• Inspiration from how previous works were written

• Outline:

Be very selective with information. Don’t aim for a long report; it’s not yet a paper, just a stepping stone to explore routes to one. Just copy a few sentences from the draft.

1. Compare Input Prompts

test_prompt_most_recent_S.ipynb

Due to several tests, there is good evidence to support the following:

• HYPOTHESIS: If the “to-output” pattern is given in the source, in many cases, the high output probability “latest subject” will be outputted compared to the 2nd place output.
• HYPOTHESIS: However, this statement holds true for fewer subjects; as the number of subjects increases, it is less likely that the “latest subject” will be outputted.

Thus, we will only stick with inputs with 2 or 3 subjects in either source or target. For the sake of preserving length consistent between inputs, we will use 2 subjects in each, and only use single tokens for subject and description.

[Due to the lack of complexity in these tests, most will not be included in the report or its appendix, but links to them will be provided]

2. Finding Important Heads for Most Recent Name Movers

3. Figuring out Head Functionality for Most Recent Name Movers