, and analyzes its annotations and examples.

7.1. Prompt Used

  • The process-aware CoT generation prompt is used for both training and inference.Source: Figure 7 · p. 13

7.2. Data Source and Metric for CoT Verification

7.3. Data Statistic

  • VIDEOP2R-CoT-162K contains 162,062 image and video visual QA pairs with perception and reasoning annotations across multiple question types.The listed types include multiple-choice, numerical, OCR, free-form, and regression questions.Source: §7.3 · p. 13, Table 3 · p. 15

7.4. Word Count and Word Cloud Analysis

  • UMAP visualizations show separate perception and reasoning annotation clusters, while word-count and word-cloud analyses suggest comparable annotation lengths but different content focuses.Perception annotations emphasize factual, video-centric terms; reasoning annotations include reflective expressions.Source: §7.3 · p. 14, §7.4 · p. 14, Figure 9 · p. 14

7.5. Annotation Examples

  • The examples illustrate separate perception and reasoning segments: visual cues support video QA answers, while image QA combines table-number extraction with mathematical reasoning and answer checking.Source: §7.5 · p. 14, Figure 11 · p. 14, Figure 12 · p. 14

8. Experiment Set up

The experiments use a two-stage training setup and evaluate VideoP2R on seven video-understanding and reasoning benchmarks. VideoP2R boosts Qwen2.5-VL-7B and exceeds Qwen2.5-VL-72B on MVBench, while remaining behind the larger model on average.

8.1. Implementation Details

  • Training limits video input to 16 frames at 128 × 28 × 28 resolution; inference uses 32 frames at 256 × 28 × 28 and top p = 0.001, temperature = 0.01, consistent with the Qwen2.5-VLA official demo.The two-stage training uses 8× NVIDIA A100 GPUs.Source: §8.1 · p. 15

8.2. Main Table Evaluation Setup

  • The evaluation covers seven benchmarks spanning spatial, temporal, multidisciplinary, and general video understanding, using each benchmark’s specified accuracy metrics and protocols.VSI-Bench overall performance averages MRA and ACC; Video-MME accuracy is reported without subtitles.Source: §8.2 · p. 15
  • For RL sampling, vLLM uses one GPU for generation and the other seven GPUs for model updates.Source: §8.2 · p. 15
  • VideoP2R significantly boosts Qwen2.5-VL-7B and outperforms Qwen2.5-VL-72B on MVBench, although it still trails the larger model on average.Open-source model results use original-paper reports; RFT model results are evaluated locally.Source: §8.2 · pp. 15–16

9. Ablation Study on Judge Model

The judge-model ablation finds improvements across judge variants and an association between judge capability and downstream performance. A perception examination compares video, text, and augmented-text inputs, with an example showing how relevant perception details support answering without direct video access.

9. Ablation Study on Judge Model

  • All judge-model variants improve over the base model; larger judges show higher perception-judgment accuracy, and Claude 3.7 has the strongest human-label agreement and overall process-aware performance.The human comparison uses 200 labeled samples, and judge capability positively correlates with downstream performance.Source: §9 · p. 16, Table 4 · p. 17

10.1. Prompt Used and Detailed Set up

  • The perception examination compares text-only, video-plus-text, and text augmented with a perception segment from VIDEOP2R or Qwen2.5-VL-7B.Table 5 compares Qwen2.5-VL-7B with and without perception-segment augmentation across video and text modalities.Source: §10.1 · p. 16, §10.2 · p. 16, Table 5 · p. 17

10.3. Examples of Qwen Inference Output

  • In the book-packing example, VIDEOP2R’s perception captures the book count and placement, while Qwen’s misperception leads to an incorrect answer even with video input.The passage attributes the clearer, answer-relevant perception to PA-GRPO and reports that Qwen answers correctly from the VIDEOP2R perception segment without direct video access.Source: §10.3 · pp. 16–17

11. RL Training Dynamics

During RL, perception and reasoning accuracy rewards rise while format compliance remains stable, and the output segments converge toward concise, informative traces. The section also reports answer-extraction reliability and examines think–answer mismatch across benchmarks.

11. RL Training Dynamics

  • Perception and reasoning accuracy rewards rise overall, while format rewards stay near 1 and both trace segments eventually become concise yet sufficiently informative.Perception and reasoning segment lengths first increase and then decrease during training.Source: Figure 15 · p. 17, §11 · p. 17

12.1. Pilot Experiment

  • Claude 3.7 extracts answers from reasoning segments with 96.5% accuracy, close to human annotators’ 95% agreement on 400 sampled responses.Most human disagreements involved unclear reasoning and apparent guessing.Source: §12.1 · p. 17

12.2. Think-Answer Mismatch Example

  • A think–answer mismatch example shows correct reasoning in the think segment paired with an incorrect final answer, illustrating a risk of final-answer-only rewards.Source: §12.2 · p. 18, Figure 16 · p. 18

12.3. Think-Answer Mismatch Results

  • Think–Answer Mismatch rates are reported across benchmarks, with all results computed on their multiple-choice subsets.Source: §12.3 · p. 18

13. More Qualitative Results of VIDEOP2R

Qualitative examples show VIDEOP2R tracking relevant visual evidence for correct reasoning, while two failure types arise from missed key frames and demands for excessive detail. Its largest reported gains occur on more challenging benchmark splits.

13.1. Success Case

  • Success examples show VIDEOP2R tracking three appearances of yellow clothing and using gestures and background context to answer correctly.Source: §13.1 · p. 18

13.2. Failure Case

  • Uniform sampling can miss question-critical frames, while overly detailed spatial questions may exceed the training length targets of [128,320], causing dropped perceptual details and reasoning errors.More sampled frames or adaptive sampling may mitigate missed frames; a dynamic length reward is proposed for future work on detail-heavy cases.Source: §13.2 · pp. 18–19, Figure 17 · p. 18, Figure 18 · p. 18

13.3. From Base Model to VIDEOP2R: A Stepwise Capability Evolution

  • A stepwise example progresses from missing later cat locations in the base model, to misjudging durations after SFT, to faithful tracking and a correct answer after RL.The final trace also rules out carpet and nest distractors; the authors say both SFT and RL are effective and necessary in this example.Source: §13.3 · p. 19

13.4. Impact of Task Difficulty

  • Gains are largest on challenging splits: +3.66% on Interpret (VideoMMMU) and +3.24% on Medium videos (Video-MME), versus +2.31% on Identify and +1.66% on Short.The authors characterize gains on simpler tasks as moderate.Source: §13.4 · p. 19

14. Impact of Model Size and Dataset Composition

On Qwen2.5-VL-3B, VIDEOP2R raises average accuracy above both its starting point and the Qwen2.5-VL-7B baseline. Its CoT dataset also outperforms Video-R1-CoT-165K under SFT and after Video-R1’s temporal GRPO.

14. Impact of Model Size and Dataset Composition

  • Average accuracy rises from 38.4% to 53.3% on Qwen2.5-VL-3B, surpassing the Qwen2.5-VL-7B baseline at 52.9%.The comparison is reported as part of the performance results on video reasoning and understanding benchmarks.Source: §14 · p. 19, Table 8 · p. 20
  • After SFT under identical settings, VIDEOP2R-CoT-162K reaches 55.6% average accuracy versus 53.9% for Video-R1-CoT-165K.The experiment trains Qwen2.5-VL-7B separately on each dataset.Source: §14 · p. 20, Table 9 · p. 21
  • After Video-R1’s temporal GRPO, VIDEOP2R-CoT-162K reaches 57.0% average accuracy versus 56.8% for Video-R1-CoT-165K, indicating generalization to another RFT pipeline.Both SFT checkpoints receive T-GRPO under identical settings.Source: §14 · p. 20, Table 9 · p. 21
Loading 2511.11113v2…