Large Language Models (LLMs) can perform sentiment analysis via natural language instructions, yet their predictions are highly sensitive to prompt phrasing.Prior work has shown that sentiment is encoded linearly in LLM representations, but the model's ability to utilize this information remains surprisingly fragile to prompt variations.To understand this behavior, we leverage Sparse Autoencoders (SAEs) to extract interpretable features from LLM activations and apply circuit-level analysis to uncover causal mechanisms underlying sentiment prediction.We identify a sentiment analysis circuit and find that prompt sensitivity may stem from task activation failure.Based on this insight, we propose a simple inference-time intervention method that amplifies circuit features to compensate for insufficient activation.Experiments across diverse datasets, templates, and languages show consistent improvements, offering an interpretable and training-free alternative to manual prompt engineering.

Offloading computationally intensive tasks from user equipments (UEs) to mobile edge computing (MEC) servers is a promising technique to boost up the computational capacity of UEs. However, MEC will incur extra energy consumption and time delays, which motivates the deployment of energy harvesting (EH) small cell networks with MEC in mobile networks. Due to the complexity of such networks, it is challenging to effectively allocate resources for UEs. In this paper, we investigate the offloading decision, wireless and computational resources allocation problem in energy harvesting (EH) small cell networks with MEC. Different from existing literatures, our research focuses on improving mobile operators' revenue by maximizing the amount of the offloaded tasks while decreasing the energy expenditure and time-delays. Besides, queues are created at the MEC server side to store the un-executed tasks in a time slot, which is used as a punishment in our utility function to avoid serious delay. Considering the varying lengths of queues, the states of EH-batteries of small base stations (SBSs) and down-link channels, the above problem is modeled as a Markov decision process (MDP). Since the states and actions in the MDP are infinite, an online and on-policy actor-critic with eligibility traces algorithm is proposed to resolve the problem. Simulation results show the proposed algorithm has superior performances compared with the policy-gradient algorithm and Q-learning.