You can find the main information about the course here.

The course is a seminar worth 6 ECTS by default. Note that the sessions are split into a 12:00 s.t.–13:30 block, a lunch break 13:30–14:15, and the second half of the session 14:15–15:45.

Schedule

session date topic
1 April 25th intro & overview
2 May 2nd core LLMs
3 May 9th prepped LLMs
4 (online) May 16th implications for linguistics
5 May 23rd implications for CogSci
  May 30th holiday
6 Jun 6th implications for society
7 Jun 13th optional initial project consultation
8 Jun 20th project launch
project work
9 Jul 18th intermediate project presentations
  Sep 1st submission deadline

Final projects

In order to get credits for this course, you need to complete a final project in groups of 3-5 students.

Important dates

The following important dates should be kept in mind!

  • June 13th: communicate your chosen project & names of your group members to us via email to Polina
  • June 13th: [optionally] attend the consultation session and talk to us about your selected project and any open questions you have
  • June 20th: [obligatory] project launch: present your selected project in a 5 minute presentation
  • July 18th: [obligatory] intermediate project presentation
  • September 1st: [obligatory] final project submission deadline

Project guidelines

Preliminary guidelines for the respective projects are the following:

  • projects to be completed in groups of 3-5
  • preliminary project presentations on July 18th
  • final project submission deadline on September 1st 23:59 ECT
  • the projects should result in submissions containing the following materials which will be the basis for grading:
    • building:
      • intermediate presentation
      • submit repository (or some other format of code sharing)
      • submit 1-2 page project report
    • testing:
      • intermediate presentation
      • submit repository (or some other format of sharing materials used for the tests)
      • submit 1-2 page project report
    • creating:
      • intermediate presentation
      • submit contents roughly corresponding to one in-depth topic discussion / group member
      • submit some visualizations or other presentation aids

Miscellaneous organizational matters

  • Feel free to use the dedicated Moodle Forum space for trying to find fellow group members.
  • Sign up for consultations during the four weeks between project launch and intermediate presentations on Moodle if you’d like to talk to us about your project progress.
  • Feel free to get in touch with us via email if you have any questions about the projects or would like to discuss your ideas.

Project ideas

As discussed in the slides for session 1, we suggest final projects that can roughly be grouped into the types “Build”, “Test” and “Create”. For projects involving coding, you can use the Python scripts accompanying single sessions as starting points for your implementations (but, of course, you don’t have to).

1. Build a “generative agent”

  • The paper on “generative agents” by Park et al. 2023 uses LLMs in innovative ways to implement agents with “deep personal characteristics”, goals (short- and long-term), planning flexibility and memory. While intended as non-player characters in immersive video games, this kind of modeling can also be used for research in computational social science simulations.
  • It would therefore be highly interesting to see how far we can reimplement and use such “generative agents”, e.g., to have argumentative or persuasive discussions with each other. This project would try to do this. It would be rather technical, exploratory, open ended, but potentially very interesting.
  • So, in this project you would reimplement a “generative agent”. You could use LangChain to implement a memory, planning, retrieval, reflection and action selection. Parts of this are already implemented in LangChain, but might be insufficiently documented or might just not work out-of-the-box. You would then have to improve on what is already there.
  • For some (simplified) agent implementation, you could try to have two or more agents interact with each other in dialogue, e.g., in bargaining or argument about a “hot topic”. It would be interesting to see whether agents argue differently based on how we define their characters, and whether they actually “change their opinion” in some sense or another.
  • Much of this is open-ended, so you need to be prepared to anticipate many technical problems and be willing to make many pragmatic executive decisions (e.g., how to measure the opinion of a generative agent) as you go along.

2. Testing statistical world knowledge and pragmatic ability in LLMs

  • A recent paper by Rohde et al. (2022) investigated human expectations of relevance of information via a simple behavioral experiment.
  • The goal of this project is to investigate whether LLMs have can be coerced to reproduce the human data, i.e., to see whether there is a sense in which LLMs have internalized statistical world knowledge and expectations of what counts as an appropriate, relevant communicative act.
  • The paper showed examples like this one to human participants (all material is included in the paper) and asked them to select one of two numbers as the missing piece in a test sentence:
    • Context: Liam is a man from the US. Liam lives down the street from Rebecca.
    • Test sentence: Rebecca thinks / announces to me that Liam has … T-shirts.
    • Task question: Choose the number which is more likely: 21 vs 29
  • The paper found that human decision makers are sensitive to the context (thinks/ announces to me, etc.) in the sense that unsolicited communicative acts raise the expectations of newsworthiness, which in turn corresponds to statistical unexpectedness (a higher number than one would normally expect).
  • Goal: Test common LLMs for their ability to reproduce human decisions for experiments 1-3 from Rohde et al.’s paper.
    • Start with zero-shot prompting.
    • Add appropriate instructions, possibly examples with Chain-of-Thought prompting.
  • Possible extension: Use LangChain to create a “pragmatic reasoning agent” in which (what you think are) relevant Chain-of-Thought steps are individually carried out by different calls to the LLM.

3. Automatic creation of experimental materials

  • With the advent of powerful generative language technology, we should ask ourselves: which tasks could we outsource to the machines; which ones should we (not) outsource?
  • One area that is relevant to science is the creation of text-based experimental material. Many studies in psychology, linguistics or elsewhere hinge on people reading text and giving (linguistic or other) responses. Often, we are interested in abstract phenomena and exemplify these phenomena with concrete vignettes. For example, if we are interested in dative alternation, we could use test sentences like:
    • Alex gave Bo the book.
    • Alex gave the book to Bo. We could also use sentences with different names, items etc.
  • When creating items, researchers in experimental psych. / linguistics may use implicit (unintentional) “aggressive design” by creating (text-based) stimulus material that is curated (e.g., by experimenter intuition) to promote the likelihood of the desired outcome (e.g., evidence for some hypothesis). Also, hand-crafting text-based stimuli can be time-consuming and effortful. AI-generated stimuli might help with both problems. But only, if we can be sure that AI-generated stimuli are valid substitutes for human-generated materials.
  • Therefore, in this project, you could look at ways of creating stimulus material for experimental studies, or benchmark data sets for LLM-testing, using automization via LLMs. The key to success are smart generation protocols (prompting, examples …) and, most importantly, good ways of quality control of the generated stimuli / test materials.
  • You could create new material similar to existing experimental or benchmark data for testing LLM performance, e.g.,
    1. a recent paper investigating LLMs Theory-of-Mind reasoning capacity
    2. Hu et al.’s pragmatic ability testing data set
    3. Argumentative strength: the corpus by Carlile et al. (2018) contains essays annotated with respect to argument components, argument persuasiveness scores, and attributes of argument components that impact an argument’s persuasiveness.
      • this might be interesting as materials for arg strength related projects
      • it is also interesting to investigate sensitivity of LLMs to argumentative strategies from the perspective of influencing the informational landscape.
      • the task would be to take the annotated passages of essays and try to vary them with respect to the quality of the single components of the argument. E.g., a passage might have an annotations with respect to persuasiveness (1 out of 4), eloquence (4 out of 6), relevance (5 out of 6). One could try to prompt LLMs in natural language to increase / decrese quality along single dimensions.
      • data can be found here (might time out first)
    4. linguistic benchmark based chain-of-thought annotations (CoT in a loose sense)
      • it would be interesting to have annotations of planning / reasoning steps accompanying solving “NLI”/”NLU” tasks from linguistic benchmarks, in a fashion similar to the STREET corpus
      • this would be helpful for fine-tuning various systems in the future, and interesting in order to understand what kinds of reasoning may be tested under the umbrella term “NLU”
      • example annotations to be produced for SWAG (original task: select best sentence continuation out of 4) (tapping into ‘world knowledge’ about likely human actions, likely reasoning, likely conversations)
        • On stage, a woman takes a seat at the piano. She a) sits on a bench as her sister plays with the doll. b) smiles with someone as the music plays. c) is in the crowd, watching the dancers. d) nervously sets her fingers on the keys.
        • CoT to be produced: Since the woman is on a stage and took a seat at the piano, it is likely that she is an artist and will perform something. She might be preparing for the performance and might open the piano or put out her music sheet next. Among the given sentences, sentence d) describes the most likely action of a person preparing to play the piano.

4. LLMs in education

  • Test current LLMs’ performance on tasks important for effective and safe employment in educational contexts, as described by Bommasani et al. (2021, p. 67ff.). These tasks include providing helpful feedback and instructions to students (in various subjects).
  • The goal of this project would be to test the applicability of current LLMs for educational purposes. A natural subject for this investigation is English or Math. The proposal focuses on testing LLMs for purposes of L2 English learning.
  • More specifically, the goal is to test whether:
    1. LLM captures learner’s mistakes (accuracy)
    2. LLM provides correct feedback, i.e., explains what is wrong and why(quality)
    3. LLM provides valuable explanation of different phenomena.
  • For instance, one could test their ability to do grammatical error correction for L2 English learners on the WI+LOCNESS corpus which contains corrected & annotated essays (download at Q&I+LOCNESS v2.1 – not where user is prompted to submit a form).
    • the corpus contains essays with various grammatical mistakes, like word order, determiner use, mistakes in tense and conditionals, and many more.
  • the task would be, e.g., to compare error correction performance and error explanations provided by LLMs to ground truth (maybe across prompting).
    • error explanations could be gathered from different sources, including online grammar learning resources.
  • @PT can provide more corpora / background on grammatical error correction systems if necessary
  • Alternative corpora of mathematical tasks and reasoning can be found in evaluations of current LLMs which are frequently tested on mathematical task solving. One good option might be STREET, which includes stepwise solutions.

5. Create a “frame problem” data set

  • In class, we discussed the “frame problem” as a foundational problem for classical AI. The example of the FP from class was that of a bomb attached to a cart in a hut. Rolling the cart out of the hut should (normally) imply that the bomb has also been removed from the hut. This is problematic for classical AI since all of the “inertia constraints” and their exceptions (things that do or do not change) need to be spelled out explicitly. But we also saw that at least some instances of LLMs give language output that suggests that some models do not suffer from the FP.
  • As far as we are aware of, the question of whether LLMs do or do not “solve” the frame problem has not been addressed. Some natural language understanding data sets (like GLUE and derivatives) contain examples that are related to the frame problem, but this is not systematically explored (AFAWCT).
  • To explore whether LLMs “solve” the frame problem, we could create a data set of stories / cases like the robot and the bomb in the hut, with dedicated questions and forced choice answers (“The bomb is in the hut” vs “The bomb is in front of the hut.”). We can then systematically test the performance of different LLMs on this new data set.

6. … projects on outreach / philo / ethics …

  • summarize the state of the art of an important discussion around LLMs by creating a:
    • term paper
    • video summary
    • podcast
  • topics can be chosen based on interest, for instance:
    • philosophical / ethical debates:
      • should AI source code be openly available?
      • what to prioritize: X-risks or Y-risks?
    • instructional video on how (not) to use LLMs in class / school