Welcome to the FLAIR Lab (Frontier Language AI Research Lab). We develop language AI agents and address fundamental challenges in robustness, reliability, safety, fairness, and adaptability to ensure that large language models (LLMs) and vision-language models (VLMs) are reliable and socially responsible. We are actively looking for self-motivated students!

Our recent research directions include:

• Robustness of LLMs and VLMs: Investigating the robustness challenges that lead to unpredictable behaviors in LLMs and VLMs, including vulnerabilities to adversarial inputs, inconsistencies across domains, and sensitivities to semantically irrelevant variations, with the goal of improving their reliability and stability in real-world applications.
• Multimodal Reasoning: Exploring how AI systems can effectively integrate, align, and reason over information from multiple modalities, such as text, images, and videos, to achieve a deeper and more coherent understanding of multimodal concepts and relationships.
• Multilingual Understanding: Studying how models acquire and represent shared knowledge across languages to enhance cross-lingual comprehension, facilitate knowledge transfer, and support equitable language understanding in globally diverse contexts.

In the era of AI, do we still need to learn how to code?

I often see people debating similar questions online: Do we still need to learn coding now that we have powerful coding agents? Should engineering interviews stop testing data structures and algorithms? If everyone can do “vibe coding”, is Computer Science still worth studying?

Before diving into these questions, I want to talk about what I consider the two core abilities of traditional programming:

1. The Understanding of Programming Language Knowledge: This includes syntax, keywords, built-in functions, and deeper programming language properries (such as memory management, or whether parameters are passed by value or by reference).

2. The Ability to Translate Abstract Problems into Machine-Executable Processes: I usually refer to this as programming logic. On a small scale, this involves deciding when to use a loop, wehn to use recursion, or choosing the most suitable data structure. On a larger scale, it is about how to efficiently integrate functions and classes, or even the communication architecture between modules.

I would like to share some tips on paper writing. A good paper is not just about solid research, the writing itself plays an important role too.

1️⃣ Define your target audience

Many students start their first paper without a clear plan in mind. In fact, you should first define your target audience and consider the expertise of them. Think about which techniques they will already know and which concepts might be new to them. Clarifying this will greatly improve your writing, especially when deciding which technical details to include and what wording is most appropriate to deliver your ideas.

2️⃣ Remember you are teaching, not reporting

This is a crucial concept. Writing a paper is not just about reporting what you did and what the results were, then passively expecting the readers to discover their value. Instead, you should proactively and strategically present your method in a structured way, moving from simple to complex and from concepts to details, to gradually guide the audience toward understanding the novelty of your work. I always use this analogy: imagine your audience as “a blank slate but smart.” They start with no specific knowledge of your topic, but they have the ability to quickly grasp new concepts. Your task is to decide what they need to know at every moment, guiding them toward a complete understanding of your work.

Should we rethink the bar for accepting a paper?

The pace of research in recent years has become incredibly fast, and I am not sure whether this is a good or a bad thing.

Looking back at my PhD years, it was completely acceptable to spend an entire year on a single project to achieve high-quality outcomes. Everyone understood that research requires significant time and effort, so each paper was considered valuable and worth careful attention. Reading papers was rewarding, and you could always learn from others’ work and save effort on exploration and trial and error. Researchers literally stood on each other’s shoulders to push the boundaries of knowledge.

This time I want to share three major red flags I notice when reviewing PhD applicants’ CVs. All three are related to publications.

❌ Not clearly marking papers that are still under review

It is completely fine to list papers that are under review on your CV, as long as they are clearly indicated. However, I have seen many applicants describe these papers in very vague ways, sometimes listing only the conference name without indicating the submission status. At first glance, this can easily make it look like the paper has already been accepted. This kind of ambiguity can be interpreted as an attempt to mislead the reader, which immediately raises concerns about the applicant’s integrity.

Reading research papers is an indispensable part of doing research. However, learning how to read papers effectively is far from easy.

Let me first list what I believe are the most important purposes of reading papers:

• First, to build a solid research foundation and develop relevant background knowledge.
• Second, to understand the research trends and challenges in the field.
• The third one is mentioned less often, but it is just as important: learning how to judge the quality and value of a paper.

Although reading more papers usually helps more with research, I would like to strongly emphasize one very important point: not every paper is worth reading in detail. This is so important so I will repeat it a few more times.

Not every paper is worth reading in detail.
Not every paper is worth reading in detail.
Not every paper is worth reading in detail.

Whenever I share my experiences applying to CS PhD programs, I’m often asked several questions:

What kind of background do you need to get into a PhD program? Do you have to have publications? Does the undergraduate school matter? How important are grades? Do applicants from interdisciplinary backgrounds have a chance?

The answer is that there is no single standard response to these questions. Different advisors value different things, and expectations also vary depending on the schools you are aiming for.

Here is one tip I often share for evaluating your target schools:

Take the time to carefully check if the schools you are interested in have labs that genuinely excite you. Then, look at the members of those labs and check what their CVs looked like when they were admitted to the PhD program.

I would like to share my perspective on what advisors care about most in PhD applications.

Passion, Curiosity, and Initiative for Research

Pursuing a PhD is a long-term commitment, and it’s inevitable to have moments of frustration along the way. Passion for research and the persistence to push through challenges are essential for success. Being proactive is also important. Students cannot rely solely on their advisor’s guidance. They should take initiative to explore independently, try new ideas, and even actively seek collaboration opportunities.

The nature of doing research is exploration and understanding, not merely in chasing state-of-the-art (SOTA) results.

Many students new to research tend to equate “doing research” with “achieving SOTA”. They believe that setting a new SOTA is the primary way to demonstrate novelty and getting published. However, this view somewhat mistakes cause for effect.

The goal of research is actually quite simple: to understand what makes an approach work, and why.

Research should aim to explain what works and what doesn’t, identify which design choices are the true driving factors behind improvements, and clarify under what conditions certain methods perform well. These fundamental insights and principles are the truly valuable and meaningful parts of research.

Are cold emails useful? In my opinion, if your resume is strong, a good cold email can definitely be beneficial.

So, what makes a good cold email?

First, you need to clearly understand the primary research focus of the advisor’s lab. For example, my lab is mainly doing NLP-related research, but I often receive emails from students expressing interest in “general AI applications”, “general machine learning”, or even areas like “pure computer vision” or “3D”. Such broad, vague, or misaligned research interests can make your email seem “templated”, and this may negatively affect the advisor’s impression of you. If you are trying to switch research fields, especially to one that differs from your prior experience, it’s important to explain why.