Paying attention to connections
In the Transformer architecture, attention captures relationships between elements in a sequence. A token’s representation depends on its context: the word “key” in a description of a door connects to different concepts than it does in a database manual.
Definitions, examples and constraints in a prompt influence the continuation. This does not mean the model has checked the facts or permanently learned new information. Using context in a request and training model weights are different processes.
A small experiment
Choose a short passage and try these two instructions:
“Explain what this says.”
“Explain this passage to a beginner in three sentences. Preserve the main claim and one limitation. Use an example to explain any unfamiliar term.”
Compare more than readability. Check whether the meaning is preserved, whether any unsupported claims appear and whether the response follows the requested format.
What to record
- The exact instruction and source passage.
- The model name, version and available generation settings.
- Several responses to each variant, including unsuccessful ones.
A detailed prompt can make an answer more useful, but it does not guarantee accuracy. Check the facts separately from the writing style.
Architecture reference: Vaswani et al., Attention Is All You Need (2017). The prompts and comparison exercise above are a BLD AI teaching example, not findings from that paper.