NeurIPS 2026: HKUST's Visual-Native Agent Solves Image Reuse in Multimodal Deep Search

Problem: Images Found During Search Cannot Be Reused

In multimodal deep search, an agent may find an image via reverse image search, but existing systems cannot pass that image to a subsequent tool (e.g., zoom or rotate) for closer inspection. The agent sees the image but lacks a mechanism to reference it in later tool calls.

Visual-Native Agent Framework

HKUST MMSense Lab proposes the Visual-Native Agent Harness . It equips the agent with nine tools: web search, paper search, image search, open webpage, zoom image, rotate image, and Python computation. All images — the original query image and any tool-returned images — are stored in a shared image bank . Each image receives a unique identifier (e.g., <image:0>, <image:1>). Subsequent tools can retrieve any image by its ID, enabling seamless chaining of visual operations.

Example: A snow-mountain photo is <image:0>. The agent crops a ridge ( <image:1>), feeds it to image search, and stores the returned clear photo as <image:3>. Even after browsing webpages, the agent can still pass <image:3> to the zoom tool for detail verification.

On-Policy Data Evolution (ODE)

To train agents that know when to search and when to re-examine images, the team designed On-Policy Data Evolution (ODE) . The process:

Evidence collection & question generation: The system picks an object verifiable from at least two independent sources, selects a relevant image with discernible details (legends, numbers, markers), gathers clues from webpages, and builds an evidence graph. Questions are formed from connected clues, with answer-leaking hints removed.

Rollout: The target model (the one slated for RL) receives the question and original image, then autonomously decides which tools to call. This full trajectory is a rollout . The model parameters stay fixed during data synthesis.

Analysis & rule revision: An analyzer compares the batch of rollouts against the evidence graph, diagnoses failure modes (e.g., missed items, misread legends), and proposes concrete rule changes (e.g., "quantity clues must be verified item-by-item via image search", "max visual operations per clue increased from 4 to 5", "image-bearing clue ratio adjusted from 50% to 40%"). Approved changes are written into the next round's generation rules.

Iterate up to 5 rounds: Each round uses the updated rules to generate 32 questions, runs rollouts with the same target model, and re-evaluates. The best rule set is then used to mass-produce training data.

Two types of rollouts serve different purposes:

RL data synthesis: Target model (post-SFT) performs rollouts. Its ability to discover new clues or get stuck determines question suitability.

SFT demonstration collection: Teacher model (GPT-5.2) performs rollouts. The focus is on whether the solution path is worth imitating (e.g., which zoom verified the answer, which search was redundant).

Case Study: 1948 UN Trusteeship Map

ODE started from a 1948 UN world map and the fact that the UN originally had 11 trust territories. The generated question: "What percentage of the original total do the trust territories shown on the map represent?" The denominator (11) was hidden; the model had to find it and count the shaded territories (numerator) from the map legend.

The target model missed several territories and incorrectly included Southwest Africa. The analyzer, comparing the rollout with the evidence graph, found that small-print place names and legend items lacked item-by-item visual verification. Consequently, ODE revised the rules: quantity clues require per-item image-search verification; max visual steps per clue increased from 4 to 5; image-bearing clue ratio dropped from 50% to 40% to allow more verification steps.

In the next round, the same model tackled a Baltimore harbor channel chart. It repeatedly zoomed and searched but could not identify the wharf. This revealed insufficient comparable image clues and excessive difficulty. ODE then reverted the image-bearing clue ratio from 40% back to 50%, demonstrating that rule changes can be rolled back based on new rollout evidence.

Training Data Composition

SFT demonstrations: 8,855 trajectories filtered from teacher model rollouts.

RL questions: 4,000 each for the 8B and 30B-A3B target models, generated using their respective best rule sets.

All three datasets cover 11 topics. Difficulty distribution (per generator's labels): "Hard" + "Expert" questions constitute ~29% of SFT data, ~62% of 8B RL data, and ~94% of 30B-A3B RL data. Stronger target models receive a higher proportion of difficult questions without narrowing topic coverage.

Experimental Results

Framework Gain

Using Qwen3-VL-30B as base model: standard search pipeline yields 24.8% average accuracy on eight benchmarks; switching to the visual-native agent framework raises it to 30.6%.

Training Gain

Keeping the framework fixed, SFT with ODE data lifts accuracy to 39.5%; subsequent RL reaches 41.5%. On MMSearch (which demands active evidence retrieval), the 8B model progresses from 47.3% (base) to 61.7% (SFT) to 66.0% (RL).

Commercial Models Also Benefit

Gemini 2.5 Flash improves from 32.0% to 35.55%; Gemini 2.5 Pro from 37.9% to 42.21%. GPT-5 improves on all seven benchmarks where both settings are reported.

Trained 30B Model Competitive with Commercial Models

Under the same visual-native framework, ODE-30B-A3B-RL achieves 41.53% average accuracy, only 0.68 points behind Gemini 2.5 Pro (42.21%). GPT-5 scores 44.41%.

Ablation Studies

Image Reuse Contribution

On MM-BrowseComp (MMBC), ~40% of questions reuse previously retrieved or cropped images — the highest reuse rate among the eight benchmarks. With ODE-8B-RL, disabling reuse (only the most recent image available) yields 7.59% accuracy. Allowing the last 4, 8, 16 images raises accuracy progressively; full history reaches 12.50%. The 306 recorded reuse events on MMBC break down into 191 zoom operations and 113 image-search operations, confirming that chaining zoom and search on the same image is critical.

Data Evolution Contribution

Fixed training sample counts: both groups use identical evidence collection, graph building, and question generation pipelines; only the rule-updating mechanism differs. One group keeps initial rules throughout; the other lets ODE revise rules each round.

SFT demonstrations (2,000 each): ODE group shows 42.9% of demos with ≥4 tool-returned images vs. 10.5% for fixed rules. Demos combining image operations and search rise from 35.3% to 54.8%. Training 8B models on these, ODE leads on 7 of 8 benchmarks (exception: SimpleVQA).

RL questions (4,000 each, same post-SFT 8B start): On MMBC, fixed-rule RL drops accuracy from 8.5% (post-SFT) to 5.8%; ODE RL raises it to 12.5%. ODE outperforms on all eight benchmarks. Extra compute is spent only on pre-training rollouts and analysis.

Behavioral Shifts Revealed by Tool-Use Statistics

After five evolution rounds, both SFT demos and 8B RL questions receive higher quality scores (Figure 8). Tool-use logs show divergent adaptation:

Teacher model (SFT demos): Average tool calls per question drop from 15.94 to 12.19, but tool-returned images rise from 1.25 to 3.75, and image-as-input calls increase from 1.06 to 2.12. The teacher demonstrates more verification per step.

8B target model (RL questions): Average tool calls rise from 6.14 to 15.33, returned images from 3.07 to 11.00. The same model needs more steps to solve the harder, better-calibrated questions.

Conclusion

The Visual-Native Agent Harness solves the image-reuse bottleneck by assigning persistent IDs to images in a shared bank. On-Policy Data Evolution closes the loop between agent behavior and training-data generation: rollout diagnostics drive concrete, reversible rule changes that yield demonstrably better SFT demonstrations and RL questions. The resulting 8B and 30B models achieve substantial gains on eight multimodal deep-search benchmarks, narrowing the gap with leading commercial models.

Paper: Towards On-Policy Data Evolution for Visual-Native Multimodal Deep Search Agents (NeurIPS 2026), https://arxiv.org/abs/2605.10832. Code: https://github.com/JoeYing1019/ODE.