Flipping Bits to Change Stance: Decision-Level Cognitive Hijacking of Large Language Models

The paper proposes the CogBias framework, which uses bit-flip attacks (BFA) to tamper with large language model weights after deployment, achieving decision-level cognitive hijacking. CogBias uses a differentiable sentiment evaluator to turn subjective preferences into optimization signals, uses a multi-objective loss to jointly constrain multiple dimensions, and builds BitScout to locate key bits, completing targeted cognitive intervention under an extremely sparse flip budget. On Llama-3.2-3B, Mistral-7B and Qwen2.5-14B, as well as in commercial recommendation and controversial factual topic scenarios, flipping a small number of bits can stably induce significant stance shifts on target topics, with limited impact on non-target tasks and the overall output distribution. The authors state that the attack requires no real-time interaction and no control over the training process, indicating that small perturbations to weight data are sufficient to break the model's high-level value alignment.