Trust Before Fusion: QIMG-7 and Source-Aware Resolution for Polluted Multimodal RAG
Introduced QIMG-7 benchmark and SATR method to enhance multimodal RAG reliability in polluted settings, improving balanced score by 11.7 points.
Key Findings
Methodology
The study introduces the QIMG-7 benchmark for evaluating multimodal retrieval pollution and proposes the Source-Aware Trust Resolution (SATR) method. SATR is training-free and compares source reliability among Parametric, Text-only, and Full-MM candidate answers to select the most trustworthy or fallback answer. This method was validated across four generator/gate stacks.
Key Results
- In the gpt-4o-mini stack, Full-MM support drops from 0.908 with clean text to 0.490 with polluted text. SATR's Field-Selector variant improves the balanced score to 0.816.
- SATR excels in handling text pollution, especially in semantic entity rewrite and image blend attacks.
- Ablation studies show explicit text-reliability modeling is the main driver of performance gains.
Significance
This research addresses the vulnerability of multimodal RAG in polluted environments. The proposed QIMG-7 benchmark and SATR method provide new evaluation and solution frameworks for academia and industry, significantly improving generation accuracy through selective trust rather than unconditional fusion.
Technical Contribution
Technical contributions include designing the QIMG-7 benchmark covering seven image attack types and proposing the training-free SATR method capable of reliability parsing and answer selection in multimodal retrieval conflicts.
Novelty
This is the first introduction of source-aware trust resolution in multimodal RAG, differing from previous text retrieval methods by addressing joint text and image pollution issues.
Limitations
- SATR's performance heavily depends on the quality of the resolver; inaccurate evidence reliability judgments may affect performance.
- Trust routing may degrade performance under polluted text in some open-weight models.
Future Work
Future work could explore more complex trust resolution models to further enhance robustness across different attack types and extend to more multimodal application scenarios.
AI Executive Summary
Multimodal retrieval-augmented generation (RAG) often relies on external evidence for knowledge-intensive tasks. However, retrieved content can be unreliable, leading to incorrect answers. Researchers have proposed the QIMG-7 benchmark and Source-Aware Trust Resolution (SATR) method to tackle the issue of multimodal retrieval pollution.
The QIMG-7 benchmark covers four datasets and seven image attack types, providing 1,760 evaluation rows. The study finds that traditional multimodal fusion methods perform poorly under polluted text, while the SATR method significantly improves generation accuracy by evaluating the source reliability of candidate answers.
The SATR method is training-free and capable of reliability parsing and answer selection in multimodal retrieval conflicts, excelling in handling text pollution. While the method relies on the quality of the resolver in some models, its contribution to enhancing the robustness of multimodal RAG is undeniable.
Deep Analysis
Background
Multimodal retrieval-augmented generation (RAG) has been widely used in knowledge-intensive QA and generation tasks. However, retrieved content may contain misinformation, especially in multimodal environments, where joint text and image pollution issues have not been fully explored.
Core Problem
The core problem faced by multimodal RAG is the potential unreliability of retrieved evidence, leading to incorrect answers. This issue is particularly pronounced in long-form generation, where early factual errors can propagate through multi-paragraph answers.
Innovation
The study introduces the QIMG-7 benchmark and Source-Aware Trust Resolution (SATR) method. QIMG-7 provides a well-controlled test environment covering various image attack types. SATR is training-free and evaluates the source reliability of candidate answers to select the most trustworthy answer.
Methodology
- �� Construct the QIMG-7 benchmark covering four datasets and seven image attack types.
- �� Introduce the SATR method, comparing source reliability among Parametric, Text-only, and Full-MM candidate answers.
- �� Achieve the best balanced score with the Field-Selector variant.
Experiments
Experiments were conducted across four generator/gate stacks using the QIMG-7 benchmark's 1,760 evaluation rows. The performance of Parametric, Text-only, and Full-MM candidate answers was compared, and ablation studies were performed.
Results
Experimental results show that the SATR method excels in handling text pollution, especially in semantic entity rewrite and image blend attacks. The Field-Selector variant improves the balanced score to 0.816.
Applications
Application scenarios include knowledge-intensive QA systems, multimodal information retrieval, and generation tasks, particularly excelling in handling unreliable evidence.
Limitations & Outlook
SATR's performance heavily depends on the quality of the resolver; inaccurate evidence reliability judgments may affect performance. Trust routing may degrade performance under polluted text in some open-weight models.
Plain Language Accessible to non-experts
Imagine you're in a kitchen preparing a big meal. You need to pick the best ingredients from a variety to cook your dishes. Multimodal retrieval-augmented generation is like choosing the right combination of ingredients, but sometimes you might get spoiled ones, which is called 'pollution.' The researchers proposed a method like a smart chef who can identify which ingredients are fresh and which need to be discarded, ensuring every dish you make is delicious.
ELI14 Explained like you're 14
Imagine you're playing a game where you need to find the right answer from different clues. Sometimes, these clues might mislead you, like a trap. The researchers invented a tool, like a superpower in the game, that helps you identify which clues are reliable and which are fake. This way, you can find the correct answer faster and win the game!
Glossary
Multimodal Retrieval-Augmented Generation (RAG)
Combines text and image retrieval to enhance the accuracy of generation models.
Used to improve the accuracy of knowledge-intensive QA.
QIMG-7 Benchmark
A test benchmark for evaluating the impact of multimodal retrieval pollution.
Tests the performance of multimodal RAG in polluted environments.
Source-Aware Trust Resolution (SATR)
A training-free method that selects answers by evaluating the source reliability of candidate answers.
Used for answer selection in multimodal retrieval conflicts.
Field-Selector
A variant of the SATR method focused on improving balanced scores.
Performed well in experiments, enhancing balanced scores.
Polluted Text
Text containing misinformation or misleading content.
Challenges the accuracy of multimodal RAG.
Open Questions Unanswered questions from this research
- 1 How to improve SATR's robustness without relying on resolver quality?
- 2 How does SATR perform in more multimodal application scenarios?
Applications
Immediate Applications
Knowledge-Intensive QA
SATR can enhance the accuracy of QA systems when handling unreliable evidence, suitable for academic research and commercial applications.
Long-term Vision
Multimodal Information Retrieval
Improving SATR could achieve higher accuracy and robustness in more multimodal application scenarios.
Abstract
Multimodal retrieval-augmented generation (RAG) is often evaluated with clean evidence, yet real retrieval can return topically relevant but unreliable content: false text and misleading images from corrupted metadata, entity swaps, typographic overlays, semantic edits, adversarial patches, blends, or style transfer. We introduce QIMG-7, a controlled benchmark for multimodal retrieval pollution in multi-sentence factual QA, spanning four datasets, seven image-attack families, and 16 paired clean/polluted regimes, for 1,760 evaluation rows per method. Across four generator/gate stacks, naive multimodal fusion is brittle: in the main gpt-4o-mini stack, Full-MM support drops from 0.908 with clean text to 0.490 with polluted text, often making Parametric fallback safer than retrieval. We propose source-aware trust resolution (SATR), a training-free approach that compares Parametric, Text-only, and Full-MM candidate answers and selects among candidate answers or falls back based on source reliability. The Field-Selector variant achieves the best balanced score, 0.816, improving over Full-MM by 11.7 points and over the Cascaded Router by 2.7 points. Ablations show that, in this text-first setting, explicit text-reliability modeling is the dominant driver of these gains. Overall, in text-first factual QA with multimodal retrieval conflict, our results support selective trust rather than unconditional fusion. Artifacts are available at https://github.com/SaadElDine/Trust_Before_Fusion.