This page contains descriptions of some projects currently open in the dlab.

If you are interested in working on one of the projects listed below, please complete this form. (Please do not send your material via email; use the form instead.)

The project descriptions are rough, and there is generally flexibility. For instance, if you are looking for an MS semester project, but are really interested in a project listed as an MS thesis (or vice versa), this might still be possible. But please note that we have a preference for thesis projects over semester projects. If you want to get involved in real research, resist the temptations of an industry-based thesis and enjoy a semester of intense intellectual challenge in a research lab instead!

We consider applications 3 times per semester (see below for the deadlines). You will hear back from us at the latest one week after the application deadline for which you applied. That said, feel free to apply at any time. We might be able to take a look at your application even before the deadline, although we can’t promise. And needless to say, the sooner you apply, the more projects will still be open.

Application deadlines:

  • For projects in a fall semester: May 15, June 15, July 15
  • For projects in a spring semester: December 1, January 1, February 1

Applications received after the latest deadline will be reviewed on a case by case basis.

(Here you can find logistic info on IC semester projects; here, on Master’s thesis projects.)

LLM prompt optimization with reinforcement learning

Ever considered becoming a prompt engineer? Don’t apply for this project then! Rather than hand-crafting prompts and manually picking demonstrations, we aim to use reinforcement learning (RL) techniques to (almost) take the human out of the loop. More formally, we are going to solve the following tasks:
Demonstration sampling: Given a task instruction I and an LLM M, we want to automatically find a set of demonstrations D that will lead M to produce good outcomes on the task I. For that, we will train a separate LM S that maps I to D, leveraging the target model M’s performance as a reward signal.
Instruction sampling: Given a task, instead of simply finding demos for a fixed instruction I, we can get even more ambitious and optimize I itself! We can bootstrap from a weak human-produced I’ and train a separate LM S that maps I’ to I, where I optimizes performance of the target model.
Project type: MS thesis or semester project
Prerequisites: Solid ML background and prior exposure to NLP. Experience with LM training (e.g., PyTorch) is a big plus.

Symbolic auto-encoding for mechanistic interpretability of language models

The ultimate goal of mechanistic interpretability is to derive an interpretation of what a unit within neural networks does. An interpretation, in its broadest sense, is a sequence of symbols that encapsulates the information of a representation. This sequence could be a paragraph in the English language, a series of mathematical equations, or a snippet of Python code. It is not yet clear which symbolic system, with its specific structure and grammar, is most effective for interpreting the inner workings of transformer models. But what if we could learn this symbolic sequence directly from the models and data? In this project, we aim to train models to learn this symbolic representation in an unsupervised manner. Our recent research on discrete sequential auto-encoding has demonstrated the feasibility of learning this specific structure by linking sequence-to-sequence models symbolically. Unlike traditional autoencoders, which encode inputs into a single latent vector, symbolic autoencoders map input sequences to hidden sequences of latents, thus enhancing the representation space through the composition of sequence elements. This project involves developing symbolic autoencoders based on transformer models on a large scale and using them to interpret units in language models. Our focus is to build upon our symbolic auto encoding framework and evaluate these models across various models on different modalities.
Project type: MS thesis or semester project
Prerequisites: Primarily curiosity and enthusiasm. While it’s beneficial to have strong programming skills and a foundation in machine learning fundamentals, particularly in sequence models, this project also offers a valuable learning opportunity. Ideally, candidates should be able to navigate large codebases, adapt quickly to new libraries such as PyTorch, PyTorchLightning, and Hydra, and possess experience in writing clean, testable code. Familiarity with machine learning fundamentals and the inner workings of sequence models like Transformers is also preferred.