InterSketch: An Interleaved Reasoning Model with Self-correcting Visual Sketch and Stepwise Reward

InterSketch introduces a new vision-language model architecture that combines visual sketches with textual reasoning in an interleaved chain-of-thought approach, moving beyond text-centric AI paradigms. The model uses self-correction mechanisms and stepwise reward functions during reinforcement learning to improve performance on complex visual reasoning tasks, reportedly outperforming proprietary models like Gemini-3-Pro.

InterSketch addresses a fundamental limitation in current vision-language models: their tendency to rely heavily on text-based reasoning while underutilizing visual information for complex tasks. Traditional VLMs process visual input but default to sequential text-only reasoning chains, creating a bottleneck for multi-step visual understanding problems. The InterSketch approach fundamentally reimagines this by creating an interleaved visual-textual chain-of-thought where intermediate visual sketches dynamically supplement textual analysis, mirroring how humans integrate perception with reasoning.

The research builds on growing recognition that chain-of-thought prompting improves LLM reasoning, now extending this principle to multimodal systems. By incorporating self-correction mechanisms and stepwise reward structures rather than relying on sparse end-task rewards, InterSketch tackles the notoriously difficult problem of training AI systems on long-horizon reasoning tasks where traditional supervision signals prove insufficient.

From a development perspective, the ability to generate intermediate visual representations during reasoning opens new possibilities for AI systems tackling design, analysis, and problem-solving tasks requiring visual-spatial understanding. The benchmarking results claiming superiority over Gemini-3-Pro suggest meaningful architectural innovations rather than incremental improvements, though peer validation through full publication remains pending.

Future developments will determine whether this interleaved approach generalizes across diverse domains or remains specialized to particular visual reasoning tasks. Integration with multimodal APIs and real-world applications in medical imaging, engineering, or scientific visualization represents the next frontier.

• →InterSketch interleaves visual sketches with text reasoning, moving beyond text-centric AI paradigms for complex visual tasks.
• →Self-correction mechanisms and stepwise reward functions enable effective training on long-horizon reasoning problems.
• →The model reportedly outperforms proprietary systems including Gemini-3-Pro on visual reasoning benchmarks.
• →Two-stage training approach combines synthesized datasets with reinforcement learning for improved multimodal reasoning.
• →Intermediate visual generation represents a significant architectural shift toward human-like visual-cognitive processing in AI.