Multi-step visual reasoning suffers when early visual misinterpretations persist as incorrect premises in subsequent reasoning. We argue that how visual evidence is represented during reasoning matters as much as how much of it is gathered. Rather than appending observations and inferences to an ever-growing dialogue, we introduce a Visual Board-centered harness that separates visual observations, uncertain interpretations, and derived reasoning within a structured, capacity-bounded, and revisable workspace. This separation makes each premise explicit and auditable, so that conflicts between what is seen and what is derived become detectable even by a text-only critic and can trigger targeted reinspection of the original image. Without task-specific training, our harness consistently outperforms standalone models across 4B, 8B, Qwen3-VL-Flash, and 32B configurations on Euclid30K, SolidGeo, and ChartQA bar charts, with gains of up to 11.68, 18.42, and 24.82 percentage points, respectively. Under identical tools, critic, and interaction limits, the structured Board outperforms shared free-form notes by 6.67 to 13.00 points and no-Board retries by 7.35 to 18.80 points while using 7.9% to 38.2% fewer tokens. Supervised fine-tuning on harness trajectories further improves direct answering and critic-free tool use, with positive transfer to chart reasoning.
Figures & tables
Figure 1: The Visual Board as a revisable reference. Upper path: an append-only interaction history carries a mistaken observation forward into every later step. Lower path: the Board keeps image facts, uncertain facts, and reasoning steps as explicit, separate premises, so the conflict becomes visible to a text-only Critic and guides targeted reinspection of the original image.
Figure 2: Overview of results. (a) Without task-specific training, the Visual Board harness improves all four backbone configurations on all three benchmarks over their standalone models. (b) On 500 Euclid30K questions with identical tools, Critic, and interaction limits, the structured Board ( ∙ ) is more accurate than shared free-form notes ( ▲ ) and no-Board retries ( × ) for every backbone, while using fewer tokens than no-Board retries (dashed arrows). (c) Supervised fine-tuning of a Qwen3-VL-4B Solver on geometry harness trajectories helps critic-free tool use far more than direct answering.
Figure 3: Overview of the Visual Board-centered harness. The VLM Solver keeps the original image available, maintains a revisable Board, and invokes focused crop or whole-image queries whenever a visual premise requires checking. A text-only Critic reviews the submitted premises and derivation to return structured feedback, without directly inspecting the image. History compression preserves both the current Board and direct image access.
Euclid30K
SolidGeo
ChartQA-Bar
System / Backbone
Base
Board
Δ
Base
Board
Δ
Base
Board
Δ
All-4B / Qwen3-VL-4B
39.40
48.80
+9.40
32.11
40.84
+8.73
57.75
77.97
+20.22
All-8B / Qwen3-VL-8B
51.40
58.75
+7.35
35.99
47.20
+11.21
57.99
81.60
+23.61
Qwen3-VL-Flash
50.80
56.90
+6.10
39.83
58.25
+18.42
68.52
86.56
+18.04
All-32B / Qwen3-VL-32B
57.40
69.08
+11.68
41.02
59.34
+18.32
61.86
86.68
+24.82
Table 1: Training-free performance compared with standalone backbones. We report accuracy in percent and improvements Δ in percentage points on Euclid30K, SolidGeo, and ChartQA-Bar [ 29 ] , comprising 500, 2,206, and 826 questions, respectively. All-4B, All-8B, and All-32B use size-matched models for the Solver, visual tools, and Critic, whereas Qwen3-VL-Flash employs Qwen3-Max as its Critic. Boldface indicates the higher accuracy in each comparison.
Setting
Euclid30K
SolidGeo
ChartQA-Bar
Standalone
39.40
32.11
57.75
Tools only
35.60
24.25
74.80
Full harness
48.80
40.84
77.97
Table 2: Training-free 4B tool-use control. Tools only uses the base 4B Solver and 4B visual backend without a Board, Critic, or post-submission visual check, whereas the full All-4B harness retains both the Board and Critic. All scores are accuracies in percent, obtained from base models without any SFT.
Method
Acc. (%)
Δ (pp)
ProReason
51.60
–
Visual Board
58.75
+7.15
Table 3: ProReason comparison on Euclid30K. Both methods use Qwen3-VL-8B-Instruct and Qwen3-8B on all 500 questions.
Figure 4: Illustrative Euclid30K trajectory under All-8B. In Round 1, reasoning_steps derives a central octahedron while image_facts records option A as a tetrahedron, giving “None of the above.” Because the two occupy separate Board fields, the text-only Critic detects the conflict and requests a vlm_query recheck. Round 2 replaces the entry and answers A, and Round 3 refines it. The final needs_revision status comes from the post-submission check, and the run ends at the round limit with answer A.
Board setting
Acc. (%)
Tokens
Time (s)
All-4B
No Board + Critic
30.00
60,636.13
64.73
Shared free-form notes
35.80
54,601.00
64.61
Concise Board
40.40
77,257.18
83.46
Default Visual Board
48.80
55,825.19
56.87
All-8B
Table 4: Board ablations on 500-question Euclid30K evaluations. Models and interaction limits are fixed within each group. Accuracy counts failed runs as incorrect, and tokens and time are per-question means. Bold marks the best accuracy in each group.
Free-form correct
Free-form wrong
Default correct
139
105
Default wrong
40
214
Table 5: Paired All-4B outcomes on 498 common Euclid30K questions. The exact gain is 100(105−40)/498=13.05 pp, with a 95% paired bootstrap interval of [8.43,17.67] pp.
Figure 5: Geometry-trained SFT improves Direct QA and critic-free tool use. Within each row, Base and SFT Qwen3-VL-4B Solvers share identical tool access, visual-tool backend, and Board format with the Critic disabled to isolate Solver fine-tuning. Board denotes the default Visual Board, and ChartQA-Bar denotes the 826-item bar-chart split. All SFT evaluations use the same 200-step checkpoint; complete numerical values are listed in the supplementary material.
Configuration
Critic
Acc. (%)
Tokens
Time (s)
All-4B
With
48.80
55,825.19
56.87
Without
36.80
44,274.37
43.79
All-8B
With
58.75
50,421.00
103.16
Without
50.80
28,374.05
77.36
Qwen3.8-Flash
With
68.47
42,775.75
61.66
Without
69.80
17,187.12
26.15
Table 6: Critic ablation with a Visual Board retained. Each pair fixes the Solver and visual backend on the 500 Euclid30K questions. Bold marks higher accuracy within each pair.