ZwZ2: Single-Image Evidence Connection Boosts Multimodal Reasoning Across Tasks

Introduction

Multimodal models can read text, recognize objects, and describe scenes, but they still struggle when a question requires linking information from different locations in an image. For example, given an image containing a boarding pass, a flight information screen, and service signs, a model must read the flight number, find the corresponding gate, and then identify the service icon at that gate. Each local detail may be clear, but the final answer depends on correctly chaining these clues.

Similar requirements appear in spatial reasoning, multi-image understanding, and multimodal agents: recognizing information is only the first step; the model must also judge which entity each piece of information belongs to, what the relationships are, and which pieces of evidence jointly support the answer. This capability is termed connecting evidence .

EvidenceConnect Data Synthesis Framework

Researchers from Shanghai Jiao Tong University, Beijing Zhongguancun Academy, Ant Group, and other institutions propose EvidenceConnect , an automated data synthesis pipeline that creates training samples requiring a model to find, associate, and integrate visual evidence within a single image.

Region Selection and Question Generation

The framework first locates candidate regions in a high-resolution image, then selects 3–7 regions that can jointly support a question. Regions are organized around a single task, such as comparing multiple objects, aggregating quantities from different locations, or following a chain of clues.

These regions are cropped into high-definition views that preserve relative spatial order, allowing a generator model to see fine-grained details and design composite questions. Three main question types are produced:

Aggregation : integrate quantities or numerical values from multiple regions to obtain a total or computed result.

Comparison and Matching : examine the full candidate set to find outliers, extreme values, or correspondences.

Dependency Chain Reasoning : use a clue in one region to decide what to inspect next, step by step, until the answer is reached.

An illustrative example chains "flight number on boarding pass" → "gate on information screen" → "service icon at that gate", forming the evidence chain T218 → C6 → plug icon.

Verification Pipeline

To ensure answer reliability and evidence necessity, EvidenceConnect applies two verification stages:

Answer Verification : Two vision-language models from different families independently answer the generated question without seeing the generator's answer or region descriptions. Only when both models reach a consistent majority opinion over multiple runs is the candidate question retained.

Evidence Verification (Drop-One Test) : Each region is masked in turn to check whether the reference answer can still be recovered. If multiple verifications show that masking a region does not change the answer, that region is deemed empirically redundant. The framework tolerates one redundant region (due to possible context overlap between adjacent crops); if two or more regions are redundant, the question is discarded.

After verification, multiple-choice conversion, and difficulty filtering, the final dataset contains 20.4K training samples .

Model Training

Using the same 20.4K samples and training configuration, the researchers apply Group Relative Policy Optimization (GRPO) — a verifiable reward-based reinforcement learning method — without supervised fine-tuning, to three base models:

Qwen3.5-4B → ZwZ2-4B

Qwen3.5-9B → ZwZ2-9B

Qwen3.6-27B → ZwZ2-27B

Experimental Results

Overall Benchmark Performance

Experiments cover six domains (fine-grained perception, spatial perception, multimodal long-context understanding, embodied spatial reasoning, multimodal agent, and general visual reasoning) and 25 benchmarks . All three ZwZ2 models exceed their respective base models on every benchmark. Average accuracy improvements:

4B : 49.41% → 54.79%

9B : 51.63% → 56.59%

27B : 55.52% → 60.51%

Cross-Task Transfer Gains

Notably, training only on single-image data yields improvements on tasks where evidence is scattered across multiple images or interaction steps:

On the multi-image long-context benchmark MileBench , ZwZ2-4B rises from 60.08% to 70.08% (+10 percentage points).

On the spatial reasoning benchmark ViewSpatial-Bench , ZwZ2-4B improves from 37.43% to 45.26% (+7.83 points).

On the embodied reasoning benchmark ERQA , ZwZ2-27B improves from 51.25% to 61.75% (+10.50 points).

On the multimodal agent benchmark AgentVista , all three model sizes show gains.

These results support the core hypothesis: learning to connect evidence within a single image transfers to tasks where evidence is distributed across multiple images and interaction steps.

Comparison with Strong Baselines

ZwZ2-27B achieves an average accuracy of 60.51% across the 25 benchmarks, outperforming GPT-5.6-Luna (54.02%), GLM-5.3-Flash (57.24%), and DeepSeek-V4.1-Flash (55.47%) under the paper's low-thinking or non-thinking mode settings, with input resolution constrained by platform limits.

Comparison with First-Generation ZwZ

A controlled study on the same Qwen3.5-4B base using the same GRPO setup compares training data sources:

ZwZ v1 data ( 73K samples ) → average accuracy 49.86% on 11 benchmarks.

EvidenceConnect data ( 20.4K samples , ~28% of v1 size) → average accuracy 53.86% , higher on 10 of 11 benchmarks .

Even with only 7K EvidenceConnect samples , average accuracy rises from the base 47.16% to 52.65%, demonstrating that organizing training questions around evidence relationships yields strong gains at smaller data scales.

Ablation Studies

Ablations on the 7K-sample scale confirm the contribution of two key design choices:

Using high-definition cropped views for data synthesis improves downstream performance.

Including the drop-one verification step further boosts average results.

Qualitative Case Studies

The paper presents concrete examples illustrating the "connecting evidence" behavior:

Cross-image consistency : Asked to find an object category present in all five images, the base model incorrectly chooses "wine glass" even though it admits some images lack it, effectively relaxing "present in all" to "present in most". ZwZ2 excludes the wine glass and correctly answers "person", which appears in every image.

Text-location binding : In a text recognition task, ZwZ2 uses the spatial description to locate the correct archway region and reads the sign text there, whereas the base model may misattribute text to the wrong location.

Complete object counting : ZwZ2 accurately counts objects across the entire image, avoiding partial counts.

Conclusion and Open Resources

ZwZ2 demonstrates a promising data-centric direction: converting the recurring evidence-organization patterns of complex tasks into easier-to-construct and verify single-image training questions. Extending this approach to video, long sequences, and interactive environments may further test the generality of "connecting evidence" as a shared training objective.

The models and the EvidenceConnect synthesis framework are open-sourced:

arXiv: 2610.37037 GitHub:

inclusionAI/Zooming-without-Zooming-v2