A Progressive Training Strategy for Vision-Language Models to Counteract Spatio-Temporal Hallucinations in Embodied Reasoning

TL;DR

This paper introduces a progressive training method for vision-language models that significantly reduces spatio-temporal reasoning hallucinations and improves dynamic reasoning capabilities.

Contribution

It develops a new Chain-of-Thought dataset and a progressive training framework combining supervised and weakly-labeled data to enhance reasoning accuracy.

Findings

Reduces forward-backward performance gap from over 70% to 6.53%.

Improves backbone accuracy in spatiotemporal reasoning tasks.

Develops models with more authentic dynamic reasoning abilities.

Abstract

Vision-Language Models (VLMs) have made significant strides in static image understanding but continue to face critical hurdles in spatiotemporal reasoning. A major bottleneck is "multi-image reasoning hallucination", where a massive performance drop between forward and reverse temporal queries reveals a dependence on superficial shortcuts instead of genuine causal understanding. To mitigate this, we first develop a new Chain-of-Thought (CoT) dataset that decomposes intricate reasoning into detailed spatiotemporal steps and definitive judgments. Building on this, we present a progressive training framework: it initiates with supervised pre-training on our CoT dataset to instill logical structures, followed by fine-tuning with scalable weakly-labeled data for broader generalization. Our experiments demonstrate that this approach not only improves backbone accuracy but also slashes the forward-backward performance gap from over 70\% to only 6.53\%. This confirms the method's ability to develop authentic dynamic reasoning and reduce the inherent temporal biases of current VLMs.