A Progressive Training Strategy for Vision-Language Models to Counteract Spatio-Temporal Hallucinations in Embodied Reasoning
For embodied AI systems requiring robust temporal understanding, this work reduces temporal biases in VLMs, enabling more reliable reasoning across time.
The paper tackles multi-image reasoning hallucination in VLMs, where a large performance gap between forward and reverse temporal queries indicates reliance on shortcuts. Their progressive training strategy reduces this gap from over 70% to 6.53%, demonstrating improved genuine spatiotemporal reasoning.
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.