Vision-language models can answer questions about images and video, but their multi-step reasoning is brittle: sampling the same input multiple times often produces different reasoning paths and conflicting final answers. A new arXiv preprint, "CASHEW: Stabilizing Multimodal Reasoning via Iterative Trajectory Aggregation," tackles this by treating reasoning as a test-time scaling problem.

CASHEW is an inference-time framework that iteratively aggregates multiple candidate reasoning trajectories into a higher-quality trace. During aggregation, it uses visual verification of object presence as a grounding signal. The authors also introduce CASHEW-RL, a learned variant trained with Group Sequence Policy Optimization. Its composite reward encourages correct answers that are grounded in minimal but sufficient visual evidence, and it adaptively allocates reasoning effort based on task difficulty.

On 13 image and video understanding and reasoning benchmarks, CASHEW and CASHEW-RL improved average accuracy by +4.1 and +5.3 percentage points respectively on Qwen3-VL-4B, and by +2.6 and +3.0 points on Qwen3-VL-8B. The results suggest that explicit aggregation and learned self-aggregation are both viable routes to more stable multimodal reasoning.