The rapid development of deep learning has been successfully applied in various fields, including aquaculture, providing new methods and ideas to realize unmanned and intelligent aquaculture. This paper focuses on the technology and research methods used in deep learning for intelligent feeding in fish farming in the past decade, and discusses the application of deep learning in intelligent feeding of fish in detail, including feeding behavior analysis, detection and tracking of live fish, growth state monitoring, residual bait identification and counting, water quality prediction, etc., and summarizes and evaluates the methods, at the same time, analyzes the technical details of the deep learning applied to intelligent precision feeding is analyzed in technical details, including data, algorithms, and evaluation performance indexes. The summarized results show that the advantage of deep learning lies in the automatic extraction of features, which also provides technical support for the construction of intelligent feeding system. However, due to the large differences in fish species, aquaculture environments and data acquisition methods, the technology is less portable, and it is still in the stage of weak artificial intelligence, which requires a large amount of data to train the model, and the cost is high, which has become a bottleneck that restricts the further development of aquaculture. Nevertheless, deep learning has still made breakthroughs in processing complex data. In summary, the purpose of this review is to provide researchers and producers with a better understanding of the current research status of deep learning in intelligent precision feeding of fish, and to provide strong theoretical support for the production process.

Prompt injection detection is commonly studied as a static offline classification problem, yet deployed LLM systems face evolving attacks and distribution shift after deployment. Static detectors are therefore poorly matched to the threat model, while routing every input to a stronger external LLM is costly and defeats the purpose of a local detector. We formulate prompt injection detection as a selective test-time adaptation problem. Our framework combines a prompt-based local detector built on masked language modeling and a learnable soft verbalizer with an entropy-based active querying mechanism that escalates only high-uncertainty inputs to an external LLM. Queried hard samples are then stored in a review window and replayed for subsequent detector updates. Empirical evaluations across multiple benchmarks show that EvoShield achieves performance on par with or even exceeding pure Large Language Model baselines, while cutting API query costs by more than 85%.