The explain-on-blackboard paradigm for interleaved visual-text reasoning

Abstract

Interleaved visual-text chain-of-thought reasoning, which generates intermediate visual artifacts to support deduction, is a promising direction for complex multimodal tasks. However, in our benchmark evaluation we observe a consistent representation tax: models perform systematically worse when translating abstract reasoning into structured spatial visualizations than when producing textual explanations alone. We conjecture that this gap is correlated with (rather than solely caused by) an alignment bias in conventional supervised fine-tuning (SFT), which optimizes outcome labels while providing little supervision on synchronization between visual semantics and textual reasoning. To empirically probe this hypothesis, we propose a pedagogical distillation pipeline building on an Explain-on-Blackboard (EoB) formulation. Inspired by dual-coding and multimedia-learning theories, teacher models generate temporally and semantically aligned teaching trajectories that comprise a discourse channel for logical explanation and a canvas channel for incremental HTML/SVG construction. A dynamic-rubric quality assurance mechanism with multi-stage filtering retains only high-fidelity trajectories for distillation. To measure progress on this problem, we further construct Cas-Bench, a 1,000-problem multidisciplinary evaluation protocol with process-level rubrics. In our experiments an 8B-parameter model fine-tuned on 4,000 distilled trajectories obtains a 69.76% relative improvement on Cas-Bench and 2–4 absolute-point gains on MMMU, MMMU-Pro, MathVision, and MathVista over its base model.