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.
Figures & tables
Figure 1. Comparison of reasoning behaviors under different reward schemes: (a) Optimizing solely for the final answer causes the model to overlook external documents. (b) Using separate rewards for retrieval and answering results in disconnected and inconsistent reasoning chains. (c) SAFE-G employs a holistic reward design that guides the model to learn a correct and coherent reasoning trajectory, ensuring the answer is faithfully anchored to the evidence.
Figure 2. Overview of the SAFE-G framework. Left: A three-stage retrieval-augmented generation framework tailored for the KB-VQA task. Right: The Evidence-grounded Reward Design employed during GRPO training.
Method
Model
Retriever
E-VQA
InfoSeek
Single-Hop
All
Unseen-Q
Unseen-E
All
Zero-shot MLLMs
BLIP-2 ( 24 )
Flan-T5XL
–
12.6
12.4
12.7
12.3
12.5
InstructBLIP ( 8 )
Flan-T5XL
–
11.9
12.0
8.9
7.4
8.1
LLaVA-v1.5 ( 28 )
Vicuna-7B
–
16.3
16.9
9.6
9.4
9.5
Qwen2.5-VL-3B ( 2 )
–
–
17.9
19.6
20.4
21.9
21.4
Table 1. VQA accuracy on E-VQA and InfoSeek. Results for our method are highlighted in light blue. ∗ marks entries that are incomparable due to differences in knowledge bases.
Method
Generator
E-VQA
Infoseek
Single-Hop
Un-Q
Un-E
All
Base model
Qwen2.5-VL-7B
35.7
39.1
38.0
35.2
ReflectiVA ( 7 )
Qwen2.5-VL-7B
71.3
56.1
55.9
56.0
SAFE-G (Ours)
Qwen2.5-VL-3B
63.1
56.7
55.5
56.1
Wiki-LLaVA ( 4 )
LLaMA-3.1-8B
46.8
51.2
50.6
50.9
ReflectiVA ( 7 )
LLaMA-3.1-8B
75.2
57.8
57.4
57.6
Table 2. VQA accuracy scores on E-VQA and InfoSeek with oracle Wikipedia pages.
Method
E-VQA
Infoseek
R@1
R@5
R@10
R@20
R@1
R@5
R@10
R@20
I → Doc.
13.2
27.7
35.5
41.7
45.6
67.1
73.0
77.9
I → Img.
13.4
31.8
41.9
48.8
43.7
64.9
72.8
79.6
I → Sum.
19.1
41.2
49.8
58.7
52.6
73.9
80.0
84.8
I → Hyb.
25.6
44.6
53.7
59.7
56.7
76.0
81.7
85.6
Table 3. Comparison of coarse-grained retrieval performance (R@K) on E-VQA and Infoseek datasets.
Method
S. R@1
S. R@5
S. R@10
S. R@20
EchoSight w/o train
0.0301
0.1036
0.1775
0.2901
EchoSight
0.2429
0.4034
0.4217
0.4423
SAFE-G (Ours)
0.2530
0.4020
0.4230
0.4407
Table 4. Fine-grained section retrieval performance on E-VQA. “S. R@k” denotes the recall of the top- k sections.
Model
Graph Retrieval
Reinforcement Learning
E-VQA
InfoSeek
PPR
w/ image
w/ evidence
w/ gating
w/ summary
Single-hop
Unseen-Q
Unseen-E
All
Qwen2.5-VL-7B
✓
✓
✓
✓
42.7
43.2
44.6
43.5
✓
✓
34.8
33.8
35.4
34.9
✓
✓
✓
38.3
41.5
41.9
41.7
✓
✓
✓
✓
43.2
44.7
45.3
45.1
SAFE-G(Ours)
✓
✓
✓
✓
✓
45.1
46.3
47.1
46.7
Table 5. Ablation study results on E-VQA and InfoSeek to validate the effectiveness of our model component. ✓indicates the inclusion of a specific module.
Figure 3. Retrieval performance (Recall@10/20) under different α settings on E-VQA and Infoseek datasets.
Figure 4. Training curves of SAFE-G-7B on E-VQA, showing the trajectories of racc , rsec , and overall reward r . All metrics increase synchronously and stabilize upon convergence, validating the effectiveness of our evidence-gated reward design.
Method
Retriever
Model
OK-VQA
Base model
–
Qwen2.5-VL-7B
64.7
KU-RAG ( 52 )
–
LLaVA-Next-7B
73.1
MMKB-RAG ( 27 )
PreFLMR
LLaMA-3.1-8B
65.4
VLM-PRF ( 18 )
EVA-CLIP-8B
Qwen2.5-VL-3B
68.6
SAFE-G (Ours)
EVA-CLIP-8B
Qwen2.5-VL-3B
70.2
VLM-PRF ( 18 )
EVA-CLIP-8B
Qwen2.5-VL-7B
77.8
Table 6. VQA accuracy on the OK-VQA benchmark. SAFE-G is evaluated at both 3B and 7B model scales and compared against zero-shot, RAG, and RL-based baselines.
Figure 5. Qualitative case studies on E-VQA (left) and InfoSeek (right). SAFE-G corrects retrieval bias via graph propagation and precisely localizes target evidence sections, with answers faithfully grounded in the cited evidence.
Method
Training
E-VQA
Infoseek
Single-Hop
Unseen-Q
Unseen-E
All
SAFE-G
Zero-shot
34.8
33.8
35.4
34.9
SAFE-G
SFT
40.4
41.7
42.9
42.3
SAFE-G (Ours)
RL
45.1
46.3
47.1
46.7
Table 7. Comparison of training paradigms on E-VQA and InfoSeek, isolating the contribution of RL over zero-shot inference and supervised fine-tuning (SFT).
Method
Model
2K
3K
4K
5K
6K
SAFE-G
Qwen2.5-VL-3B
37.2
38.4
39.2
39.7
39.5
SAFE-G
Qwen2.5-VL-7B
43.5
44.3
45.1
44.9
45.3
Table 8. Sensitivity of SAFE-G to RL training data size on E-VQA (Single-hop), evaluated at both 3B and 7B model scales with training set sizes ranging from 2K to 6K samples.
Knowledge-based visual question answering combines image understanding with external or learned knowledge to answer questions requiring facts beyond the image.
Knowledge-based Visual Question Answering (KB-VQA) requires Multimodal Large Language Models (MLLMs) to identify and combine fine-grained visual cues with retrieved textual evidence. However, retrieval often introduces noisy and partially relevant content, while images contain distracting visual regions, causing pretrained MLLMs to overlook the evidence that actually supports the answer. To address this, we introduce Look Twice (LoT), a training-free inference-time framework that turns the model's own internal attention into an explicit multimodal evidence-selection mechanism. LoT first leverages the model's internal attention patterns to identify query-relevant image regions and textual sentences, filters attention sinks and distracting content, and reformulates the input to explicitly highlight the selected evidence before answer generation. The method requires no parameter updates, auxiliary models, or architectural modifications. Across four KB-VQA benchmarks and ten off-the-shelf MLLMs ranging from 2B to 38B parameters, LoT improves every evaluated backbone, with average gains of up to +12.5 accuracy points. It also provides further gains when combined with established context-refinement strategies, yielding additional improvements over already refined inputs. These results establish LoT as a general and effective mechanism for enabling pretrained MLLMs to exploit available multimodal evidence more accurately. Source code is publicly available at https://aimagelab.github.io/LoT/.
Multi-modal Retrieval-Augmented Generation (RAG) has emerged as a highly effective paradigm for Knowledge-Based Visual Question Answering (KB-VQA). Despite recent advancements, prevailing methods still primarily depend on images as the retrieval key, and often overlook or misplace the role of Vision-Language Models (VLMs), thereby failing to leverage their potential fully. In this paper, we introduce WikiSeeker, a novel multi-modal RAG framework that bridges these gaps by proposing a multi-modal retriever and redefining the role of VLMs. Rather than serving merely as answer generators, we assign VLMs two specialized agents: a Refiner and an Inspector. The Refiner utilizes the capability of VLMs to rewrite the textual query according to the input image, significantly improving the performance of the multimodal retriever. The Inspector facilitates a decoupled generation strategy by selectively routing reliable retrieved context to another LLM for answer generation, while relying on the VLM's internal knowledge when retrieval is unreliable. Extensive experiments on EVQA, InfoSeek, and M2KR demonstrate that WikiSeeker achieves state-of-the-art performance, with substantial improvements in both retrieval accuracy and answer quality. Our code will be released on https://github.com/zhuyjan/WikiSeeker.
Yingjian Zhu, Xinming Wang, Kun Ding +3
School of Artificial Intelligence, University of Chinese Academy of Sciences · State Key Laboratory of Multimodal Artificial Intelligence Systems (MAIS), Institute of Automation, Chinese Academy of Sciences · School of Artificial Intelligence, University of Science and Technology Beijing
Knowledge-based Visual Question Answering (KB-VQA) requires models to answer questions by integrating visual information with external knowledge. However, retrieved knowledge is often noisy, partially irrelevant, or misaligned with the visual content, while internal model knowledge is difficult to control and interpret. Naive aggregation of these sources limits reasoning effectiveness and reduces answer accuracy. To address this, we propose MaS-VQA, a selection-driven framework that tightly couples explicit knowledge filtering with implicit knowledge reasoning. MaS-VQA first retrieves candidate passages and applies a Mask-and-Select mechanism to jointly prune irrelevant image regions and weakly relevant knowledge fragments, producing compact, high-signal multimodal knowledge . This filtered knowledge then guides the activation of internal knowledge in a constrained semantic space, enabling complementary co-modeling of explicit and implicit knowledge for robust answer prediction. Experiments on Encyclopedic-VQA and InfoSeek demonstrate consistent performance gains across multiple MLLM backbones, and ablations verify that the selection mechanism effectively reduces noise and enhances knowledge utilization.
Xianwei Mao, Kai Ye, Sheng Zhou +4
Zhejiang University, Hangzhou, China · Alibaba Group, Hangzhou, China