ODE: Strategy Data Evolution Method for Vision-Native Multimodal Deep Search Agents

This paper proposes the ODE method, which solves the problem of visual evidence reuse in multimodal search through a vision-native framework (Image Bank Reference Protocol) and addresses static training data via a closed-loop data generator (ODE). It significantly improves agent performance across 8 benchmark tests—for example, the average score of Qwen3-VL-8B increased from 24.9% to 39.0%, surpassing Gemini-2.5 Pro (37.9%).

Multimodal deep search requires agents to chain tool calls, analyze images, and perform complex reasoning, but current systems face two major bottlenecks:

1. Temporary Visual Evidence: Existing tool frameworks treat images as one-time outputs, so intermediate visual evidence cannot be reused by subsequent tools;
2. Static Training Data: Data built through fixed processes cannot adapt to the dynamic evolution of strategy capabilities, leading to resource waste.

The paper proposes a vision-native agent framework, with the core being the Image Bank Reference Protocol:

• Register images returned by tools as addressable references and store them in an "Image Bank";
• Subsequent tools can access historical images via reference IDs, enabling reuse of visual evidence in the reasoning chain;
• Solve multi-step visual reasoning problems (e.g., map annotation → analysis), avoid repeated image generation/transmission, and improve efficiency and information integrity.

ODE is a closed-loop data generator synchronized with strategy training, with the core process:

1. Current strategy rollout generates execution trajectories;
2. Analyze trajectories to identify success/failure patterns;
3. Generate targeted training data to strengthen weak links;
4. Train the strategy with new data and repeat. It supports data curation for Supervised Fine-Tuning (SFT) and Reinforcement Learning (RL), covering the full training lifecycle of agents.

In 8 multimodal deep search benchmark tests, ODE showed significant effects:

• Qwen3-VL-8B: Average score increased from 24.9% to 39.0% (+56%), surpassing Gemini-2.5 Pro (37.9%);
• Qwen3-VL-30B: Average score increased from 30.6% to 41.5%. This proves that dynamic data generation can bridge the gap in model scale.

1. Image Bank Reuse: In iterative visual optimization tasks, avoid repeated overhead and maintain information integrity;
2. Rollout Feedback Advantage: Generate data based on the actual performance of the strategy, which is more aligned with task requirements than static synthesis;
3. Dynamic Adaptation: Static data cannot adjust difficulty as the strategy improves, while ODE's round-by-round refinement avoids resource waste.

Application Prospects:

• Extend to general agent training and Vision-Language Model (VLM) optimization;
• Reduce multimodal data annotation costs;
• Support continuous learning (iterative optimization after deployment). Limitations and Directions:
• Optimize computational overhead;
• Verify open-world generalization ability;
• Improve interpretability;
• Adapt to multi-agent collaboration scenarios.