Zooming without Zooming v2: Connecting Evidence for Generalizable Multimodal Perception
Multimodal perception—the ability of AI systems to understand and integrate information from multiple sensory modalities, especially vision and language—is fundamental to building intelligent agents capable of complex reasoning and interaction. Two particularly challenging tasks in this domain are spatial perception and multimodal long-context understanding. Spatial perception involves recognizing how objects relate spatially within a scene, while long-context understanding requires integrating evidence across multiple images or dialogue turns.
However, training AI models to excel at these tasks is hindered by the scarcity of high-quality, targeted training data. Annotating spatial relations or multi-turn visual dialogues is expensive and error-prone, and existing synthetic data often lack the compositional complexity needed for robust generalization.
The Core Insight: Connecting Evidence
The authors of the paper propose a unifying concept called connecting evidence as the atomic capability underlying fine-grained perception, spatial perception, and long-context multimodal understanding. This capability entails:
- Identifying relevant visual facts within an image or sequence
- Binding these facts to the correct entities
- Recovering the relations among these entities
- Composing the gathered evidence into a coherent answer
By focusing on this shared ability, the authors argue that it is possible to train models more effectively and enable generalization across diverse multimodal tasks.
Method: EVIDENCECONNECT Framework
To operationalize this insight, the authors introduce EVIDENCECONNECT, an automated data synthesis framework designed to generate training examples that require connecting evidence across multiple regions within a single image. The key components of the method are:
- Multi-region question synthesis: From high-resolution, visually complex images, the system selects groups of 3 to 7 related regions. These groups support compositional questions that require aggregation, comparison, or chaining of visual evidence. For example, a question might ask to identify an object based on a chain of spatial relations or to count items that share a particular attribute.
- Privileged evidence layouts: To facilitate question generation, the framework creates zoomed-in crops of the selected regions arranged on a compact canvas. These privileged views reveal fine-grained details that are not visible at the full-image scale and are used only during data construction.
- Cross-model answer anchoring: To ensure answer reliability, multiple vision-language models (VLMs) from different families independently answer the synthesized questions. Only questions with majority agreement on the answer are retained.
- Adaptive drop-one region verification: The framework tests whether the answer depends on every selected region by masking each region in turn and verifying if the answer changes. Questions with redundant regions (whose removal does not affect the answer) are discarded to maintain evidence dependence.
- Multiple-choice conversion and difficulty filtering: Verified questions are converted into multiple-choice format with distractors, enabling reinforcement learning with verifiable rewards (RLVR). Examples that are too easy (always answered correctly by a screening model) are filtered out to maintain training signal quality.
During training, models receive the full image with colored outlines marking the selected regions but do not have access to the privileged zoomed-in views. This setup requires the model to perceive and connect evidence independently.
Experimental Evidence and Results
The authors trained three model sizes (4B, 9B, and 27B parameters) based on the Qwen vision-language models using RLVR on the synthesized EVIDENCECONNECT data alone, without supervised fine-tuning or benchmark-specific question formats. The resulting models are called ZwZ2-4B, ZwZ2-9B, and ZwZ2-27B.
They evaluated these models on a comprehensive suite of 25 benchmarks spanning six domains:
- Fine-grained perception (e.g., VStar, HR-Bench)
- Spatial perception (e.g., TreeBench, CountQA)
- Multimodal long-context understanding (e.g., MileBench, MemLens)
- Embodied spatial reasoning (e.g., ERQA, ViewSpatial-Bench)
- Multimodal agent tasks (e.g., AgentVista, ScreenSpot-Pro)
- General visual reasoning (e.g., MMStar, BabyVision)
Key findings include:
- ZwZ2 models improved over their base Qwen counterparts on all 25 benchmarks at every scale, with overall accuracy increases of approximately 5 points.
- Fine-grained and spatial perception benchmarks saw gains of 5.2 to 6.3 points on average, with some individual benchmarks improving by over 10 points.
- Despite training only on single images, ZwZ2 models showed improvements on multi-image long-context understanding and multimodal agent tasks, suggesting effective transfer of the connecting evidence capability.
- ZwZ2-27B achieved the highest overall accuracy (60.51%) compared to advanced models like GPT-5.6-Luna and GLM-5.3-Flash.
- Compared to other synthetic training datasets targeting related tasks, EVIDENCECONNECT data yielded superior improvements, even with fewer training examples.
Ablations and Analysis
The authors conducted ablation studies to assess the contributions of key components:
- Drop-one verification: Removing this step during data synthesis reduced average accuracy by about 1.6 points, indicating its importance for ensuring evidence dependence.
- Privileged views: Synthesizing questions from zoomed-in privileged views rather than original images improved average accuracy by about 1.3 points, highlighting the value of detailed visual context during question generation.
Additional analyses showed that:
- Training data scale positively influenced performance, with larger datasets yielding better results.
- ZwZ2 models demonstrated better alignment of attention weights with ground-truth evidence regions, indicating improved focus on relevant visual cues.
- Performance gains persisted across varying context lengths in long-context benchmarks, though gains diminished at the longest contexts, consistent with training on single images.
Limitations and Future Directions
The authors acknowledge two main limitations:
- The current framework synthesizes questions from single images and does not directly model temporal or cross-frame evidence spread across video sequences. Extending synthesis to multi-frame or video contexts is a natural next step.
- The experiments focus on models up to 27B parameters and on reinforcement learning fine-tuning stages. Whether the approach benefits larger models or earlier training phases such as pre-training remains to be explored.
Practical Implications
This work suggests a promising direction for improving multimodal perception by focusing on the shared capability of connecting evidence. By synthesizing compositional, multi-region questions with rigorous verification, the approach provides scalable, high-quality training data that can enhance a model’s ability to integrate dispersed visual facts.
Practitioners developing vision-language models or embodied AI agents may find EVIDENCECONNECT useful for pretraining or fine-tuning models to better handle spatial reasoning and long-context understanding, even when annotated data are scarce. The demonstrated transfer from single-image training to multi-image and interactive tasks is particularly encouraging for building more generalizable multimodal systems.
“These results establish EVIDENCECONNECT as an effective framework for supervising evidence connection and suggest that strengthening this shared capability leads to more generalizable multimodal perception.” — Authors
Sources
AI-assisted research summary. Check the original paper for complete methods, results, and limitations.