adding new games to clembench

[bakuzen]

Adding 5 games:

1. memory - a game where the LM is given contact info for 50 people (first name, last name, work, emial, and note where note is a unique attribute, like hobby) and then is asked about contact info based on the note.
2. memory_narrative - same game as memory, but all information about contacts is given in the pre-prompt
3. memory_turns - information about new contacts is given at each dialogue turn. One question about a contact is also asked about a contact given at a prior turn.
4. memory_narrative_turns - questions about contacts is asked for at each turn, but all information about contacts is given in the preprompt.

The four games show how well an LM can fish out information about a contact given a complete preprompt, but also if a LM can "remember" information during the dialogue outside of the prompt and game history.

5. simplesnake - a game for evaluating LLMs on 2D text-based spatial reasoning tasks.

[phisad]

Hi @bakuzen, do you have baseline results for memory? I cannot imagine that it is not played perfectly by an LLM, since all information is given in the prompt. Is this like a need-in-the-haystack task or do the contacts include some controllable level of ambiguity?

[bakuzen]

Big models have no problem, but when it gets up to, say 50 contacts medium and smaller models struggle. Likely due to lack of context window size, but there are limits if one wants to use even a good-sized model to fish through information like this. As for controllable ambiguity, we try to set it up where all individuals have a unique note, but we don't explicitly check.

[phisad]

OK, sorry for the late response. But I just now realize that this is the wrong repository. You should merge into clembench.