CASHEW: Stabilizing Multimodal Reasoning via Iterative Trajectory Aggregation

TL;DR

This paper introduces CASHEW, a framework that stabilizes multimodal reasoning by aggregating multiple reasoning trajectories and grounding them in visual evidence, significantly improving performance across various benchmarks.

Contribution

It proposes a novel inference-time and learned aggregation approach to stabilize multimodal reasoning, with training via Group Sequence Policy Optimization for robustness.

Findings

Up to +23.6 percentage points on ScienceQA

Up to +8.1 percentage points on EgoSchema

Significant performance improvements across 13 benchmarks

Abstract

Vision-language models achieve strong performance across a wide range of multimodal understanding and reasoning tasks, yet their multi-step reasoning remains unstable. Repeated sampling over the same input often produces divergent reasoning trajectories and inconsistent final predictions. To address this, we introduce two complementary approaches inspired by test-time scaling: (1) CASHEW, an inference-time framework that stabilizes reasoning by iteratively aggregating multiple candidate trajectories into higher-quality reasoning traces, with explicit visual verification filtering hallucinated steps and grounding reasoning in visual evidence, and (2) CASHEW-RL, a learned variant that internalizes this aggregation behavior within a single model. CASHEW-RL is trained using Group Sequence Policy Optimization (GSPO) with a composite reward that encourages correct answers grounded in minimal yet sufficient visual evidence, while adaptively allocating reasoning effort based on task difficulty. This training objective enables robust self-aggregation at inference. Extensive experiments on 13 image understanding, video understanding, and video reasoning benchmarks show significant performance improvements, including gains of up to +23.6 percentage points on ScienceQA and +8.1 percentage points on EgoSchema.