In each presentation class session, 9 teams will present.
In order to finish the class on time, we will strictly follow the timing of 120 mins / 9 teams = 13 mins per team.
This consists: 9 mins for the presentation + 4 mins for discussion
Practice and time your presentations before!
Every member should present at least once for the proposal and the midterm. It is encouraged but not required for everyone to present in each presentation.
Please make a team submission of the slides to Gradescope. Only one member per team should submit.
Your proposal report should be 3-6 pages for content and can have unlimited pages for references and the appendix.
Your midpoint report should be 3-8 pages for content and can have unlimited pages for references and the appendix.
Your final report should be 4-9 pages for content and can have unlimited pages for references and the appendix.
Overleaf project.
TAs will share an Overleaf project link with each team that will allow up to 10 collaborators. The link will be a copy of this proposal report template: link.
Free feel to reorganize the folders and files. For example, you can have a "proposal" folder or proposal.tex file for the proposal report; a "midterm" folder or midpoint.tex file for the midpoint report.
Please make sure you address the components in the guidelines below.
Please make a team submission of your report to Gradescope. Only one member per team should submit.
Title: This summarizes the problem and the novel solution that your project studies and can change later on.
Abstract: A few sentences summarizing your problem statement, proposed solution, and experimental setup. Describe the initial findings, if available.
1 Introduction
Problem and motivation
What problem are you trying to solve? Why is it important?
Have others tried to solve the same problem? What are the limitations of prior work?
What research questions are you trying to answer? How can the takeaways of your research be reused?
Proposed approach
What novel methods will you initially try?
Briefly, what data, evaluation, and models will you use?
How are they different from prior work or baseline methods? How do they address the limitations of prior work or baseline methods?
2 Related Work
Mention no less than 4 relevant papers with 2-3 sentences per paper. Then describe how your work is related to and meanwhile differs from each of them.
3 Method
What initial analyses will you conduct on baseline methods, if applicable?
What are your initial proposed methods? Why are these the right choices for your problem?
If your project establishes a new benchmark, then you will describe the evaluation dataset construction methodology, data quality control methodology (if you collect or synthesize new data), and evaluation methodology (e.g., metrics, LLM-as-a-judge).
4 Experimental Setup
4.1 Datasets
Data source
What datasets do you plan to use?
Do they already exist? Are they publicly available?
Do you need to create or preprocess a dataset, and if so, what are the steps?
Data splits
How should they be split into train, validation (development), and test sets? Has their been off-the-shelf splits, and if so, do they satisfy your needs?
You should use the same fixed data splits to train, validate, and test the baselines and novel models, whenever you want to make a fair comparison among these methods.
Use the validation set for finding hyperparameters, such as learning rate and epochs. Do not train or test on the validation set.
Use the test set only for the final report. Never use this set for training or development.
4.2 Evaluation Methods
How you will measure progress?
What are the evaluation sets to use?
What are the evaluation metrics to use?
Are these fair, correct, and comprehensive for the problem?
You should use the same fixed evaluation datasets and metrics to make a fair comparison among methods.
You may need multiple evaluation metrics to evaluate different aspects of the models.
4.3 Baselines / Methods to benchmark
What are the baseline methods that you will set up or reproduce?
For baselines from existing literature, will you directly use results reported in other papers, run inference on public model checkpoints, use existing codebase to re-train and evaluate baseline models, or implement the baselines yourself?
If your project is about benchmarking existing methods, describe which existing methods your will evaluate and analyze.
4.4 Implementation details
What off-the-shelf assets (data, model checkpoints, code) are available for you to build your project off of?
What are some software libraries that you will use?
What computational resources will you need for initial experiments?
4.5 Timeline
Provide a rough week-by-week timeline.
What would you like to accomplish by the midterm?
What would you like to accomplish by the end of the semester?
4.6 Possible challenges and mitigations
What is hard about your task and research question?
What are the foreseeable challenges that may prevent you from concluding the proposed study?
What are your contingency plans if things turn out to be harder than expected or experiments do not go as planned?
Title and Abstract
Same as the proposal report
1 Introduction
Components in the proposal report + a summary of initial results and findings so far
2 Related Work
Same as the proposal report
3 Method
Describe the motivation(s), overall pipeline(s), and details of your proposed methods so far.
4 Experiments
4.1 Datasets, 4.2 Evaluation Methods, 4.3 Baselines, 4.4 Implementation Details
Same requirements as Section 4.1-4.4 in the proposal report.
4.5 onwards: Show initial results and analyses of the baselines and your proposed methods in tables or figures. Discuss the findings.
If you have multiple result tables or figures that have different experimental setups, you need to describe the setup for each set of experiments. Feel free to reorganize the subsections flexibly according to your project needs.
No need to include the timeline, challenges, and mitigations, since there are less than 3 weeks left between the midpoint report due date and the final report due date.
Same as the midpoint, except that there should be more results, analyses, and findings.