Visual agents answer questions about ultra-high-resolution photos by zooming into the relevant regions. However, existing agents often answer from common sense instead of reading the evidence. Reinforcement learning (RL) on the final answer cannot correct this, because a guess earns the same reward as reading. We propose Cross-Input Relative Policy Optimization (CrossRPO), which trains agents on contrast sets. A contrast set asks one question about several versions of an image that differ only in a small region with the evidence and have different answers, so only reading the evidence solves the whole set. CrossRPO rewards solving the whole set, and we prove that it does not reinforce a set answered from common sense. We build contrast sets for training (ViRA-Train) and evaluation (ViRA-Bench) by editing one small object in images of 21 to 120 megapixels, and we photograph further sets with real changes (ViRA-Bench-Real) to exclude editing traces. We train two Visual Reasoning Agents, ViRA-9B and ViRA-27B, with CrossRPO on only about 3,500 items of ViRA-Train each. CrossRPO achieves 63.3% set accuracy on ViRA-Bench at the maximum resolution, against 59.3% for GRPO. ViRA-9B is the most accurate open model of its size on all benchmarks, and ViRA-27B outperforms Claude Haiku 5.5 and GPT-6-Luna, both at their highest thinking level, on most benchmarks.
Figures & tables
Figure 1: Examples of ViRA-Bench and ViRA-Bench-Real. Each set shows the original photo (O, cut to 16:9) with the object in a yellow box. The insets show the object in O and the two variants (A, B), with the correct answers. Below each set are the original’s answer and the number of baseline settings, out of 15 (Appendix D.1 ), that give this answer to all three versions at 1 MP. Questions are shortened.
Figure 2: GRPO and CrossRPO. The policy πθ samples the rollouts, and the reward r is the correctness c of the answer. (a) GRPO compares the rollouts of one input s . (b) CrossRPO compares rollouts within each version sv of a set and multiplies the result by 1+λwv , where wv comes from the rewards of the other versions.
Figure 3: Statistics of ViRA-Bench and ViRA-Bench-Real. (a) Size of the 466 ViRA-Bench photos, and every ViRA-Bench-Real photo has 45.4 MP. (b) Cumulative distribution of the object’s area as a share of the photo, with dots at the medians.
Max Res.
ViRA-Bench
ViRA-Bench-Real
Model
(MP)
0.5 MP
1 MP
2 MP
4 MP
Max
0.5 MP
1 MP
2 MP
4 MP
Max
Open 7–9B models
Qwen3.5-9B
16.8
0.9 (2K)
2.2 (4K)
3.8 (6K)
7.0 (12K)
12.2 (49K)
0.0 (2K)
0.0 (3K)
0.0 (6K)
2.6 (12K)
7.7 (49K)
ZwZ2-9B
16.8
2.1 (4K)
3.8 (5K)
9.2 (8K)
13.9 (14K)
24.1 (51K)
0.0 (3K)
0.0 (5K)
0.0 (8K)
2.6 (14K)
17.9 (50K)
Vision-OPD-9B
16.8
2.1 (2K)
5.1 (3K)
8.0 (6K)
12.3 (12K)
20.5 (49K)
0.0 (2K)
0.0 (3K)
1.3 (6K)
2.6 (12K)
24.4 (49K)
Open 7–9B models + tool
Table 1: Set accuracy (%) at five pixel budgets, with the tokens per set in gray. Max : each model’s maximum resolution ( Max Res. ), which also caps the other budgets. Δ : gain of ViRA over its base model with a zoom tool. Green and red cells: ViRA-27B is more or less accurate than the proprietary model. Bold: best in each group. ViRA results are means of three runs. Tables 16 and 17 give the per-image accuracy and the ablations, and Appendix E gives the protocol.
V ∗
HR-Bench 4K
HR-Bench 8K
MME-RW-Lite
TreeBench
O3-Bench
MMStar
Hallusion
Model
1 MP
Max
1 MP
Max
1 MP
Max
1 MP
Max
1 MP
Max
1 MP
Max
Max
Max
Open 7–9B models
Qwen3.5-9B
74.9
82.2
64.9
83.2
60.2
79.2
52.8
55.5
44.2
46.7
27.2
53.3
70.8
74.1
ZwZ2-9B
82.2
90.1
72.9
87.9
65.4
84.8
57.1
61.3
49.6
48.4
26.1
59.4
76.3
78.5
Vision-OPD-9B
80.6
94.2
67.9
83.0
63.2
80.9
57.6
61.5
46.7
47.7
30.1
53.6
72.7
74.1
Open 7–9B models + tool
Table 2: General benchmarks, all models : accuracy (%) at 1 MP and at each model’s maximum resolution ( Max ). Tools, + tool, Δ and the colors are as in Table 1 . O3-Bench photos have a median of 19 MP, and V ∗ and HR-Bench photos have 3.4 to 38.8 MP. Most MMStar and HallusionBench images are below 1 MP (medians 0.18 and 0.53 MP), so we report them only at the maximum resolution. ViRA results on V ∗ and HR-Bench are means of three runs. Bold: best in each group.
Figure 4: Where the agents look on one contrast set of ViRA-Bench at 1 MP. Each tile is the agent’s last crop, or the whole photo if it received no crop. For the open models, colors show the attention of the final reply to each part of the tile, relative to a request for a general description of the image [ 82 ] , so red marks where the reply looks more than usual. Appendix H.2 gives the details and 14 other sets.
ViRA-Bench
ViRA-Bench-Real
V ∗
HR-4K
HR-8K
MME-RW
TreeBench
O3-Bench
MMStar
Hallusion
Advantage
0.5
1
2
4
Max
0.5
1
2
4
Max
1
Max
1
Max
1
Max
1
Max
1
Max
1
Max
Max
Max
CrossRPO ( λ=0.5 )
38.7
46.9
55.6
57.5
63.3
31.2
49.6
57.7
68.8
77.4
92.0
96.2
87.5
91.0
84.7
90.1
63.6
65.3
56.3
57.0
50.4
67.0
77.1
79.6
Joint term
λ=0 (per-version RLOO)
31.3
41.2
50.0
52.9
58.7
23.5
41.9
56.8
65.4
70.1
93.0
95.6
87.8
92.9
85.7
89.8
62.4
64.1
54.3
55.3
43.5
64.1
78.8
80.2
λ=1
33.4
44.2
51.3
56.1
56.6
32.9
49.6
59.0
67.1
69.2
89.5
94.6
85.4
89.8
83.4
89.1
60.2
64.0
52.8
53.6
47.8
62.3
77.4
78.3
λ=2
40.9
47.9
53.8
55.7
58.2
52.1
59.0
68.8
69.7
74.8
87.3
93.9
82.4
88.9
81.8
86.8
60.3
62.6
52.3
53.1
40.3
58.3
76.5
77.0
Table 3: Ablations : set accuracy (%) on ViRA-Bench and ViRA-Bench-Real, and accuracy (%) on the general benchmarks. The column heads are pixel budgets in MP, and Max is the maximum resolution. All models are trained from Qwen3.5-9B. Results on ViRA-Bench, ViRA-Bench-Real are means of three runs.
Base
λ=0
Pooled GRPO
Pooled RLOO
CrossRPO
ConFiQA, all three correct
58.6
81.0 ± 1.3
79.6 ± 1.8
80.2 ± 1.7
81.9 ± 3.1
strong-prior questions (493)
64.3
90.5 ± 0.4
91.5 ± 3.5
90.3 ± 1.0
91.5 ± 2.0
O correct
76.0
87.1 ± 1.7
86.4 ± 2.2
87.0 ± 1.6
87.8 ± 2.3
A correct
75.7
90.6 ± 1.2
90.5 ± 1.1
90.0 ± 1.2
91.1 ± 1.5
B correct
71.8
90.6 ± 1.5
91.0 ± 0.4
89.4 ± 1.7
91.0 ± 1.6
A or B with the original fact
2.4
0.1 ± 0.1
0.1 ± 0.1
0.2 ± 0.1
0.0 ± 0.0
Table 4: Text contrast sets (ConFiQA, K=3 , Qwen3.5-9B): accuracy (%) and the rate of answering with the original fact (lower is better). NQ-Swap and FaithEval were not trained on. Strong-prior questions: the base model answers them correctly without the passage. Each method is trained three times with different random seeds, and each run and the base model are evaluated three times. Cells give the mean and the standard deviation over the three training runs. Bold: best in each row.
All versions correct
Accuracy per version
K
λ=0
Pooled GRPO
Pooled RLOO
CrossRPO
λ=0
Pooled GRPO
Pooled RLOO
CrossRPO
3
81.0 ± 1.3
79.6 ± 1.8
80.2 ± 1.7
81.9 ± 3.1
89.4 ± 1.1
89.3 ± 1.2
88.8 ± 1.4
90.0 ± 1.8
4
79.2 ± 3.0
81.2 ± 2.4
79.0 ± 1.2
82.0 ± 3.1
89.7 ± 1.3
90.7 ± 1.0
89.4 ± 0.7
91.5 ± 1.0
8
73.4 ± 3.8
79.7 ± 0.7
76.8 ± 4.2
81.4 ± 3.5
91.0 ± 1.4
92.6 ± 0.8
91.4 ± 0.9
93.4 ± 1.3
16
77.9 ± 1.7
76.5 ± 2.3
78.7 ± 5.6
78.9 ± 1.0
93.2 ± 1.1
93.0 ± 1.4
94.5 ± 0.9
94.6 ± 0.8
Table 5: Text sets with K versions (ConFiQA, Qwen3.5-9B): accuracy (%) on the test questions at the K used in training. For K>3 , the joint term of CrossRPO uses pairs of other versions (Eq. ( 15 )). Each method is trained three times with different random seeds, and each run is evaluated three times. Cells give the mean and the standard deviation over the three training runs. Bold: best in each row and group.
Appendix figures & tables27 assets
Supplementary material from the paper’s appendix.
Appendix
Method
K
Answers
λ
Joint term (D1)
Baseline (D2)
Scale (D3)
RLOO [ 2 ]
1
–
0
–
leave-one-out
none
Dr. GRPO [ 44 ]
1
–
0
–
group mean
none
GRPO [ 55 ]
1
–
0
–
group mean
per group
NoisyRollout † [ 43 ]
1
same
0
–
pooled with noisy copies
per group
COMET [ 89 ]
2
same
0
–
pooled over versions
per group
DNA-Train [ 27 ]
2
differ
0
–
per version
per version, ℓ1
Appendix
Table 6: Prior methods within the framework of Appendix A . K , Answers and λ give the special case of the objective Eq. ( 15 ) that a method optimizes (Proposition A.41 ). (D1)–(D3) are the design choices of Appendix A.4 : how the credit for answering all versions is computed (Theorem A.23 and Corollary A.24 ), which rollouts form the baseline (Propositions A.25 , A.26 and A.27 ), and what the advantages are divided by (Proposition A.29 ). A group mean and a leave-one-out mean differ by a constant factor (Lemma A.7 ). † Not covered by Eq. ( 15 ): NoisyRollout scores rollouts sampled on a noisy copy under the clean input, and BPO rewards a rollout for agreeing or disagreeing with an anchor rollout. CRPO weights the copy by waug and matches Eq. ( 15 ) for waug=1 .
Requirement
CrossRPO
Alternative in prior work
What the alternative loses
Target
set accuracy ∏vpv for K≤3 , all-correct triples for K>3
per-version accuracy; agreement
a blind policy reaches pv=1 on one version, and agreement is maximized by blind policies; the set accuracy loses its weight as K grows (Propositions A.11 and A.12 , Corollary A.31 and Proposition A.32 )
Unbiasedness
leave-one-out within each version
division by group statistics
a bias from statistics that contain the rollout itself [ 44 ] ; CrossRPO is unbiased for every R (Theorem A.19 )
Variance (D1)
all combinations
pairing by index
covariance larger by E[\Cov(g^Πb∣O)]⪰0 ; first-order variance ratio ρK(p)→0 as K grows (Theorem A.23 and Corollary A.24 )
Variance (D2)
per-version baseline
baseline pooled over versions
an extra Δv2 from the difficulty gap between versions (Proposition A.25 )
Credit (D2)
per-version baseline and wv
baseline pooled over versions
answers from the prior get +K−1 instead of 0, and credit shifts toward easier versions (Propositions A.26 and A.27 )
Scale (D3)
one statistic per batch
standard deviation per group or set
a cost difference of any size ϵ>0 becomes the advantage −N−1 (Proposition A.29 )
Appendix
Table 7: Summary of Appendix A : each requirement, the choice CrossRPO makes, the alternative used in prior work, and what the alternative loses, with the result that shows it.
Members of a set
Shared by the members
Shortcut whose set accuracy is bounded
Effect of the joint term
Versions of one photo that differ in the evidence (contrast sets)
scene outside Ω , question
answering from the scene and common sense without inspecting the region
bounds the shortcut by a fraction K−k (Proposition A.40 (a))
Different photos with the same question and different answers [ 17 ]
question
answering from the question alone
bounds the shortcut by a fraction K−k (Proposition A.40 (a))
Transformed copies with the same answer
everything but the transform
none: the answers are not distinct
rewards answering every copy correctly, but a shortcut can do so
Independent items
nothing
none (Proposition A.40 (b))
in expectation a constant reweighting of RLOO, plus variance (Proposition A.38 )
Appendix
Table 8: Which inputs to group (Appendix A.10 ). The joint term separates reading from a shortcut only on sets whose members share the shortcut and have pairwise different answers.
Set accuracy
λ
0.5 MP
1 MP
2 MP
4 MP
Max
ViRA-Bench
0.5
38.7
46.9
55.6
57.5
63.3
0
31.3
41.2
50.0
52.9
58.7
ViRA-Bench-Real
0.5
31.2
49.6
57.7
68.8
77.4
0
23.5
41.9
56.8
65.4
70.1
Appendix
Table 9: Effect of the joint term. CrossRPO ( λ=0.5 , ViRA-9B) is compared with one group per image ( λ=0 ), and both are trained from Qwen3.5-9B for 227 steps. Top: set accuracy (%) on ViRA-Bench (762 sets) and ViRA-Bench-Real (78 photographed sets). Bottom: accuracy (%) on the general benchmarks V ∗ (191 questions) and HR-Bench (800 each). All numbers are means of three evaluation runs. As Propositions A.38 and A.40 predict, the gain appears on sets and not on single images.
Step
Model
Sets
Propose the object
Claude Sonnet 5.5
480
Gemini 3.1 Pro
49
GPT-6.1-Terra
233
Write the question and the edits
Claude Opus 5.5
529
GPT-6-Astra
233
Edit the crop
GPT-5.4 Image 2
480
Appendix
Table 10: Models in the pipeline of ViRA-Bench and the number of sets that each one handles.
ViRA-Bench
Real
ViRA-Train
Contrast sets (questions)
762
78
6,457
Image sets
544
78
4,462
Photos
466
234
3,797
Photo size (MP)
24–108
45.4
21–120
Versions per set
3
3
2–3
Region that differs
0.010%
–
0.022%
Appendix
Table 11: Dataset statistics (medians for the region and the object; the region and the object’s share are percentages of the photo, and its longer side is measured at 1 MP). Real: ViRA-Bench-Real. An image set is one object with its versions, and two questions can share one image set.
0.5 MP
1 MP
2 MP
4 MP
Max
Agent
O
A/B − O
O
A/B − O
O
A/B − O
O
A/B − O
O
A/B − O
Qwen3.5-9B + tool
80.1
− 0.3
80.6
− 0.3
84.0
− 0.3
86.5
− 0.1
85.7
+0.5
ZwZ2-9B + tool
77.8
− 0.5
78.7
+0.4
83.1
− 0.1
85.0
+1.2
85.8
− 0.1
λ=0
86.5
+0.7
89.6
+0.3
91.2
− 0.3
92.9
− 0.1
94.3
+0.4
ViRA-9B
87.6
− 0.4
89.9
+0.4
92.4
0.0
94.2
− 0.1
96.2
+0.2
Qwen3.8-27B + tool
75.5
− 1.1
81.5
− 0.3
86.9
− 0.3
89.5
− 0.3
90.6
+0.1
Appendix
Table 12: Editing traces and the first zoom on ViRA-Bench (762 sets). For each budget, the first column gives the share (%) of original (O) images whose first zoom covers the object, and the second gives the share on the variants (A, B) minus the share on O, in points. The three versions share the question and the position of the object, and only A and B are edited. ViRA and λ=0 are means of three evaluation runs. Claude Haiku 5.5 takes at most 3.8 MP, so its 4 MP columns repeat its max columns.
Source
K
ViRA-9B
ViRA-27B
Answer
Contrast sets
2–3
1,664
1,338
multiple choice
Commons detail questions
1
599
600
open
Multi-evidence, photos
1
859
949
choice or open
Multi-evidence, maps
1
170
175
choice or open
Synthetic charts (5 types)
1
200
200
open
Math, ThinkLite-VL [ 68 ]
1
150
139
open
Appendix
Table 13: Training data of ViRA-9B and ViRA-27B: number of items per source. Math: 70 geometry, 40 table and 40 arithmetic problems for ViRA-9B, and 70, 29 and 40 for ViRA-27B. K : number of versions of an item. Answer : form of its answer.
Setting
Value
Base model
Qwen3.5-9B / Qwen3.8-27B
Thinking
off
Framework
verl 0.10 [ 58 ] , vLLM 0.32 rollouts
Hardware
8 H200 GPUs
Training time
12.7 h (9B) / 29.6 h (27B)
Steps
227 (9B) / 212 (27B)
Appendix
Table 14: Training hyperparameters of ViRA.
Model
Checkpoint or API model
Training data
Ours
ViRA-9B
trained from Qwen/Qwen3.5-9B
3,642
ViRA-27B
trained from Qwen/Qwen3.8-27B
3,401
Open models
Qwen3.5-9B
Qwen/Qwen3.5-9B
–
ZwZ2-9B
inclusionAI/ZwZ2-9B
20.4K
Appendix
Table 15: Evaluated models : checkpoints, API model names and training data. Training data: the number of samples that train a model from its base model, as its paper reports it. a+b : supervised fine-tuning and RL. ≥ : only the supervised stage is reported.
Figure 5: Set accuracy vs. tokens for every model on ViRA-Bench. Each model is labeled at its last marker, and the labels of proprietary models give the thinking level. Filled markers are pixel budgets of 0.5–4 MP, hollow markers are each model’s max resolution, and dashed lines are open models without a zoom tool. Claude Haiku 5.5 has no 4 MP marker because its max resolution is 3.8 MP. Tokens are summed over all turns and versions of a set.
Figure 6: Set accuracy vs. tokens for every model on ViRA-Bench-Real. Each model is labeled at its last marker, and the labels of proprietary models give the thinking level. Filled markers are pixel budgets of 0.5–4 MP, hollow markers are each model’s max resolution, and dashed lines are open models without a zoom tool. Claude Haiku 5.5 has no 4 MP marker because its max resolution is 3.8 MP. Tokens are summed over all turns and versions of a set.
ViRA-Bench
ViRA-Bench-Real
Model
0.5 MP
1 MP
2 MP
4 MP
Max
0.5 MP
1 MP
2 MP
4 MP
Max
Open 7–9B models
Qwen3.5-9B
31.3
33.8
37.0
39.9
46.2
28.6
29.5
32.5
35.0
45.7
ZwZ2-9B
30.7
34.3
39.6
44.2
54.3
27.8
27.8
29.9
35.5
52.6
Vision-OPD-9B
32.6
36.3
40.0
44.2
52.7
27.8
28.2
32.5
35.9
56.4
Open 7–9B models + tool
Appendix
Table 16: Per-image accuracy (%), all models , on ViRA-Bench and ViRA-Bench-Real. Each version of a set counts separately. The models and the budgets are as in Table 1 . ViRA results are means of three runs. Bold: best in each column within each group.
Advantage
0.5 MP
1 MP
2 MP
4 MP
Max
ViRA-Bench
CrossRPO ( λ=0.5 )
38.7 (65.1) (12K)
46.9 (71.1) (16K)
55.6 (75.8) (26K)
57.5 (78.2) (44K)
63.3 (81.3) (147K)
Joint term
λ=0 (per-version RLOO)
31.3 (59.7) (11K)
41.2 (66.8) (15K)
50.0 (72.5) (24K)
52.9 (74.7) (42K)
58.7 (78.6) (144K)
λ=0 , one version per set
31.1 (59.3) (10K)
40.5 (67.3) (15K)
49.3 (72.5) (23K)
53.6 (74.7) (41K)
58.1 (78.0) (147K)
λ=1
33.4 (61.4) (16K)
44.2 (68.9) (22K)
51.3 (73.7) (35K)
56.1 (76.5) (60K)
56.6 (78.1) (192K)
Appendix
Table 17: Ablations on ViRA-Bench and ViRA-Bench-Real : set accuracy with per-image accuracy in parentheses (%), and tokens per set in gray. All results are means of three runs. The rows are the variants of Table 20 , which Appendix G.1 describes. The rows under Per-version GRPO ×(1+λwv) multiply the advantage of per-version GRPO by this factor (Appendix G.3 ). Appendix G.4 describes the row λ=0 , one version per set .
Perception
Reasoning
Advantage
OCR
RS
DT
MO
AD
All
OCR
DT
MO
AD
All
Overall
Maximum resolution
CrossRPO ( λ=0.5 )
95.2
71.3
96.0
55.8
52.9
68.8
85.0
89.0
60.7
46.0
59.9
65.3
Joint term
λ=0 (per-version RLOO)
95.2
69.3
96.0
53.3
50.3
67.1
84.0
92.0
62.7
44.2
59.6
64.1
λ=0 , one version per set
96.4
68.7
93.0
57.4
53.1
68.9
88.0
89.0
56.7
46.0
59.5
65.2
Appendix
Table 18: Ablations on MME-RealWorld-Lite by task and domain : accuracy (%) at the maximum resolution and at 1 MP. The columns are those of Table 25 , and the rows are those of Table 17 .
Accuracy ↑
Advantage at 1 MP
O
A
B
Zoomed
On obj. ↑
Zooms
CrossRPO ( λ=0.5 )
77.4
68.6
67.2
99.9
86.6
1.07
Joint term
λ=0 (per-version RLOO)
75.0
63.4
61.9
100.0
88.3
1.06
λ=0 , one version per set
78.2
61.9
61.6
100.0
90.0
1.02
λ=1
77.6
66.8
62.4
97.7
82.1
1.61
Appendix
Table 19: Reading the evidence, ablations at 1 MP on ViRA-Bench, with the columns of Table 23 . O , A and B give the accuracy (%) on the original and on the two variants, Zoomed the share of images with at least one zoom, On obj. the share whose last zoom covers the object, and Zooms the zooms per image. The rows are those of Table 17 . All results are means of three runs.
Variant
Baseline
Factor
Scale
CrossRPO ( λ=0.5 )
version
1+0.5wv
batch
λ=0 (per-version RLOO)
version
1
batch
λ=0 , one version per set
version
1
batch
λ=1
version
1+wv
batch
λ=2
version
1+2wv
batch
Random partners
version
1+0.5w~v
batch
Appendix
Table 20: Settings of the ablations. Baseline : the rollouts that the baseline of a rollout averages, those of its own version or those of the whole set. Factor : the multiplier of the difference between the reward and the baseline. Scale : the divisor of the advantages, the root mean square of the batch or the standard deviation of the set or of the version. wv is the cross-input weight (Eq. ( 4 )), and w~v is this weight computed from random partners. One version per set changes the training data: each contrast set keeps one version.
V ∗
HR-4K
HR-8K
MME-RW
TreeBench
O3-Bench
MMStar
Hallusion
Advantage
1
Max
1
Max
1
Max
1
Max
1
Max
1
Max
Max
Max
CrossRPO ( λ=0.5 )
92.0
96.2
87.5
91.0
84.7
90.1
63.6
65.3
56.3
57.0
50.4
67.0
77.1
79.6
Per-version GRPO, λ=0.5
88.7
95.8
83.9
87.8
80.8
88.6
56.1
59.2
54.6
53.1
43.5
67.8
78.3
82.9
Per-version GRPO, λ=1
87.3
94.2
85.7
91.7
83.2
89.8
63.1
64.5
51.4
50.1
35.9
53.9
75.0
77.1
Per-version GRPO
89.5
95.3
86.0
90.8
84.6
88.8
59.5
60.6
53.6
54.8
43.8
58.6
75.1
79.2
Appendix
Table 21: Per-version GRPO with the cross-input weight, general benchmarks : accuracy (%) at 1 MP and at the maximum resolution ( Max ), in the columns of Table 3 . The rows Per-version GRPO, λ multiply the advantage of per-version GRPO by 1+λwv (Appendix G.3 ).
V ∗
HR-4K
HR-8K
MME-RW
TreeBench
O3-Bench
MMStar
Hallusion
Run
1
Max
1
Max
1
Max
1
Max
1
Max
1
Max
Max
Max
CrossRPO ( λ=0.5 )
92.0
96.2
87.5
91.0
84.7
90.1
63.6
65.3
56.3
57.0
50.4
67.0
77.1
79.6
λ=0 (per-version RLOO)
93.0
95.6
87.8
92.9
85.7
89.8
62.4
64.1
54.3
55.3
43.5
64.1
78.8
80.2
λ=0 , one version per set
90.9
95.5
86.0
92.2
82.5
90.2
63.3
65.2
56.3
56.5
43.8
63.2
78.7
78.7
Appendix
Table 22: Training on one version per set, general benchmarks : accuracy (%) at 1 MP and at the maximum resolution ( Max ), in the columns of Table 3 . The last row trains on one version of each contrast set with the advantage of λ=0 (Appendix G.4 ).
Accuracy ↑
Model at 1 MP
O
A
B
Zoomed
On obj. ↑
Zooms
Open 7–9B models + tool
Qwen3.5-9B + tool
63.4
52.5
49.9
99.1
62.6
1.82
ZwZ2-9B + tool
68.9
60.5
55.8
97.1
64.8
1.56
Vision-OPD-9B + tool
42.7
35.3
29.3
4.3
3.4
0.08
Qwen2.5-VL-7B + DeepEyes tool
47.0
24.9
19.0
49.0
18.8
2.45
Appendix
Table 23: Reading the evidence, all models at 1 MP on ViRA-Bench, for every model with a zoom tool. O , A and B give the accuracy (%) on the original and on the two variants, Zoomed the share of images with at least one zoom, On obj. the share whose last zoom covers the object, and Zooms the zooms per image. ZoomEye is a tree search, so its Zoomed and On obj. refer to the crop that its final answer uses, and its Zooms counts the crops that the search visits. ViRA results are means of three runs. Table 19 gives the ablations.
Figure 7: Case study on ViRA-Bench at 1 MP: seven models on 14 sets (Figures 7 to 10 ). A row shows one variant or all three versions of a set. It gives the question, the whole photo with the object in a yellow box, the object at full resolution in each version, and for every model the last crop that it received with its answer. A yellow box in a crop marks the object, and a model that received no crop of the photo is shown with the whole photo. For the open models, colors show where the final reply looks in the shown image (Appendix H.2 ). Claude Haiku 5.5 and GPT-6-Luna run at max thinking.
Figure 8: Case study (continued). The layout is as in Figure 7 . Baselines give the original’s answer on a variant, and both ViRA agents answer every shown version correctly.
Figure 9: Case study (continued). The layout is as in Figure 7 . (a, b) The proprietary models have the object in their last crop but give the original’s answer. (c, d) They zoom up to six times without covering the object. (e) ViRA-9B gives this answer on every version.
Figure 10: Case study (continued). The layout is as in Figure 7 . ViRA-9B zooms once and gives the original’s answer on the variants, whereas ViRA-27B zooms several times and answers correctly.
ViRA-9B
Qwen3.5-9B + tool
1 MP
Max
1 MP
Max
Wrong answers (% of images)
28.9
18.7
44.8
37.1
Share of the wrong answers (%)
No answer
0.7
0.5
16.6
6.4
Answered without zooming
0.4
0.1
1.5
3.4
No crop covers the object
21.0
10.0
22.8
19.1
Appendix
Table 24: Error types on ViRA-Bench (2,286 images of 762 sets). A crop covers the object if its box on the photo contains at least half of the object’s box. No answer includes rollouts that reach the limit of six calls. ViRA-9B numbers are means of three runs, and Qwen3.5-9B + tool is run once.
Perception
Reasoning
Model at Max
OCR
RS
DT
MO
AD
All
OCR
DT
MO
AD
All
Overall
Open 7–9B models
Qwen3.5-9B
91.2
58.7
94.0
40.8
44.0
59.4
76.0
79.0
42.7
38.2
49.6
55.5
ZwZ2-9B
92.0
60.7
90.0
48.0
55.4
64.8
86.0
90.0
46.0
43.2
55.7
61.3
Vision-OPD-9B
94.0
60.7
96.0
51.7
53.7
66.3
82.0
75.0
60.7
39.2
54.0
61.5
Open 7–9B models + tool
Appendix
Table 25: MME-RealWorld-Lite by task and domain at the maximum resolution : accuracy (%) of every model in Table 2 . The domains are OCR in complex context (OCR), remote sensing (RS), diagrams and tables (DT), monitoring (MO) and autonomous driving (AD). Perception has 1,169 questions (OCR 250, RS 150, DT 100, MO 319, AD 350), and reasoning has 750 (OCR 100, DT 100, MO 150, AD 400). All is the accuracy on all questions of the task, and Overall is the accuracy on all 1,919 questions, as in Table 2 . Tools, + tool, Δ and the colors are as in Table 1 . Bold: best in each column within each group. Table 18 gives the ablations.
Perception
Reasoning
Model at 1 MP
OCR
RS
DT
MO
AD
All
OCR
DT
MO
AD
All
Overall
Open 7–9B models
Qwen3.5-9B
85.2
50.7
83.0
39.8
43.4
55.7
76.0
72.0
42.7
37.5
48.3
52.8
ZwZ2-9B
86.4
55.3
82.0
44.5
50.3
59.8
82.0
76.0
46.0
42.2
52.8
57.1
Vision-OPD-9B
88.0
56.7
84.0
48.0
53.4
62.4
79.0
66.0
50.7
38.8
50.1
57.6
Open 7–9B models + tool
Appendix
Table 26: MME-RealWorld-Lite by task and domain at 1 MP : accuracy (%). The columns, rows, colors and bold are as in Table 25 .
What does it take to build a visual reasoner that works across charts, science, spatial understanding, and open-ended tasks? The strongest vision-language models (VLMs) suggest that broad visual reasoning is within reach, yet their closed data and reinforcement learning (RL) pipelines make their gains difficult to study, reproduce, or extend. We introduce Vero, a family of fully open VLMs that match or exceed existing open-weight models across diverse visual reasoning tasks. We scale RL data and rewards across six broad task categories, constructing Vero-600K, a 600K-sample dataset from 59 datasets, and designing task-routed rewards that handle heterogeneous answers. Across VeroEval, our 30-benchmark suite, Vero-600K outperforms existing RL datasets under controlled comparisons. Applied to five starting models, Vero variants gain 2.9-5.4 points on average over their initial models. Notably, Vero-Qwen3I-8B, trained on the Instruct model, surpasses Qwen3-VL-8B-Thinking by 3.8 points on average without additional distillation. Systematic ablations reveal that different task categories elicit distinct reasoning patterns and that broad gains depend on learning them jointly rather than in isolation. All data, code, and models are publicly available.
Retrieval-Augmented Generation (RAG) extends Large Vision-Language Models (LVLMs) with external visual knowledge. However, existing visual RAG systems typically rely on generic retrieval signals that overlook the fine-grained visual semantics essential for complex reasoning. To address this limitation, we propose UniDoc-RL, a unified reinforcement learning framework in which an LVLM agent jointly performs retrieval, reranking, active visual perception, and reasoning. UniDoc-RL formulates visual information acquisition as a sequential decision-making problem with a hierarchical action space. Specifically, it progressively refines visual evidence from coarse-grained document retrieval to fine-grained image selection and active region cropping, allowing the model to suppress irrelevant content and attend to information-dense regions. For effective end-to-end training, we introduce a dense multi-reward scheme that provides task-aware supervision for each action. Based on Group Relative Policy Optimization (GRPO), UniDoc-RL aligns agent behavior with multiple objectives without relying on a separate value network. To support this training paradigm, we curate a comprehensive dataset of high-quality reasoning trajectories with fine-grained action annotations. Experiments on three benchmarks demonstrate that UniDoc-RL consistently surpasses state-of-the-art baselines, yielding up to 17.7% gains over prior RL-based methods.
High-resolution visual question answering often fails because a multimodal model does not acquire the small, spatially localized evidence needed to answer a question. Sequential zooming can recover detail, but it asks the main model to choose a region before obtaining a reliable overview. We introduce VPS, a visual parallel-search framework in which a main agent first invokes grid_search to inspect image tiles in parallel with question-conditioned sub-agents, and then adaptively invokes zoom_in on a precise or merged region. The same interface supports both training-free inference and post-training of the main and sub-agents. Across five benchmark splits and three model sizes, VPS improves mean accuracy over dedicated zoom-only search in 14 of 15 same-model comparisons, with gains up to 8.0 points and especially strong improvements for smaller main models. ZoomBench retains an approximately 3.2-point gain at every tested size. We further develop a supervision pipeline with hint-free verification and a paired role-specific GRPO surrogate for learning the controller and tile-reader roles. SFT improves observed accuracy on all five benchmark splits, including a 4.17-point gain on HR-Bench 4K. Role-specific RL further reshapes search behavior: main-only RL reduces mean tool use from 2.65 to 2.11 with similar pass@1 in an internal four-response evaluation, while external accuracy changes are mixed. Joint training reveals an asymmetry between local evidence reading and global search control. Together, these results support VPS as an effective inference-time scaffold and a trainable decomposition for visual evidence acquisition.
Xijia Tao, Yihua Teng, Xinyu Fu +5
The University of Hong Kong · Huawei Research · City University of Hong Kong +1