SlideJev: Order-Stable and Tissue-Grounded Jev Decisions for Whole-Slide Pathology
Authors: Lei Wu, Jiashuai Liu, Di Zhang, Chunze Yang, Jiusong Ge, Junbo Lu, Yongqi Liang, Wenjie Zhao, Ni Zhang, Chen Li, Zeyu Gao, Qidong Liu
Organizations: School of Computer Science and Technology, Xi’an Jiaotong University · Institute of Intelligent Medicine, Chinese Academy of Medical Sciences · University of Cambridge
Whole-slide visual question answering (VQA) offers a natural-language interface to pathology, allowing diagnostic and phenotypic questions to be answered directly from tissue slides. However, current vision-language models (VLMs) rely heavily on linguistic priors: their answers are sensitive to the order of the candidate answers, and can be reached through shortcuts that bypass the tissue evidence. To address these challenges, we introduce SlideJev, a framework that separately verifies candidate answers against the slide and learns to ground these judgments in tissue evidence. To connect parallel candidate evaluation with visual evidence learning, we adopt Jev and score each candidate by the difference between support and rejection log-likelihoods from the original language-model head. We compute these scores in parallel using separate prompts with the same format and no option-position labels, then normalize them for answer selection. To ground these scores in tissue evidence, candidate-conditioned reading and comparison guide gated updates to visual tokens across decoder layers. With the backbone frozen, we train these visual modules through the same scores using question-answer and paired-image supervision. Across three whole-slide pathology VQA benchmarks, SlideJev achieves the highest mean overall accuracy and 100% answer consistency across four option orders. Image-control experiments further show that SlideJev achieves the largest average accuracy gains from correct images over mismatched and image-free inputs across the three datasets. Our code is publicly available at https://github.com/leowu-ai/SlideJev.
Figures & tables
Figure 1: Reliability challenges and SlideJev. (a) Order sensitivity and weak visual dependence. (b) Jev links parallel verification and gated updates for stable, evidence-grounded decisions.
Figure 2: Overview of SlideJev . A shared visual source supplies CONCH patch memory and an initial visual prefix. Candidates are verified in parallel using separate, identically formatted prompts; only the first and last of K inputs are shown. At each decoder layer, candidate-conditioned reads retrieve patch evidence, and candidate responses are compared on the union of selected patches. The enlarged inset illustrates how soft gates control visual-token updates before each frozen Qwen decoder block. The Jev readout contrasts support and rejection log-likelihoods from the original language-model head to produce candidate scores.
Model
TCGA
BRACS-3
Microscopy
Diagnosis
Clinical
Overall
ACC
BA
(a) General-purpose VLMs
Gemma-4-E4B-it
48.94 −5.41
42.02 −4.28
60.67 −5.57
44.99 −4.67
49.08 −0.74
41.63 −0.72
Qwen3.6-27B
63.18 −2.58
40.58 −3.97
71.18 −2.71
48.75 −3.50
53.91 +0.97
39.31 +1.35
Qwen3.8-27B
65.57 −3.13
46.26 −0.72
71.97 −1.28
53.23 −1.45
50.78 −0.51
36.26 −1.38
GPT-6.1-sol
69.40 −1.36
53.24 −0.61
71.02 +0.96
58.88 −0.75
54.69 +1.93
43.30 +1.74
Table 1: Results on SlideBench-VQA (TCGA) and BRACS-3. Each cell reports the four-order mean (%); smaller gray numbers indicate the difference from the original-order result in percentage points (pp). TCGA columns report accuracy, as answer choices vary by question. For BRACS-3 transfer, ACC reports accuracy and BA denotes balanced accuracy, computed as the mean recall across the three classes.
Model
Histology
Biomarkers
Molecular subtype
Overall
Subtype
Grade
ER
PR
HER2
ACC
BA
(a) General-purpose VLMs
Gemma-4-E4B-it
87.04 −0.08
52.46 +0.16
20.10 −0.08
27.58 −0.49
45.93 −0.19
18.68 +0.91
42.30 +0.01
37.09 +0.36
Qwen3.6-27B
22.25 +0.25
30.12 −2.35
48.26 −2.38
36.38 −3.95
40.55 −2.00
12.98 −0.15
33.08 −1.80
37.83 −2.55
Qwen3.8-27B
22.79 −0.30
41.99 −0.68
32.03 +2.55
25.59 +0.07
42.62 −0.18
15.89 0.00
31.75 +0.20
38.00 +0.40
GPT-6.1-sol
84.44 +0.30
58.82 −0.75
80.26 −0.05
74.21 +0.31
42.40 −0.15
26.85 −0.89
58.51 −0.19
39.95 +0.02
Table 2: Results on BCNB. Each cell reports the four-order mean (%); smaller gray numbers show changes from the original-order result in percentage points (pp). Task columns and ACC report accuracy; BA denotes balanced accuracy. The two HER2 tasks are combined.
Figure 3: Answer consistency across four option orders. SlideJev maintains 100% consistency across all three benchmarks, while the compared baselines remain sensitive to option order.
Figure 4: Correct-image accuracy gains. Horizontal and vertical axes show original-order ACC gains (pp) over mismatched and image-free inputs, respectively, averaged equally across BCNB, SlideBench-VQA, and BRACS-3.
Variant
ACC
Cons.
Image gain
Wrong
No image
w/o evidence module
47.95 0.00
100.00
−0.02
−0.17
w/o comparison
55.90 0.00
100.00
+4.87
+7.78
w/o parallel inputs
55.24 −5.15
33.26
+2.37
+14.44
w/o paired supervision
53.75 0.00
100.00
−0.33
+5.60
Generative readout
46.88 0.00
99.88
+1.47
+12.82
Table 3: Ablation study of SlideJev. Results are averaged equally across the full BCNB, TCGA, and BRACS-3 evaluation cohorts. ACC: four-order mean accuracy (%); gray values: changes from original-order ACC (pp). Cons.: answer consistency (%). Image gains: original-order ACC gains (pp) over mismatched (Wrong) and image-free (No image) inputs.
Figure 5: Additional SlideChat baseline comparisons. FT denotes same-data fine-tuning. Non-voting methods report four-order mean ACC; Vote aggregates predictions across the four orders. Panel (d) reports the equal-weight mean across datasets. Dashed lines indicate SlideChat; vertical scales vary by panel.
Figure 6: Case studies of regional attention and Jev readout. (A) Lung carcinoma subtyping. (B) Glioma differential diagnosis. Heatmaps show layer-averaged attention for the predicted candidate, with six patches ranked by attention. Bars show percentage deviations from each patch’s cross-candidate mean attention; red and blue indicate above- and below-mean values. Jev scores appear below each case. Cases A and B include image-free and mismatched-image predictions, respectively.
College of Computer Science, Sichuan University · Department of Pathology and Institute of Clinical Pathology, West China Hospital, Sichuan University · Department of Computer Science, School of Computing, National University of Singapore +1