Chain-of-thought (CoT) reasoning lets multimodal large reasoning models (MLRMs) break complex visual problems into intermediate steps. But this capability also exposes them to a threat called CoT hijacking, where an adversary implants a backdoor through poisoned fine-tuning. The attack can then manipulate the model's reasoning process, potentially leading to incorrect or malicious outputs.

The arXiv preprint "Detecting Chain-of-Thought Hijacking in Backdoored Multimodal Large Reasoning Models" (arXiv:2610.16598) describes this vulnerability. The abstract explains that CoT reasoning enables MLRMs to solve complex visual problems, but also makes them susceptible to hijacking. The paper's focus is on detecting such backdoors, though the full methodology is not detailed in the available excerpt.

As multimodal reasoning models become more capable and widely used, securing their internal reasoning steps is increasingly important. This work underscores the need for detection techniques that can identify when a model's chain-of-thought has been compromised, helping to preserve reliability and trust in these systems.