Recent multimodal large language models (MLLMs) support Thinking with Images, invoking visual tools such as zooming and cropping to inspect image regions during inference. Yet these systems remain brittle in fine-grained reasoning: to acquire a decisive detail, a model must ground its attention on the correct region, but knowing which region is correct presupposes having already observed that detail. We identify this circular dependency as the grounding paradox, show that grounding errors are rarely self-corrected within a single trajectory---once a misleading region is inspected, all subsequent reasoning conditions on that observation and the error propagates to the final answer---and observe that because each trajectory constructs its own evidence, answer-level aggregation discards the very information that distinguishes trajectories. We propose Test-Time Scaling over Perception (TTSP), a closed-loop framework that treats perception as the unit of scalable inference and allocates compute along two axes: Entropy-Gated Perceptual Exploration samples diverse trajectories and uses critical-token entropy to withhold evidence the model cannot commit to, while Evidence-Guided Iterative Refinement distills validated observations into a correctable Evidence Ledger that steers later rounds to re-inspect unresolved regions. Across high-resolution and general multimodal benchmarks, TTSP consistently outperforms strong test-time scaling baselines, while improving grounding quality with favorable token efficiency.
Figures & tables
Figure 1: The grounding paradox and its resolution. Left: a single trajectory must commit to a region under incomplete evidence and inherits any early error. Right: TTSP converts the circularity into a gated search across trajectories, coupled through a correctable evidence ledger.
Figure 2: Overview of TTSP . Each round draws fresh and guided traces ( coverage ), gates them on critical-token entropy ( selection ), and refines an Evidence Ledger whose spatial directives redirect the next round ( utilization ); the answer is a reliability-weighted vote over all retained traces.
V ∗ Bench
HR-Bench-4K
HR-Bench-8K
TreeBench
MME-RealWorld-Lite
Method
Attr.
Spat.
Ovr.
FSP
FCP
Ovr.
FSP
FCP
Ovr.
Perc.
Reas.
Ovr.
Perc.
Reas.
Ovr.
Reference systems, not budget-matched
GPT-4o
72.2
60.5
67.5
66.8
63.3
65.0
60.8
58.5
59.6
61.7
38.3
46.9
54.4
48.3
52.0
Thyme
83.5
80.3
82.2
91.0
63.0
77.0
86.5
57.5
72.0
57.0
28.9
39.3
59.1
49.1
55.2
DeepEyes
91.3
88.2
90.1
91.3
59.0
75.1
86.8
58.5
72.6
57.7
29.3
39.8
55.4
46.8
52.1
Pixel-Reasoner
83.5
76.3
80.6
86.0
60.3
72.9
80.0
54.3
66.9
54.3
30.1
39.0
53.1
45.6
50.1
Table 1: High-resolution and fine-grained visual reasoning. All test-time scaling methods within a backbone block use the same trajectory budget as TTSP ; reference systems are not budget-matched. Best result per column within each block is in bold.
Table 4
Figure 3: Scaling behavior of TTSP along (a) backbone size, (b) perception width K , and (c) perception depth N .
Figure 6Figure 7
Appendix figures & tables11 assets
Supplementary material from the paper’s appendix.
Appendix
Perception
Reasoning
Method
Attr.
Mat.
Phys.
Ret.
OCR
Ovr.
Persp.
Ord.
Cont.
Cntm.
Comp.
Ovr.
GPT-4o
51.7
61.5
65.2
43.8
69.1
61.7
18.8
38.6
48.8
72.4
43.2
38.3
DeepEyes
62.1
53.8
65.2
68.8
51.5
57.7
11.8
24.6
36.6
51.7
47.7
29.3
Pixel-Reasoner
58.6
61.5
65.2
50.0
48.5
54.3
14.1
31.6
39.0
44.8
40.9
30.1
Qwen3-VL-4B-Instruct
Base
44.8
61.5
52.2
62.5
69.1
60.4
15.3
29.8
48.8
75.9
47.7
36.3
Appendix
Table 4: Category-level results on TreeBench. Perception covers attributes (Attr.), material (Mat.), physical state (Phys.), object retrieval (Ret.), and OCR. Reasoning covers perspective transformation (Persp.), ordering (Ord.), contact and occlusion (Cont.), spatial containment (Cntm.), and comparison (Comp.).
Perception
Reasoning
Method
OCR
RS
DT
MO
AD
Ovr.
OCR
DT
MO
AD
Ovr.
Pixel-Reasoner
89.6
52.0
86.0
38.9
30.9
53.1
71.0
72.0
46.0
32.5
45.6
DeepEyes
90.0
52.7
89.0
43.3
33.4
55.4
76.0
69.0
44.0
35.0
46.8
Qwen3-VL-4B-Instruct
Base
89.6
50.7
85.0
37.3
34.3
53.4
73.0
75.0
42.7
31.0
44.8
SC
91.2
53.3
89.0
38.6
36.9
55.5
78.0
78.0
43.3
32.8
47.0
Appendix
Table 5: Category-level results on MME-RealWorld-Lite. RS, DT, MO, and AD denote remote sensing, diagram and table understanding, monitoring, and autonomous driving, respectively.
Split
RTWI
R1
R2
R3
R4
Perception
40.65
37.83
40.38
42.03
45.29
Reasoning
45.82
42.79
44.51
47.04
49.22
Appendix
Table 6: Grounding mIoU on TreeBench after each round of TTSP , compared with the strongest matched-budget baseline. R1 to R4 denote the refinement rounds; each round is measured over the same number of retained traces.
V ∗ Bench
HR-Bench-4K
HR-Bench-8K
Signal
Attr.
Spat.
FSP
FCP
FSP
FCP
Avg.
Answer probability
93.9
92.1
96.0
75.3
93.0
74.5
87.5
Probability margin
93.0
93.4
96.5
75.8
93.5
74.5
87.8
Mean token probability
94.8
92.1
96.0
76.3
93.0
74.0
87.7
Length-normalized probability
93.9
92.1
96.5
76.3
93.5
74.5
87.8
Cross-trace agreement
93.9
92.1
96.0
76.5
93.5
74.8
87.8
Appendix
Table 7: Accuracy when alternative trace-reliability signals replace critical-token entropy inside TTSP , with all other settings fixed. Results use Qwen3-VL-4B-Instruct. Avg. is the mean of the six reported metrics.
V ∗ Bench
HR-Bench-4K
HR-Bench-8K
TreeBench
Method
Attr.
Spat.
Ovr.
FSP
FCP
Ovr.
FSP
FCP
Ovr.
Perc.
Reas.
Ovr.
DeepEyes-7B
Base
87.8
81.0
85.1
90.0
58.5
74.2
86.5
57.3
71.9
57.0
35.9
43.7
SC
89.6
82.3
86.7
91.3
61.0
76.2
88.0
61.0
74.5
61.1
37.1
45.9
CISC
90.4
82.3
87.2
91.3
59.0
75.2
88.0
61.8
74.9
61.1
37.5
46.2
TRT
89.6
81.0
86.2
90.0
61.8
75.9
87.8
59.0
73.4
60.4
37.1
45.7
Appendix
Table 8: Generalization of TTSP to additional multimodal backbones on high-resolution benchmarks. Absolute scores are not comparable with Table 1 , since the backbones and inference setups differ.
Method
Easy
Medium
Hard
Overall
Base
56.0
35.4
34.9
40.9
SC
64.0
40.2
42.5
47.2
CISC
59.7
38.6
39.6
44.6
TRT
59.7
39.6
40.6
45.3
Self-Ref.
61.0
39.6
38.7
45.3
Self-Cer.
59.0
39.4
37.7
44.4
Appendix
Table 9: Results on VisualProbe with Qwen3-VL-8B-Instruct.
Setting
Value
Perception depth N
4
Perception width K
8
Fresh ratio α
0.4
Gate ratio ρ
0.4
Entropy truncation ktop
256
Critical positions k′
128
Appendix
Table 10: Default settings of TTSP , shared across all backbones and benchmarks.
Figure 8: Prompt template for perceptual exploration.
Figure 9: Prompt template for ledger-guided exploration.
Figure 10: Prompt template for the Evidence Ledger update.
Figure 11: Specification of the crop tool exposed to the model.
Visually grounded reasoning links visual evidence to multi-step conclusions and actions, aiming to make multimodal systems accurate, auditable, and robust.
Qwen Business Unit of Alibaba · School of Artificial Intelligence, Nanjing University · National Key Laboratory for Novel Software Technology, Nanjing University +5