V-Zero: Gating On-Policy Distillation with Contrastive Visual Evidence for Fine-Grained Visual Reasoning
Listen
Radio episode about this paper
Transcript
Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.
Tom: I'm Tom.
Jane: And I'm Jane, here with Lu, Meng and our in-house large language model, Lalam.
Tom: We're looking at the paper: "V-Zero: Gating On-Policy Distillation with Contrastive Visual Evidence for Fine-Grained Visual Reasoning".
Paper summary: Tom: So to wrap up this V-Zero paper, Haoxiang Sun and his team have proposed an answer label free framework for visual reasoning using contrastive evidence gating within on policy distillation. They show that by comparing positive and negative visual evidence views from a teacher side, we can gate the dense token level distillation process.
Jane: It’s about moving away from needing human annotated answer labels by using this contrastive mechanism to evaluate trajectory reliability during training. The core finding is that this method improves fine grained visual reasoning by an average of three point one points compared with a Qwen three point five base model, all while being significantly more efficient than prior methods.
Lu: The paper demonstrates that this contrasts with the limitations of standard on policy distillation, which only aligns tokens locally and can’t score the correctness of the full trajectory once a mistake is made. V-Zero addresses that by using those paired views to gate supervision.
Meng: For deployment, it means we can potentially train these models faster and cheaper because we aren't relying on massive amounts of human labeled reasoning traces for supervised fine tuning. The efficiency gains are significant, over five times faster than some supervised fine tuning methods and over ten times faster than reinforcement learning baselines.
Jane: In simpler terms, this paper shows a clever way to train an AI to reason visually by having the model learn from its own experience while using a teacher's comparison of different visual views to guide it away from mistakes without needing human answers. That’s the essence of V-Zero: Gating On-Policy Distillation with Contrastive Visual Evidence for Fine-Grained Visual Reasoning.
Conclusion: Tom: So, V-Zero is this new framework for visual reasoning that uses contrastive evidence gating to prune reasoning tokens.
Jane: It seems like they’re trying to get around needing those expensive human answer labels by using visual comparisons instead.
Tom: Exactly, it reinterprets on policy distillation as a negative-free stopgradient alignment objective. It’s about cutting out the exploration part that takes so much time with reinforcement learning.
Jane: So the core idea is building a training signal by contrasting positive and negative views of visual evidence to gate that token-level distillation process.
Tom: And they found this contrastive gating strengthens the model for rollouts where the tokens are better supported by those positive views. It discards the negative view, which is smart.
Jane: The result they showed is an average improvement of three point one points on fine-grained visual reasoning compared to a Qwen three point five base model.
Tom: That number's concrete, that’s what I like to see. And they also said this method is over five times faster than previous supervised fine tuning methods and ten times faster than reinforcement learning baselines.
Jane: It shows that you can get better reasoning performance without needing massive amounts of labeled data, which is a big deal for deployment.
Tom: True, because we’re seeing these gains from training-time visual evidence crops rather than relying on ground-truth answer labels during the final fine tuning step.
Jane: But what this really means for us is that we can build more efficient AI that still has strong visual perception across different benchmarks like HR four K and ZoomBench.
Tom: Yeah, it seems to be a solid method for improving how AI learns to see and reason about complex visuals without the heavy overhead of traditional supervision.
Jane: We’ll keep digging into those ablation studies next to see exactly where that gating mechanism provides the most benefit for different visual tasks.
Haoxiang Sun, Zhihang Yi, Langxuan Deng, Yuhao Zhou, Peiqi Jia, Jian Zhao, Li Yuan, Jiancheng Lv
Sichuan University · –Xi’an Jiaotong University · –TeleAI of China Telecom · –Peking University
cs.CV
Submitted: 2026-06-24
Updated: 2026-11-18
Comments: revised title and abstract, added more ablation studies, and updated training details and results
Code: https://github.com/eVIgroup-SCU/V-Zero
License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/
Importance score: 98/100
The gist: Fine-grained visual reasoning requires multimodal large language models (MLLMs) to identify task-relevant visual evidence and ground their reasoning in local image regions The single most informative…
Key concepts
- On-Policy Distillation (OPD)
- This is a training method where a student model learns from an older 'teacher' model by observing its behavior on actual data trajectories. V-Zero reinterprets OPD as an objective that can be improved by considering visual evidence, rather than relying solely on textual answers.
- Contrastive Evidence Gating
- This mechanism uses paired positive and negative views of a visual input to create a 'gate' for token-level distillation. It compares the log-probabilities generated under these different views to determine which parts of the visual evidence are most reliable, allowing the model to focus its learning on high-quality information.
- Trajectory-Level Evidence Score
- This score aggregates the contrastive evidence gap across an entire reasoning path (trajectory). By measuring how much better a positive visual view is compared to a negative one for different tokens in a sequence, V-Zero creates an overall advantage score that guides the distillation process.
- Answer-Label-Free Framework
- V-Zero operates without needing any pre-annotated textual answers or labels during training. It derives its learning signals purely from visual evidence comparisons, making it more efficient and flexible than methods dependent on ground-truth text.
Terminology
Summary
Fine-grained visual reasoning requires multimodal large language models (MLLMs) to identify task-relevant visual evidence and ground their reasoning in local image regions The single most informative sentence about the paper's subject and finding is V-Zero, an answer-label-free framework for visual reasoning with contrastive evidence gating.
Pruning redundant reasoning tokens
V-Zero proposes an answer-label-free framework for visual reasoning with contrastive evidence gating This approach avoids costly exploration from reinforcement learning and heavy dependence on textual supervision by using no annotated textual answer labels The framework reinterprets On-Policy Distillation (OPD) as a negative-free stop-gradient alignment objective This perspective reveals that OPD's potential is limited by the absence of trajectory-level discrimination for erroneous paths.
How the training signal is built
V-Zero builds its training signal by contrasting paired positive and negative visual evidence views to gate dense token-level distillation The student samples on-policy trajectories from the full image, while the teacher replays them with additional paired visual evidence views The teacher computes sampled-token log-probabilities under two additional views, a positive view z+ and a negative view z−. These evaluations are used to compute the teacher-side visual evidence gap, denoted as Delta (g)k = l(g)+,k − l(g)-,k This gap is aggregated into a trajectory-level evidence score p(g), which is then normalized to create a trajectory-level evidence advantage a(g). p. 3 This advantage is converted into a non-negative stopgradient contrastive evidence gate w(g). p. 3
Where the gains come from
The primary mechanism for improvement is the contrastive evidence gating, which strengthens OPD for rollouts whose tokens are better supported by the positive visual evidence view The V-Zero objective discards the negative view and distills only from the positive teacher view using this gated signal Experiments show that V-Zero improves fine-grained visual reasoning by an average of 3.1 points compared to the Qwen3. p. 15-4B base model These gains come from training-time visual evidence crops rather than ground-truth answer labels. p. 1
Efficiency and Generalization
V-Zero achieves significant efficiency gains, being more than 5× faster than previous supervised fine-tuning methods and more than 10× faster than reinforcement learning baselines The framework preserves strong generalization capabilities across multiple visual reasoning benchmarks Ablation studies confirm that removing the gate weakens perception-average performance and degrades VStar, HR-4K, and ZoomBench scores. Furthermore, increasing the group size from G = 4 to G = 8 improves the perception-average score as well as HR-4K, HR-8K, and ZoomBench. p. 5 The training step analysis shows that the perception average improves substantially after training and peaks at step 60. p. 6
Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.
Improvements for AI systems
- Contrastive Evidence Gating for On-Policy Distillation
This mechanism uses paired positive and negative visual evidence views to evaluate student-sampled trajectories
to estimate rollout reliability. This gate strengthens OPD for rollouts whose tokens are better supported by the positive visual evidence view and suppresses rollouts whose teacher support is not improved by that evidence.
- Answer-Label-Free Fine-Grained Visual Reasoning
V-Zero improves fine-grained visual reasoning without annotated textual answer labels
by using contrastive evidence gating to modulate dense token-level distillation.
This allows the system to learn from trajectories sampled by the student itself
while avoiding costly external supervision.
- Enhanced Trajectory-Level Discrimination
The framework addresses a limitation of standard OPD by estimating rollout reliability through paired positive and negative teacher-side visual evidence views and using trajectory-level contrastive evidence gates.
This ensures that the distillation process assesses whether the trajectory as a whole is approaching the correct answer,
rather than just local token corrections.
- Efficiency Gains Over Prior Methods
V-Zero achieves significant speedups, being more than 5x faster than previous supervised fine-tuning methods and more than 10x faster than reinforcement learning baselines.
This demonstrates that V-Zero provides high performance gains while cutting training cost by over 5x relative to SFT methods and over 10x relative to RL baselines.
- Improved Performance on Visual Reasoning Benchmarks
The system can improve the Qwen3.5-4B backbone by an average of 3.1 points on fine-grained visual reasoning benchmarks while preserving strong generalization.
It reaches top-tier performance among visually grounded reasoning systems, showing that teacher-side visual evidence comparisons can provide a practical training signal for visual reasoning without annotated textual answer labels.
- Scalability Through Rollout Group Size
Increasing the number of sibling rollouts from G = 4 to G = 8 improves the perception-average score and performance on benchmarks like ZoomBench, indicating that a larger rollout group provides a more informative within-prompt comparison for the trajectory-level contrastive evidence gate.
Sources
- On-Policy Distillation of Language Models: Learning from Self-Generated Mistakes
- Qwen3-VL Technical Report
- Are We on the Right Way for Evaluating Large Vision-Language Models?
- SFT Memorizes, RL Generalizes: A Comparative Study of Foundation Model Post-training
- Gemini 2.5: Pushing the Frontier with Advanced Reasoning, Multimodality, Long Context, and Next Generation Agentic Capabilities
- GRIT: Teaching MLLMs to Think with Images
- Revisiting On-Policy Distillation: Empirical Failure Modes and Simple Fixes
- DeepEyesV2: Toward Agentic Multimodal Model
- Reinforcement Learning via Self-Distillation
- Video-OPD: Efficient Post-Training of Multimodal Large Language Models for Temporal Video Grounding via On-Policy Distillation
- Rethinking On-Policy Distillation of Large Language Models: Phenomenology, Mechanism, and Recipe
- Self-Distillation Enables Continual Learning
- Thinking with Images for Multimodal Reasoning: Foundations, Methods, and Future Frontiers
- Pixel Reasoner: Incentivizing Pixel-Space Reasoning with Curiosity-Driven Reinforcement Learning
- Zooming without Zooming: Region-to-Image Distillation for Fine-Grained Multimodal Perception
- Self-Distilled RLVR
- MM-Vet: Evaluating Large Multimodal Models for Integrated Capabilities
- Thyme: Think Beyond Images
- MME-RealWorld: Could Your Multimodal LLM Challenge High-Resolution Real-World Scenarios that are Difficult for Humans?
- Self-Distilled Reasoner: On-Policy Self-Distillation for Large Language Models
Related papers
- TEA: Text Encoder Alignment for Robust Concept Erasure in Text-to-Image Models
- Do Egocentric Video-Language Models Capture Both Hand- and Object-Centric Cues?
- Aura: Consistent Multi-Subject Video Generation via VLM-Grounded Semantic Alignment
- PVRF: All-in-one Adverse Weather Removal via Prior-modulated and Velocity-constrained Rectified Flow
- DLWM: Diverse Latent World Models for Efficient Multimodal Reasoning
- Out of Sight, Still in Mind: Token Compression for Omni-LLMs