Thinking with Images in Large Multimodal Reasoning Models

3D spatial reasoning remains challenging: it requires both accurate answers and structured reasoning in the form of executable code that invokes tools. Data scarcity is a bottleneck: benchmarks like Omni3D-Bench offer only hundreds of questions. This thesis addresses that gap with a tool-augmented approach: the model receives dense scene captions, generates executable code that calls tools, and is trained with GRPO using verifiable accuracy rewards. We contribute (1) a larger dataset from structured Omni3D annotations, (2) an oracle setting that separates tool error from reasoning error, and (3) a GRPO methodology with composite rewards (format, execution, accuracy, code quality, consistency) and ablations. On Omni3D-Bench (501 questions), GRPO improves accuracy from 19.5% (SFT) to 29.1% (full reward), and format and execution compliance rise from ~30% to near-perfect levels. The oracle reaches 35.6%, exposing a gap of 6.5 percentage points under standard tools and indicating imperfect 3D grounding as a bottleneck. Reward ablations confirm that consistency and code quality add value. The work demonstrates substantial improvements in both accuracy and output quality, while identifying main bottlenecks and a path for future improvements.