The application provides a large
language model generation control method and
system based on time conditioning fine-tuning, comprising the following steps: obtaining training text with time attributes, disassembling the training text into fact units, and labeling available time for each fact unit; for the same query, constructing multiple training samples under different historical time boundaries, wherein each training sample includes the query, a time boundary, a set of available facts under the time boundary, and a set of prohibited facts; encoding the time boundary into a
time condition vector; inserting a trainable
time gate adapter into a pre-trained large
language model, the
time gate adapter receiving the
time condition vector and generating a gate coefficient for modulating the hidden representation of the large
language model. This method does not rely on a single prompt template, does not require permanent deletion of model knowledge, and does not require additional secondary review in each reasoning, thus having good stability,
scalability and deployment efficiency.