VideoChat-R1.5: Visual Test-Time Scaling to Reinforce Multimodal Reasoning by Iterative Perception

Thank you for your valuable feedback and constructive suggestions. We appreciate your recognition that our VTTS framework's core idea of mimicking human hierarchical attention through iterative refinement on high-confidence spatio-temporal regions is conceptually sound and applicable to various scenarios, as well as your acknowledgment of our extensive experimental validation across multiple tasks including Multimodal Video Understanding, Multimodal Image Understanding, Grounded Video Question Answering, Temporal Grounding, Spatial Grounding, and Object Tracking. Below, we address each of your comments and questions in detail.

Table 1: Trade-off Between VTTS Performance and Inference Time

Questions 1: Use SOTA expert models for clue finding

As shown in Table 2, in the QA task, the clue finding method of VTTS is stronger than the traditional model. The reason is our method specifically targets implicit clues that are not directly detectable by standard perception models, which typically recognize explicit entities or actions but fail to reason about contextual or temporal dependencies. In contrast, our approach leverages the rich prior knowledge of VLMs to interpret these implicit clues and guide effective clue extraction, non-training strategies perform poorly in such cases.

Table 2: Abaltion on Clue finding models

Questions 2: Why is the performance of sft worse than that of GRPO?

While both SFT and GRPO aim to train models to identify implicit clues, GRPO’s use of L1 loss—versus SFT’s cross-entropy loss—proves more effective for continuous or numerical clues (e.g., time stamps, durations, counts). Cross-entropy treats outputs categorically, so predictions like 20.1 and 21.0 incur similar penalties when the ground truth is 20.0, even though 20.1 is semantically much closer. In contrast, L1 loss in GRPO directly penalizes small deviations, enabling fine-grained, metric-aware learning. This leads to more accurate clue extraction, especially in temporal or quantitative reasoning, where precise estimation is critical for downstream perception. As a result, GRPO consistently outperforms SFT, which can degrade performance due to its coarse, discrete supervision on inherently continuous clues.

Questions 3: Why is single-step perception not as good as multi-step perception?

During the initial step, the model samples visual information uniformly without leveraging any learned or predicted visual clues. At this stage, the model lacks sufficient guidance to focus on relevant regions, resulting in suboptimal perception. However, in subsequent iterations, the model can utilize implicit visual clues derived from its own predictions to perform non-uniform, adaptive sampling—effectively refining its focus on more informative regions. This iterative refinement allows the model to progressively improve its perception and prediction accuracy. Therefore, multiple perception steps are essential to fully exploit the learned visual patterns and achieve strong performance.

Questions 4: Single and multiple inference time

We give the inference time of VTTS model and baseline models in Table 1.

Questions 5: Cold Start

In our work, we did not employ a cold start phase before reinforcement learning, as the base Qwen model already exhibits strong instruction-following capabilities and a solid foundation in reasoning. Thanks to this strong pre-training, the model is able to generate partially correct output formats and meaningful reasoning trajectories solely through well-designed prompts, even before fine-tuning. After just a few steps of reinforcement training, the model quickly stabilizes and produces nearly perfect formatting and coherent reasoning chains. This rapid convergence suggests that a separate supervised fine-tuning (SFT) cold start phase is not necessary in our setup, and the policy initialization is sufficiently strong to enable effective reinforcement learning.

Questions 6: Visualization

We have included detailed case studies in the supplementary material, which contain visualizations of the model’s output reasoning paths. These examples illustrate how our model progressively perceives visual content, refines its reasoning, and arrives at the final prediction.

Major Weakness 1: More comparisons with other models

In Table 1, we provide more experimental results, showing that the VTTS model is ahead of other models, including the baseline (videochat-r1) based on qwen2.5vl.

Major Weakness 2: Innovations in reinforcement learning

While our approach does involve modifying the reward function, it enables a novel training paradigm for multi-step perception and reasoning, which goes beyond a simple reward redesign. Specifically, we introduce a multi-reward structure—comprising answer reward, spatiotemporal reward, and format reward. This composite reward signal guides the model to learn not only what the correct output is, but also how to arrive at it through structured, step-by-step reasoning. The integration of these rewards within the reinforcement learning framework facilitates the emergence of interpretable and controllable reasoning behavior, which is not typically achieved through standard reward setups. Thus, our work contributes a meaningful and effective advancement in the application of RL to complex reasoning tasks.

Major Weakness 3: Relationship between performance and inference time

As shown in Table 1, VTTS strikes a strong balance between efficiency and performance. On MMVet, VTTS-7B (Single Infer) achieves 1.35s inference time—comparable to LLaVA-CoT (1.41s) and MM-Eureka (1.39s), slightly above the baseline (1.08s)—but with significantly higher performance (67.2%). On VideoMME, VTTS-7B (Single Infer) takes 13.21s (vs. 11.26s baseline) yet achieves 65.2%, outperforming existing methods. Thus, VTTS improves performance with acceptable latency, offering a more effective trade-off than current long-chain inference approaches.

Major Weakness 4: Ways to continue improving performance

We show the failure case in Appendix Fig. 9. Despite the design of a reasonable reward function, the model still finds it difficult to accurately locate some spatiotemporal clues. This shows that the current training data and SFT stage are still insufficient in modeling spatiotemporal perception capabilities, and further data expansion and training strategy optimization are needed.

Minor weakness:

We thank the reviewer for the careful feedback. We acknowledge the minor errors and will correct "Fune" to "Fine" in Table 5 and "Glue" to "Clue" in Figure 3 in the final version. Additionally, we will improve the figure resolution to ensure all text is clear and legible. These revisions will be made promptly in the camera-ready copy.

We sincerely appreciate the reviewer's thoughtful and constructive feedback on our work. We are grateful for the positive recognition of our key contributions, including the reasonableness and effectiveness of our iterative perception method, the well-designed integration of reinforcement learning with spatio-temporal supervision, and the value of our proposed VTTS datasets for the research community. These encouraging comments motivate us to further improve our work and address the remaining concerns raised by the reviewer. Below, we address each of your comments and questions in detail.

Table 1: Trade-off Between VTTS Performance and Inference Time

Reply to Q1: Computational overhead of the iterative inference：

As shown in Table 1, VTTS achieves a good balance between computational overhead and performance compared to the direct output baseline and other long-chain inference methods. Specifically, VTTS-7B (Single Infer) achieves an average inference time of 1.35 seconds on MMVet, slightly higher than the direct output baseline's 1.08 seconds. This is similar to other long-chain inference methods, such as LLaVA-CoT (1.41 seconds) and MM-Eureka (1.39 seconds), and its performance (67.2%) far exceeds these methods. Furthermore, while multi-inferring increases inference time, it also improves performance by 1.1%. For the video task (VideoMME), VTTS-7B (Single Infer) achieves an inference time of 13.21 seconds, higher than the baseline's 11.26 seconds, yet achieves a performance of 65.2%, significantly outperforming other methods. Therefore, VTTS's Single Infer method improves performance while maintaining similar inference time to existing CoT methods. Meanwhile, Multi-infer achieves significant performance improvements while incurring moderate computational overhead, outperforming most existing long-chain inference methods.

Reply to Q2, W1, W2: Performance compared with other reasoning models:

As shown in Table 1, VTTS-7B achieved a performance of 68.3% on MMVet (images) and 67.1% on VideoMME (videos) in multi infer mode, significantly outperforming other post-inference training methods (such as LLaVA-CoT, Insight-V, and MM-Eureka) and test-time expansion methods. For example, LLaVA-CoT achieved a performance of 60.3% on MMVet, while Insight-V and MM-Eureka achieved performances of 64.7 and 65.9, respectively, both lower than VTTS. Furthermore, VTTS also outperformed the VideoChat-R1 model (62.4) on video tasks. This performance gain demonstrates the effectiveness of the test-time expansion strategy, indicating that VTTS has stronger reasoning capabilities for complex visual tasks.

We sincerely appreciate the reviewer's thoughtful feedback and recognition of our work's significance in scaling multimodal reasoning through iterative, adaptive computation. Thank you for positively evaluating our VTTS approach with ITP and reinforcement learning, as well as acknowledging VTTS-80K as a valuable resource with comprehensive annotations. We are grateful for your appreciation of our diverse task evaluation and quantitative improvements. Your constructive feedback has strengthened our work, and we thank you for your careful consideration. Below, we address each of your comments and questions in detail.

Table 1: Trade-off Between VTTS Performance and Inference Time

Questions 1: Why is the performance of sft worse than that of GRPO?

While both SFT and GRPO aim to train models to identify implicit clues, GRPO’s use of L1 loss—versus SFT’s cross-entropy loss—proves more effective for continuous or numerical clues (e.g., time stamps, durations, counts). Cross-entropy treats outputs categorically, so predictions like 20.1 and 21.0 incur similar penalties when the ground truth is 20.0, even though 20.1 is semantically much closer. In contrast, L1 loss in GRPO directly penalizes small deviations, enabling fine-grained, metric-aware learning. This leads to more accurate clue extraction, especially in temporal or quantitative reasoning, where precise estimation is critical for downstream perception. As a result, GRPO consistently outperforms SFT, which can degrade performance due to its coarse, discrete supervision on inherently continuous clues.

Questions 2: The right trade-off for deployment

We provide an ablation experiment in Figure 4 in the main text, discussing that as the number of perceptions increases, we find that the second round of multi inference has the greatest benefit, and the benefits thereafter gradually decrease, with almost no additional benefit after the third or fourth rounds.

Weakness 1: Use SOTA expert models for clue finding

As shown in Table 2, in the QA task, the clue finding method of VTTS is stronger than the traditional model. The reason is our method specifically targets implicit clues that are not directly detectable by standard perception models, which typically recognize explicit entities or actions but fail to reason about contextual or temporal dependencies. In contrast, our approach leverages the rich prior knowledge of VLMs to interpret these implicit clues and guide effective clue extraction, non-training strategies perform poorly in such cases.

Table 2: Abaltion on Clue finding models

Weakness 2: Inference Time of VTTS.

As shown in Table 1, our model has similar inference costs as other CoT models in single infer mode, while its performance is improved and outperforms other CoT models. In multi infer mode, although the inference time is longer, the performance is further improved.

Weakness 3: More Results on other MLLMs.

As shown in Table 1, we have conducted preliminary experiments applying VTTS to InternVL3-8B, using the same VTTS-80K training set to obtain InternVL3-8B-VTTS. The results are summarized in Table 1 below:

Table 3: Results on InternVL3

These results show that VTTS consistently improves performance on InternVL3-8B under both single and multiple inference settings. This indicates that our method is not limited to Qwen2.5-VL and can generalize to other strong MLLMs with different architectures and training paradigms.

At the same time, we found during the training process that InternVL3 is worse than Qwen in terms of instruction following ability, and the preference disappears during RL training, so the performance improvement of InternVL3 is less than that of Qwen.

Weakness 4: More Results on Vision-side TTS baselines

In Table 1, we provide more experimental results. You can see that our model has improved capabilities compared to visual reasoning models such as Insight-V and MM-Eureka. Since the base model used by V* and Visual COT is relatively weak, we did not compare them. At the same time, in Table 2, we provide a baseline for random clue selection. You can see that VTTs is more capable of selecting clues than random and expert models.

Weakness 5: Problems in writing

We thank the reviewer for the feedback. In the camera-ready version, we will carefully proofread the paper to correct inconsistent tense and plurality, ensuring improved clarity and professional presentation.

We sincerely thank the reviewers for their valuable feedback and insightful comments. We appreciate the reviewers' recognition of our novel VTTS approach for scaling perceptual computation at test time, the consistent performance improvements across diverse tasks using Qwen2.5-VL models, our high-quality VTTS-80K benchmark dataset with comprehensive multi-level annotations, and our thorough ablation studies providing meaningful insights into various components. Below, we address each of your comments and questions in detail.

Questions 1: More Results on Other MLLM

As suggested, we have conducted preliminary experiments applying VTTS to InternVL3-8B, using the same VTTS-80K training set to obtain InternVL3-8B-VTTS. The results are summarized in Table 1 below:

Table 1: Results on InternVL3

These results show that VTTS consistently improves performance on InternVL3-8B under both single and multiple inference settings. This indicates that our method is not limited to Qwen2.5-VL and can generalize to other strong MLLMs with different architectures and training paradigms.

At the same time, we found during the training process that InternVL3 is worse than Qwen in terms of instruction following ability, and the preference disappears during RL training, so the performance improvement of InternVL3 is less than that of Qwen.

Weakness 1: Gain From Visual Scaling

Table 2: Abalation Study on Visual Scaling

To disentangle the effect of visual scaling from repeated language generation, we ablate VTTS under different inference settings in Table 2. While multi-inference with text-only aggregation yields a modest gain (+0.7% over single infer), the full VTTS approach with visual resampling achieves a +1.9% improvement, outperforming text-only multi-inference by +1.2%. This clear performance gap demonstrates that the gains primarily stem from visual test-time scaling—i.e., leveraging diverse visual representations—rather than repeated language generation alone.

Weakness 2: The essence of the ROI selection process

Our ROI policy is designed to capture spatio-temporal implicit cues in response to questions, leveraging the rich priors of VLMs to interpret and localize such cues. While trained on VTTS-80K, the policy learns task-agnostic reasoning patterns—i.e., how to find relevant spatio-temporal evidence given a question—enabling generalization across diverse downstream tasks rather than overfitting to dataset-specific structures.

Weakness 3: High Inference Cost

We acknowledge that iterative perception increases inference costs. In future work, we plan to optimize engineering efficiency by caching previously computed visual tags and other inference optimizations to significantly reduce redundant computations and improve runtime performance.

Weakness 4: Quality of VTTS-80K

We acknowledge the potential limitations imposed by LLMs in the annotation pipeline. However, we employ careful filtering, multi-round verification to ensure high-quality annotations. Moreover, the spatiotemporal cues and QA of the dataset are partially derived from the original annotations, so the search for implicit spatiotemporal capabilities is not limited to LLMs.

Thank you for your reply. We provide the trade-off between inference latency and performance in the table below.

Table 1: Trade-off Between VTTS Performance and Inference Time

Our model incurs inference costs comparable to other TTS models in single inference mode, while achieving superior performance. In multi-inference mode, although the computational time increases, we observe further performance gains, demonstrating the effectiveness of our approach in leveraging additional reasoning steps. This trade-off between inference cost and accuracy provides a favorable balance, particularly in settings where performance is prioritized.