Prospective Students

Oct 23, 2025 • Zhiming Zhou

Foundations Laboratory for Artificial Intelligence Research (FLAIR)

I’m looking for TALENTED students with GENUINE enthusiasm for FUNDAMENTAL research in machine learning/artificial intelligence.

• Overall:
  • I have a very high standard for research and require students to do RIGOROUS and IN-DEPTH research INDEPENDENTLY.
  • As the entrance assessment, you will need to CLEARLY explain a recent top-tier paper, while sharing your THOUGHTS on it.

• Our goal:
  • Research that pursues perfection and truth (which will naturally lead to publications in top conferences/journals).
  • I personally prefer theoretical/theory-driven research on fundamental topics in the field of machine learning.

• What will be your research topic?
  • We DO NOT assign topics to students in principle.
  • Find your own research topic by reading papers from top conferences/journals.
    • You may find papers via https://papercopilot.com/top-venues/.
    • You may use https://www.connectedpapers.com/ to find more papers related to a certain paper or topic.
    • I will always be there if you need any guidance.
  • Directions that you may give priority to:
    • Generative models: more principled generative modeling frameworks or their optimization/generalization techniques.
    • LLMs: techniques that improve the performance of (multimodal) LLMs or enhance their reliability.
    • AGI: mechanisms that give the model the ability to remember, reason, evolve on its own, and keep learning over time.
    • Architecture: more principled network building blocks, topologies/connections, and configs.
    • Optimization: parameter-free/adaptive optimizer, optimizers that consider characteristics of neural networks, etc.
    • Generalization: direct optimization of generalization, new mechanisms that ensure or improve generalization.
    • Efficiency: techniques that lower the resource requirements or cost of training or inference.
  • When you have read more than 50 papers and still haven’t found a suitable topic, I may give my suggestions.

• The recommended way of doing research:
  1. Understanding the state-of-the-art methods and/or theories.
  2. Identifying their major limitations or open problems in the field (focus on the high-level idea instead of the details).
  3. Attempting to address (one of) the identified limitations or problems in a principled way.

• Guidelines for reading a paper:
  • L1: Overall Idea (REQUIRED)
    • What problem is this paper concerned with?
    • What limitations of existing work motivate this study?
    • What is the main contribution of this paper? — Try to explain the high-level idea clearly.
  • L2: Detailed Analysis (whenever possible)
    • How does the proposed method work?
      • (If applicable) How is it trained? How is it used after training? — Understand the algorithm and data flow.
    • Why would it work as intended and better than the existing?
    • Why is it designed as such? Does it have to be like that? Are there any aspects that seem unjustified?
  • L3: Critical Assessment (have a try)
    • What is your assessment of this work? Is there any limitation? Can you do better in any aspect?
  • Warnings and tips:
    • Papers may contain minor errors. You may focus on the overall idea while dismissing some details.
    • Papers may contain misleading statements. You may focus on the mathematics, the most precise part.
    • Papers may not fully or clearly describe their methods. You may need to infer or supplement certain details.
    • Some ideas or designs may, in fact, lack solid justification. You should make your own judgments.
    • LLMs can hallucinate, while incorrect views may also spread widely. DO NOT accept any claim at face value.
    • For mathematical results, it is important to build an intuitive understanding beyond the formal derivations.

• My style:
  • I might be generally nice and supportive, but I can be RATHER CRITICAL of research.
  • You will be required to build a DEEP and CLEAR understanding of your research topic and related works.
    • This may take quite a while, but it is a prerequisite for us to start the real research.
  • I will guide and force you to think deeply, clearly, and logically. Hope we have the same belief that:
    • Superficial understanding may lead to specious conclusions.
    • Deeper thinking can bring simpler/better solutions.

• Management:
  • There will be a weekly meeting, which is typically online and one-on-one.
  • When presenting papers or other materials:
    • You NEED to prepare slides that SUMMARIZE and RESTRUCTURE the materials according to your understanding.
  • You will be required to write a diary report, which summarizes (along with the approximate time you spend on each part):
    • New things you learn (from papers or other materials) and your concrete understanding of them;
    • New thoughts about your research topic or related contents;
    • New experiments you have done as well as their analysis;
    • New codes you have written and the issues it deals with;
    • The issues that you are trying to address and your new attempts towards addressing them;
    • …
    • If nothing is done, then: The reason why you do not have any progress today.
  • There will be an ENTRANCE ASSESSMENT and a PROBATIONARY period of one semester for every newcomer.

• Characteristics that I value most:
  • [5’] The ability to think deeply.
  • [5’] Clear and strict logic.
  • [5’] Good intuition and critical thinking.
  • [4’] Good at math and programming.
  • [4’] Good (self-)learning ability.
  • [4’] Confidence and perseverance.

• Principle of authorship:
  • We determine the authorship according to the contribution. The tentative score-based rules are as follows:
    • Idea: 1 ~ 3 points, depending on the importance.
    • Theory: 0 ~ 4 points, depending on the importance and workload.
    • Coding & Experiments: 0 ~ 4 points, depending on the workload.
    • Writing: 1 ~ 2 points, depending on the workload.
  • The points of each part will be allocated to the participants according to their contribution.
  • For conferences/journals that typically do not regard authors as of equal contribution:
    • The authors with higher contribution points have a higher priority in choosing the position in the author list.
    • For authors that share similar contributions, the relatively lower-ranked can argue for a claim of equal contribution.
      • E.g., their difference in contribution points is less than 1.
  • In principle, people with too little contribution will not be regarded as one of the authors.
    • E.g., get less than 1 point, even though somehow participated in.