Step-level credit assignment added to Interleave-RL raises the reasoning score from 46.3% to 66.3% on GeoVAD-Bench¶

Claim · All claims
Evidence · 1 study · causal r2 · q2
Currency · newest study 2026 · not yet checked against its sources · not reviewed by a person

Subclaims¶

q2 i? Adding step-level credit assignment on top of Interleave-RL yields the strongest ablation performance, with reasoning increasing from 46.3% to 66.3% and answer accuracy from 60.3% to 62.6%. → 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

Ablation study on GeoVAD-Bench under the Auto-Aux setting comparing base model, SFT-only, Interleave-RL without SCA, and full model (Table 9). The article reports the Reasoning score "increases from 46.3% to 66.3%" when SCA is added.

"Adding SCA on top of Interleave-RL yields the strongest overall performance. In particular, the Reasoning score increases from 46.3% to 66.3%, indicating that step-level process supervision effectively directs optimization toward correcting individual reasoning steps."

Discussion¶

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
• Reasoning with ground-truth auxiliary diagrams consistently outperforms reasoning from the original diagram alone on GeoVAD-Bench — related
• Four process-level error categories account for roughly nine in ten attributed failures in VCoT geometry solving — related
• Correct final answers are associated with stronger performance on all four process-level metrics, with reasoning-process correctness showing the clearest separation — related