Vision-language models (VLMs) have achieved remarkable success across diverse tasks. However, concerns about their trustworthiness persist, particularly regarding tendencies to lean more on textual cues than visual evidence and the risk of producing ungrounded or fabricated responses. To address these issues, we propose Saliency-R1, a framework for improving the interpretability and faithfulness of VLMs reasoning. Specifically, we introduce a novel saliency map technique that efficiently highlights critical image regions contributing to generated tokens without additional computational overhead. This can further be extended to trace how visual information flows through the reasoning process to the final answers, revealing the alignment between the thinking process and the visual context. We use the overlap between the saliency maps and human-annotated bounding boxes as the reward function, and apply Group Relative Policy Optimization (GRPO) to align the salient parts and critical regions, encouraging models to focus on relevant areas when conduct reasoning. Experiments show Saliency-R1 improves reasoning faithfulness, interpretability, and overall task performance.
Figures & tables
Figure 1: Main motivation of this work. Different thinking processes might focus on distinct regions of an image, even if they arrive at the correct answer. Unfaithful thinking processes either focus on irrelevant parts of the image or fail to consider the image.
Figure 2: Overview of our method. (a) Illustration of saliency map techniques based on logits decomposition. (b) Illustration of attention rollout for generating saliency maps with thinking tokens as the bottleneck. (c) GRPO with saliency maps alignment reward.
COCO Caption
OpenPSG
Method
Deletion ↓
Insertion ↑
Deletion ↓
Insertion ↑
5%
15%
30%
5%
15%
30%
5% %
15%
30%
5%
15%
30%
CAM [ 97 ]
86.19
82.44
78.34
-0.10
2.47
28.02
87.29
82.50
76.56
-7.12
-2.33
16.10
Grad-CAM [ 57 ]
86.19
82.44
78.34
-0.10
2.47
28.02
87.29
82.50
76.56
-7.12
-2.33
16.10
Raw Attention
84.44
78.90
73.50
-0.12
-0.08
15.95
85.07
77.42
68.68
-7.13
-6.46
10.56
Attention Rollout [ 1 ]
84.75
71.57
58.48
-0.11
0.03
38.43
87.84
73.55
59.33
-8.38
-2.75
24.36
Table 1: Faithfulness experiment results. We use Qwen2.5-VL-3B-Instruct for experiments. Our saliency map technique achieves comparable or better faithfulness to SOTA methods regarding the deletion and insertion metrics. We leave the results on GranDf dataset in the Appendix. The best metric is bold and the second best metric is underlined .
Figure 3: Qualitative evaluation of interpretability. We present example responses and their corresponding saliency maps from the base model, the SFT-tuned model (Saliency-R1-CI), and saliency-R1. The ground-truth bounding box is highlighted in red . Due to space constraints, some nonessential parts of the model responses are omitted; the full versions are provided in the Appendix.
Method
Energy-PG (%) ↑
PG (%) ↑
Qwen2.5-VL-7B [ 6 ]
19.26
24.49
Saliency-R1-CI
23.65
29.04
Vision-R1
25.32
32.58
Saliency-R1-zero
26.27
32.58
Saliency-R1-7B
30.08
38.63
Table 3: Evaluation of Interpretability. We calculate the energy-PG and PG metric on the test set of saliency-r1-8k dataset.
Figure 4: Ablation Studies. Top : Average metrics on 9 VQA benchmarks. Bottom : Metrics on MME.
Appendix figures & tables6 assets
Supplementary material from the paper’s appendix.
Appendix
GranDf
Method
Deletion ↓
Insertion ↑
5%
15%
30%
5%
15%
30%
CAM [ 97 ]
81.71
78.21
74.14
-4.80
3.86
22.96
Grad-CAM [ 57 ]
81.71
78.21
74.14
-4.80
3.86
22.96
Raw Attention
79.72
73.02
66.58
-5.58
-2.06
15.91
Attention Rollout [ 1 ]
83.56
70.17
58.88
-4.94
1.36
25.98
Appendix
Table 4: Faithfulness experiment results. We use Qwen2.5-VL-3B-Instruct for experiments. Our saliency map technique achieves comparable or better faithfulness to SOTA methods regarding the deletion and insertion metrics. The best metric is bold and the second best metric is underlined .
Accuracy
Method
Clean
Foreground ↓
Background ↑
-
σ=25
σ=50
σ=25
σ=50
Qwen2.5-VL-7B
70.8
60.8
63.9
50.5
55.4
Saliency-R1
74.4
58.0
66.4
48.9
59.0
Appendix
Table 5: Counterfactual test results. We inject Gaussian noise with different σ to the foreground and background of the images, and prompt the model to answer visual questions.
Method
POPE σ=25
POPE σ=50
MME σ=25
MME σ=50
VQA-RAD
ReasonVQA
Qwen2.5-VL-7B
85.1
81.1
2317
2223
73.1
42.2
Saliency-R1
86.9
83.2
2360
2252
75.4
56.9
Appendix
Table 6: Robustness Benchmark. We compare Saliency-R1 with base model on several benchmarks that show the robustness and generalizability of the method. We inject Gaussian noise with different σ to the images of POPE and MME.
Method
CAM
Attention Rollout
ATTN-LRP
TAM
Ours
s/samples
2.29
3.89
4.52
5.56
2.59
Appendix
Table 7: Throughput analysis. We report the average time to generate one saliency map using the saliency-r1-8k dataset.
Method
POPE
MFP-3B
86.5
RE-ALIGN-7B
87.1
MFCD-7B
87.4
Saliency-R1-3B
87.5
Saliency-R1-7B
88.1
Appendix
Table 8: Comparison with hallucination-reduction baselines.
Figure 5: Additional examples of the saliency maps generated by our proposed saliency map techniques, and the corresponding questions and responses. The examples are generated using Saliency-R1-7B.