APPO: Attention-guided Perception Policy Optimization for Video Reasoning

Henghui Du; Chang Zhou; Xi Chen; Di Hu

cs.CVFeb 27, 2026

APPO: Attention-guided Perception Policy Optimization for Video Reasoning

Authors: Henghui Du, Chang Zhou, Xi Chen, Di Hu

Organizations: Gaoling School of Artificial Intelligence, Renmin University of China, Beijing · AI Technology Center, Online Video Business Unit, Tencent PCG

Abstract

Complex video reasoning, actually, relies excessively on fine-grained perception rather than on expert (e.g., Ph.D, Science)-level reasoning. Through extensive empirical observation, we have recognized the critical impact of perception. In particular, when perception ability is almost fixed, enhancing reasoning from Qwen3-8B to OpenAI-o3 yields only 0.7% performance improvement. Conversely, even minimal change in perception model scale (from 7B to 32B) boosts performance by 1.4%, indicating enhancing perception, rather than reasoning, is more critical to improve performance. Therefore, exploring how to enhance perception ability through reasoning without the need for expensive fine-grained annotation information is worthwhile. To achieve this goal, we specially propose APPO, the Attention-guided Perception Policy Optimization algorithm that leverages token-level dense rewards to improve model's fine-grained perception. The core idea behind APPO is to optimize those tokens from different responses that primarily focus on the same crucial video frame (called intra-group perception tokens). Experimental results on diverse video benchmarks and models with different scales (3/7B) demonstrate APPO consistently outperforms GRPO and DAPO (0.5%~4%). We hope our work provides a promising approach to effectively enhance model's perception abilities through reasoning in a low-cost manner, serving diverse scenarios and demands.

Figures & tables

Panel of Figure 2 — Figure 2: The Perception-Reasoning curves on SEED-Bench-R1 [ 9 ] and Perception-Test [ 10 ] benchmarks, quantifying the impact of perception vs. reasoning ability on overall performance. Each point in the curve represents the performance achieved by combining specific perception and reasoning ability. In particular, we first prompted four perception models with progressively enhanced abilities (including Qwen2.5-VL-3/7/32B [ 11 ] and Gemini-2.0-flash [ 12 ] ) to describe video content in detail. Subsequently, the other four reasoning models with varying capabilities (including Qwen3-4/8B, Qwen3-235-A22B-thinking [ 13 ] , and OpenAI-o3 [ 1 ] ) were used to think and answer questions based on the descriptions provided by each perception model, respectively, yielding $4\times 4$ cross-combination results. (a) For SEED-Bench-R1 benchmark, we evaluate on $2K$ Level-1 samples. (b) For Perception Test benchmark, we randomly select $1K$ samples from different videos for evaluation. (c) The performance comparison of GRPO, DAPO and our APPO on SEED-Bench-R1 benchmark across different scales models, demonstrating the significant improvements brought by enhanced perception.

Figure 3: The overview of APPO algorithm, which primarily consists of two core steps: attention-guided frame selection and intra-group perception tokens re-weighting. Firstly, a group of $G$ responses are divided into two sets based on the reward scores, and the final target frames are selected from these two sets based on attention weights. Building upon these frames, those tokens from different responses focusing on the same frame are assigned into a group and optimized with token-level weights.

Figure 4: The illustration of intra-group perception tokens. The tokens within the same group primarily focus on the same crucial video frame.

Method	SEED-Bench-R1				NExTGQA		Perception Test	VSI-Bench	MVBench	NExT-QA
Method	L1 (In-Dist.)	L2 (OOD)	L3 (OOD)	Avg.	mIoU	Acc@QA	Perception Test	VSI-Bench	MVBench	NExT-QA
Qwen2.5-VL-3B
Base Model	28.2	29.4	27.0	28.0	10.1	40.3	42.9	29.2	50.8	59.4
+ SFT	32.6	33.7	28.6	31.6	10.5	63.7	58.2	32.9	57.1	72.4
+ GRPO	35.3	35.7	31.0	34.0	10.3	70.7	62.1	34.8	61.5	74.2
+ DAPO	36.6	37.5	31.8	35.3	10.6	70.9	62.5	36.7	62.8	76.1

Table 1: The comparison results across SFT, GRPO, DAPO and our APPO based on diverse video benchmarks and Qwen2.5-VL-3/7B models. All experiments are conducted on the same dataset and experimental settings to ensure a fair comparison. For evaluation, each frame is processed at a resolution of

224\times 224

. The video is sampled at

1fps

while maintaining the maximum of

60

frames.

Method	Size	Training Data	SEED-Bench-R1				NExTGQA		Perception Test	VSI-Bench	MVBench	NExT-QA
Method	Size	Training Data	L1 (In-Dist.)	L2 (OOD)	L3 (OOD)	Avg.	mIoU	Acc@QA	Perception Test	VSI-Bench	MVBench	NExT-QA
Video Reasoning Models
TW-GRPO [ 8 ]	7B	$1K$	30.2	29.0	25.6	28.3	17.7	53.2	54.9	25.1	52.9	70.2
VideoChat-R1 [ 25 ]	7B	$18K$	33.3	34.4	27.8	31.8	22.0	63.2	57.7	19.9	55.8	74.5
GRPO-CARE [ 7 ]	7B	$260K$	29.9	28.3	26.0	28.1	18.0	51.5	57.1	24.0	54.6	66.4
Video-R1 [ 5 ]	7B	$260K$	30.9	32.3	28.8	30.7	14.9	64.7	57.0	35.8	63.9	76.5

Table 2: The comparison results with video reasoning models on diverse video benchmarks. Our APPO is trained on a subset of

34K

from Video-R1-

260K

dataset. All methods employ the same base model Qwen2.5-VL-7B for zero-shot evaluation on these benchmarks. For evaluation, each frame is processed at a resolution of

224\times 224

. The video is sampled at

1fps

with the maximum of

30

frames.

Panel of Figure 5 — Figure 5: The comparison of generation entropy, grad norm and reward scores during training process with GRPO and DAPO.

Figure 6: The ablation results for different attention layers.

Panel of Figure 7 — Figure 7: The ablation results for different hyperparameters in APPO based on SEED-Bench-R1 benchmark and Qwen2.5-VL-7B model.

Appendix figures & tables13 assets

Supplementary material from the paper’s appendix.

Appendix

Figure H : The attention weights of response tokens to video frames. Question: “What did the mouse write on the blackboard?". The answer mainly appears in the last part of the video. The x-axis represents the video frame index (total $34$ frames). The y-axis shows the response tokens. The darker the color, the higher the attention score.

Figure I : The illustration of four intra-group perception tokens. The same token may attend to multiple crucial frames, as shown in the red solid circles in the figure.

Figure J : The attention weights of response tokens to video frames. Question: “When the blue cat turned his head to look the second time, what was the small kitten doing?". The answer mainly appears in the middle part of the video ( $15s\sim 20s$ ). The x-axis represents the video frame index (total $34$ frames). The y-axis shows the response tokens. The darker the color, the higher the attention score.

Attention Difference	0s	0 $\sim$ 3s	3 $\sim$ 6s	6 $\sim$ 9s	9 $\sim$ 12s	12 $\sim$ 15s	$>15s$
Sample Nums	20.1%	23.5%	17.0%	12.6%	7.8%	6.2%	12.8%

Appendix

Table C: Differences in focused frames on NExTGQA test set.

Hyper-parameters	Value
Training
Batch size	16
Gradient Accumulation Steps	1
Warmup	False
Rollout Numbers	8
Rollout Temperature	1.0

Appendix

Table D : The hyperparameters used during training and evaluation.

Data Resource	Sample Nums.
LLaVA-Video	$\sim 5K$
STAR	$\sim 11K$
Perception Test	$\sim 6K$
NExT-QA	$\sim 5K$
CLEVRER	$\sim 6K$

Appendix

Table E : The detailed composition of

34K

training subset selected from Video-R1-

260K

RL training data.

Strategy	L1 (In-Dist.)	L2 (OOD)	L3 (OOD)
Hard	49.1	50.7	48.4
Soft	50.5	51.3	49.3
All	50.2	50.9	47.3

Appendix

Table F: Ablation results for different select strategies on SEED-Bench-R1 Benchmark.

Model	SEED-Bench-R1			Perception Test	MVBench
Model	L1 (In-Dist.)	L2 (OOD)	L3 (OOD)	Perception Test	MVBench
Qwen2.5-VL 3B	0.238	0.319	0.279	0.301	0.257
Qwen2.5-VL 7B	0.103	0.183	0.201	0.163	0.095

Appendix

Table G: Standard deviation results of Tab 1.

Figure K : The prompt template for temporal video grounding task.

Figure L : The prompt template for RL training.

Model Scale	Training Efficiency		Percentage
Model Scale	DAPO	APPO (Ours)	Percentage
3B (30 frames)	0.325	0.264	81%
7B (16 frames)	0.462	0.429	93%

Appendix

Table H: The training efficiency comparison between DAPO and our APPO. Training efficiency refers to the training samples processed per second.

Panel of Figure 20 — Figure M : Attention comparison between DAPO and our APPO. X-axis : video frames. Y-axis : generated tokens during reasoning.

Figure N : The visualization example from SEED-Bench-R1 Level-3 testset. The correct option is “D. turn off sink tap". Our APPO algorithm successfully paid attention to the critical information of handwashing, which made subsequent logical reasoning more coherent.

Video Reasoning

Video reasoning studies how AI systems use video evidence to infer events, actions, temporal relations, causes, and future states.

See all 210 papers

Perceive-to-Reason: Decoupling Perception and Reasoning for Fine-Grained Visual Reasoning
Jul 1, 2026Hongxing Li, Xiufeng Huang, Dingming Li +11Visual Reasoning VLM Reasoning
1Zhejiang University · 2Alibaba Group
PerceptionComp: A Video Benchmark for Complex Perception-Centric Reasoning
Mar 27, 2026Shaoxuan Li, Zhixuan Zhao, Hanze Deng +9Video Reasoning Temporal Video Understanding
Tsinghua University · Nanyang Technological University · University of Washington
Perception Before Reasoning: Dynamic Latent Reasoning for Video Understanding and Question Answering
Aug 4, 2026Haotian Xia, Zilin Xiao, Junbo Zou +2Video QA Latent Visual Reasoning
Department of Computer Science, Rice University · College of Sciences, Georgia Institute of Technology

Jul 1, 2026J/K move · Enter open · S save

Perceive-to-Reason: Decoupling Perception and Reasoning for Fine-Grained Visual Reasoning

Hongxing Li, Xiufeng Huang, Dingming Li, Wenjing Jiang+10

1Zhejiang University · 2Alibaba Group

Visual Reasoning VLM Reasoning Visually Grounded Reasoning Fine-Grained Visual Perception RL for VLMs

Open paper AskPDF arXiv

Fine-grained visual reasoning remains challenging for vision-language models, especially when small but critical visual cues are buried in high-resolution images. Existing approaches rely on repeated cropping or test-time visual search to introduce local evidence, but they typically do not explicitly distinguish perception from reasoning. In this paper, we propose Perceive-to-Reason (P2R), a unified framework that formulates fine-grained visual reasoning as a two-stage process: the model first localizes question-relevant evidence as a Perceiver, and then answers the question as a Reasoner based on the annotated image and cropped regions. To better align training with this decoupled formulation, we further introduce Perception-Reasoning Alternating GRPO (PRA-GRPO), a role-aware reinforcement learning strategy that alternates between perception-focused and reasoning-focused updates using only final-answer supervision. Built on top of Qwen3-VL-Instruct-2B/4B/8B, P2R consistently improves performance across model scales. In particular, P2R-4B achieves 93.2% on V-Star, 81.9% on HR-Bench-4K, and 80.5% on HR-Bench-8K, substantially outperforming its corresponding backbone. Further experiments show that the benefits of P2R extend beyond high-resolution benchmarks to broader multimodal reasoning tasks. These results suggest that explicitly decoupling perception from reasoning provides an effective framework for fine-grained visual reasoning.

Abstract

Video Reasoning

Explore similar work

Perceive-to-Reason: Decoupling Perception and Reasoning for Fine-Grained Visual Reasoning

PerceptionComp: A Video Benchmark for Complex Perception-Centric Reasoning

Perception Before Reasoning: Dynamic Latent Reasoning for Video Understanding and Question Answering