Reasoning with ground-truth auxiliary diagrams consistently outperforms reasoning from the original diagram alone on GeoVAD-Bench¶

Claim · All claims
Evidence · 1 study · causal r2 · q2
Currency · newest study 2026 · checked against its sources 2026-10-20 (automated, 1 of 1 entry) · not reviewed by a person · recheck by 2029: tool and model accuracy changes quickly

Subclaims¶

q2 i? Across evaluated models, final-answer accuracy under GT-Aux consistently exceeds No-Aux, with reference-diagram gains of 3.0 to 7.0 percentage points as printed in Table 4. → Zhitong Dong 2026

Evidence¶

Zhitong Dong 2026¶

Zhitong Dong, Jicai Pan, Yingguo Gao, Jingting Ding, Hao Chen, Jinjie Gu. (2026). Beyond Generation and Accuracy: Diagnosing and Enhancing Visual Chain-of-Thought for Geometry Problem Solving. arXiv preprint. https://arxiv.org/abs/2609.12606 doi:10.48550/arxiv.2609.12606

q2 · i? · causal · r2

Benchmark evaluation on GeoVAD-Bench comparing final-answer accuracy under three controlled auxiliary conditions (Table 4). The article reports "GT-Aux consistently outperforms No-Aux"; printed reference gains are +3.3 (GPT-5.4 + GPT-image2), +3.0 (Qwen3-VL-8B + Qwen-image-edit), and +7.0 (SenseNova-U1-8B) percentage points.

"Across the evaluated models, GT-Aux consistently outperforms No-Aux, showing that validated auxiliary diagrams provide useful information for geometry problem solving."

Discussion¶

Learner Variables¶

• Geometric Reasoning — outcome: instruction changes it

Related Claims¶

• GeoWeave-8B achieves the highest open-source final-answer accuracy and process average on GeoVAD-Bench, gaining 25.3 and 30.4 percentage points over its base model — related
• Step-level credit assignment added to Interleave-RL raises the reasoning score from 46.3% to 66.3% on GeoVAD-Bench — related
• Step-linked visual grounding externalizes spatial reasoning: synchronized diagrams that reveal auxiliary lines step by step reduce split-attention effort — related
• Correct final answers are associated with stronger performance on all four process-level metrics, with reasoning-process correctness showing the clearest separation — related
• Viewpoint-aware spatial reasoning is a persistent bottleneck with poor calibration of spatial uncertainty — related