PaperlayerView exact revision on arXiv

Source-linked AI summary

SAFE-G: Structure-aware Faithful Evidence-guided Generation for Knowledge-based Visual Question Answering

Long Shu, Shuochen Liu, Wei Chen, Junda Lin, Zhi Zheng, Huijun Hou, Tong Xu

arXiv:2608.21796v1cs.CVcs.AI

TL;DR

Existing KB-VQA systems struggle to filter noise in structurally complex contexts and to ensure reasoning remains faithful to retrieved evidence. SAFE-G combines coarse-to-fine multimodal graph retrieval with evidence-grounded reinforcement learning, and establishes state-of-the-art performance across benchmarks.

  • Problem

    Existing retrieval-augmented KB-VQA methods struggle to aggregate complementary evidence and consistently ground reasoning in retrieved context.

  • Method

    SAFE-G combines structure-aware multimodal graph retrieval with evidence-grounded reinforcement learning that couples answer correctness to evidence usage.

  • Results

    SAFE-G establishes state-of-the-art performance across Encyclopedic-VQA, InfoSeek, and OK-VQA benchmarks.

  • Takeaways & Limitations

    SAFE-G localizes precise evidence without costly trainable re-rankers and enforces adherence to retrieved context during generation.

  • Takeaways & Limitations

    SAFE-G uses α=0.9 in all experiments, retaining a small image-similarity contribution because text summary matching is dominant but image similarity benefits robust retrieval.

Abstract

from arXiv · show

Knowledge-based Visual Question Answering (KB-VQA) aims to answer queries that necessitate reasoning over external knowledge sources beyond the visual content. Typically, current methods fuse multimodal features to retrieve external information, subsequently leveraging Multimodal Large Language Models (MLLMs) to derive answers from the retrieved evidence. However, these methods often struggle to capture structural associations within complex contexts to effectively filter noise. Furthermore, they frequently fail to ensure that the reasoning process remains strictly faithful to the retrieved evidence. To address these challenges, we propose SAFE-G, a Structure-Aware Faithful Evidence-guided Generation framework, which enables precise evidence localization and trustworthy reasoning. Specifically, we first employ a coarse-grained hybrid search fusing visual and textual modalities to recall candidate documents, and subsequently implement a structure-aware fine-grained graph retrieval that captures structural dependencies to filter noise and pinpoint precise evidence. Moreover, we introduce a reinforcement learning (RL) strategy with an evidence-grounded reward that assigns credit to correct answers only when the selected evidence is correct. This strict alignment constraint compels the model to anchor its response in the retrieved context, effectively enhancing its capability to locate evidence via multimodal features and perform faithful reasoning. Extensive experiments on the Encyclopedic-VQA and InfoSeek benchmarks demonstrate that SAFE-G outperforms prior methods by a margin of 8.9% and 3.5%, substantially enhancing the overall reasoning accuracy. Our source code is publicly available at: https://github.com/MINE-USTC/SAFE-G.

1 Introduction

KB-VQA combines visual understanding with external knowledge, but existing systems struggle to aggregate structurally related evidence and keep generation grounded in retrieved context. SAFE-G couples structure-aware multimodal retrieval with evidence-grounded reinforcement learning, achieving state-of-the-art performance on Encyclopedic-VQA and InfoSeek.

  • KB-VQA integrates visual content with external knowledge to answer questions requiring information beyond image pixels.
  • Existing retrieval-augmented frameworks struggle to aggregate complementary evidence across passages and to ensure reasoning remains grounded in retrieved context.
  • SAFE-G couples structure-aware multimodal retrieval with evidence-grounded generation to address retrieval noise and unfaithful reasoning.
  • Fine-grained graph retrieval aggregates relevant evidence and localizes information without requiring a trainable re-ranker.
  • Evidence-grounded reinforcement learning couples answer correctness with evidence usage to enforce adherence to retrieved context.
  • SAFE-G achieves state-of-the-art performance on the Encyclopedic-VQA and InfoSeek benchmarks.

2 Related Work

Related work develops multimodal retrieval and reinforcement-learning approaches for knowledge-intensive visual reasoning. SAFE-G differs by using training-free graph propagation with visual-semantic signals to connect evidence fragments and filter noise structurally.

  • KB-VQA research has expanded beyond visual-only VQA because MLLMs struggle with knowledge-intensive queries requiring precise information beyond pre-training.
  • Visual-language retrieval methods align multimodal queries with heterogeneous documents through fine-grained cross-modal interactions, object-level features, or reflective relevance assessment.
  • SAFE-G uses a multimodal structure-aware graph mechanism with training-free Personalized PageRank to aggregate structurally central and visually consistent evidence.
  • RL-based multimodal methods can improve reasoning without massive human labeling, while GRPO reduces computational burden by eliminating a concurrent value-function critic.

3 Preliminaries

KB-VQA requires retrieving and reasoning over relevant evidence from an external knowledge base rather than answering solely from image content. The task is evaluated through answer quality and document- or section-level retrieval recall.

  • KB-VQA answers multimodal queries by retrieving and reasoning over relevant evidence from an explicit external knowledge base.
  • The knowledge base contains Wikipedia documents with metadata, associated images, and section-level textual units combining section titles with their content.
  • Given an image-question query, a retriever selects a query-specific document subset and an MLLM predicts the answer conditioned on the query and retrieved knowledge.
  • Answer quality uses BEM for E-VQA and VQA Accuracy for InfoSeek, while retrieval quality uses document- and section-level Recall@K.

4 Method

SAFE-G combines coarse multimodal retrieval, structure-aware graph refinement, and evidence-grounded reinforcement learning for KB-VQA. The framework progressively localizes relevant evidence and couples answer correctness with evidence validity.

  • Overview: SAFE-G uses three stages: coarse multimodal retrieval, fine-grained structure-aware graph retrieval, and reinforcement learning with evidence-grounded rewards.The stages identify candidate documents, propagate relevance across related sections, and train faithful generation.
  • Coarse-grained Multimodal Hybrid Retrieval: Coarse hybrid retrieval matches query images against document summaries and images, combining visual and semantic signals to select candidate documents.Summaries provide compact document representations, while CLIP-based matching produces the fused relevance score used to retrieve top-K_d candidates.
  • Structure-aware Graph Retrieval: Structure-aware graph retrieval connects sections and entities through containment and extracted relational edges, allowing propagation to aggregate complementary evidence across documents.The schema-less heterogeneous graph links otherwise disjoint sections through shared entities.
  • Multimodal-aware Section Prior: Multimodal-aware initialization combines dense textual relevance with Stage 1 signals to guide Personalized PageRank from visually and semantically relevant anchors.Section and entity nodes receive query-dependent weights, and PPR propagates relevance through entity–relation connections.
  • Evidence-grounded Reward Design: Evidence-grounded reinforcement learning awards answer correctness only when cited evidence matches the ground-truth source, while also rewarding evidence selection and format compliance.Structured outputs expose reasoning, citations, and answers for evaluating the composite reward.

5 Experiments

The experiments use E-VQA and InfoSeek to evaluate knowledge-based visual question answering across distinct datasets and official answer-quality metrics.

  • Datasets: E-VQA contains 221K question–answer pairs, up to five images per question, and approximately 1M/13.6K/5.8K train/validation/test instances.Its retrieval corpus contains about 2M Wikipedia-derived pages with sectioned text and associated images.
  • Datasets: InfoSeek contains about 1.3M image–question–answer triplets aligned with roughly 11K Wikipedia entities or pages.Its Wikipedia-based knowledge base contains 6M entries, while experiments retrieve over a 100K-page subset.
  • Evaluation Metrics: E-VQA answer quality is measured with BEM, whereas InfoSeek uses VQA Accuracy under its standard evaluation setting.The two metrics provide benchmark-specific measures for semantic answer matching and open-ended answer evaluation.

5.2 Implementation Details

Implementation uses multimodal document retrieval, parameter-efficient generator tuning, and GRPO training under fixed resource and sampling settings.

  • Retrieval Details: Coarse retrieval first filters 40 documents by summary matching, then selects 20 candidates using a fused score after image encoding with EVA-CLIP-8B.FAISS performs the two-stage retrieval, while LLaMA-3.3-70B-Instruct supports OpenIE extraction in the fine-grained stage.
  • Generator Training Details: Qwen2.5-VL-3B/7B generators are fine-tuned with LoRA while keeping the vision transformer frozen.The LoRA configuration uses r=64 and α=64.
  • Generator Training Details: GRPO training uses a 1 × 10^-5 learning rate, cosine decay, eight rollouts per query, and a maximum generation length of 600 tokens.Training samples include 4K E-VQA instances and 4K ReflectiVA instances for InfoSeek, using eight NVIDIA RTX A6000 GPUs.

5.3 Main Results

SAFE-G achieves strong results across E-VQA and InfoSeek, with gains attributed to fine-grained evidence localization and evidence-grounded reasoning rather than model capacity alone.

  • 5.3.1 VQA Results: SAFE-G is evaluated against zero-shot MLLMs and retrieval-augmented systems including CC-VQA, ReflectiVA, and VLM-PRF.The comparison covers both zero-shot and retrieval-augmented baselines across E-VQA and InfoSeek.
  • 5.3.1 VQA Results: Zero-shot MLLMs struggle on both benchmarks, while standard RAG improves performance but does not strictly align generation with retrieved context.Recent RL-based methods address this gap only partially by optimizing final answer correctness.
  • 5.3.1 VQA Results: 45.1% on E-VQA with Qwen2.5-VL-7B surpasses CC-VQA by 3.7 points and VLM-PRF by 8.0 points.SAFE-G-3B reaches 39.2%, exceeding VLM-PRF-7B at 37.1%; on InfoSeek, gains over CC-VQA and VLM-PRF are also reported across model scales.
  • 5.3.2 Results with Oracle Documents: 35.7% for the oracle-document zero-shot baseline on E-VQA trails SAFE-G by over 47 points.The oracle setting shows that providing the correct Wikipedia page alone does not resolve noise filtering in long contexts.
  • 5.3.2 Results with Oracle Documents: 82.8% versus 72.9% gives SAFE-G a 9.9-point advantage over ReflectiVA with the same Qwen2.5-VL-7B generator.The comparison attributes the gap to evidence localization and faithful reasoning rather than generator capacity.
  • 5.3.3 Coarse-grained Hybrid Retrieval Results: Hybrid retrieval improves R@K on both datasets by combining image–image and image–summary matching.The resulting candidate set supports subsequent structure-aware refinement.
  • 5.3.3 Coarse-grained Hybrid Retrieval Results: SAFE-G’s graph retrieval is compared with EchoSight’s task-specific Q-Former reranker under the same Stage 1 candidate pool.The evaluation measures section-level recall on E-VQA’s single-hop split.

5.4 Ablation Study

Ablations show that visual priors, evidence-grounded rewards, evidence gating, and summaries each contribute to SAFE-G’s retrieval or generation performance.

  • Multimodal Prior Drives Evidence Localization: Removing image-guided graph initialization reduces E-VQA accuracy to 42.7%, a 2.4-point drop from the full model.Text-only structural connectivity is insufficient to align evidence localization with visual query context.
  • Retrieval Fusion Weight: Figure 3 evaluates Recall@10/20 under different α settings on E-VQA and InfoSeek.Pure image retrieval can be distracted by visual ambiguity, while pure summary retrieval can suffer semantic confusion.
  • Evidence-grounded Reward Enforces Faithful Reasoning: Adding the evidence-grounded reward raises performance from 34.8% without RL to 38.3%.The reward couples evidence selection with answer correctness and encourages grounding in retrieved context.
  • Evidence Gating Prevents Ungrounded Reward: Evidence gating further improves performance from 38.3% to 43.2% by conditioning accuracy credit on evidence correctness.This prevents parametrically correct but ungrounded answers from receiving accuracy reward.
  • Summaries Align Retrieval with Visual Queries: The full model reaches 45.1%, outperforming the variant without summaries by 1.9 points.Document summaries provide compact representations that better align coarse multimodal retrieval with visual queries.

5.5 Further Analysis

Further analyses examine multimodal fusion, cross-benchmark transfer, training paradigms, reward dynamics, efficiency, and RL data sensitivity. Results support combining visual and textual retrieval signals with evidence-grounded reinforcement learning, while showing diminishing returns from additional RL data.

  • Effect of Fusion Weight α: α=0.9 yields the best coarse-retrieval performance, while moving to text-only retrieval at α=1 causes a slight decline.Summary matching is dominant, but retaining a small image-similarity contribution benefits robust retrieval.
  • Generalization to OK-VQA: SAFE-G achieves 79.8% on OK-VQA with Qwen2.5-VL-7B, surpassing VLM-PRF by 2.0 points and the zero-shot model by 15.1 points.With the 3B backbone, SAFE-G reaches 70.2%, exceeding VLM-PRF by 1.6 points.
  • Inference Efficiency: SAFE-G requires 4.70 seconds per query for coarse retrieval and PPR propagation, approximately 34% higher than a representative trainable RAG baseline.The online pipeline takes 3.67 seconds for coarse retrieval and 1.03 seconds for graph propagation, while avoiding model-training costs.
  • Effectiveness of Reinforcement Learning: RL outperforms SFT by 4.7 points on E-VQA and 4.4 points on InfoSeek under identical graph-retrieved evidence.Zero-shot performance is 34.8% on E-VQA and 34.9% on InfoSeek, while SFT reaches 40.4% and 42.3%, respectively.
  • Reward Dynamics: Evidence-gated reward training suppresses accuracy feedback when section evidence is incorrect, then converges after the model learns document and section localization.The accuracy reward remains low initially because r_sec=0 suppresses it regardless of answer correctness.
  • RL Data Sensitivity: SAFE-G-7B improves from 43.5% at 2K RL samples to 45.1% at 4K, then shows diminishing returns through 45.3% at 6K.SAFE-G-3B follows a similar pattern and reaches near-peak performance at 5K samples.

5.6 Case Study

Case studies show SAFE-G correcting visually induced retrieval bias and locating evidence buried within long articles. Its graph propagation and evidence-grounded reward support answers derived from the selected sections.

  • E-VQA: Graph retrieval elevates the correct evidence after Stage 1 ranks a visually similar distractor above the gold document.Propagation across entity-linked sections enables fine-grained visual-textual consistency verification.
  • InfoSeek: PPR-based propagation identifies the InfoSeek answer in a deeply buried “Inscriptions” section by aggregating cross-passage structural cues.The evidence-grounded reward requires the model to cite that section before deriving the final answer.

6 Conclusion

SAFE-G combines structure-aware graph retrieval with faithfulness-driven reinforcement learning to address retrieval noise and ungrounded hallucinations in Knowledge-based VQA. Across three benchmarks, it establishes state-of-the-art performance over existing baselines.

  • Framework: SAFE-G integrates training-free Personalized PageRank retrieval with multimodal priors and evidence-grounded GRPO that couples answer correctness with evidence selection.This design isolates precise evidence and anchors reasoning in verified knowledge without costly re-ranker training.
  • Results: SAFE-G establishes a new state-of-the-art across Encyclopedic-VQA, InfoSeek, and OK-VQA, outperforming existing baselines by substantial margins.The conclusion reports this result across all three evaluated benchmarks.
Loading 2608.21796v1…