What Is a Research Idea?

I have noticed that many people who are just getting started in research have misconceptions about what a research idea is. Adding something to a learning objective or a model architecture, and assuming, based on intuition alone, that it will improve experimental performance, does not constitute a complete research idea—even if it actually improves performance. If someone who has worked with me for a long time still presents something like this to me as a research idea, I would feel deeply saddened.

A complete research idea tells a coherent story. There is a practical background and significance, previous efforts to address the problem, and limitations in those efforts. Next comes your analysis of those limitations, the thinking (or, better yet, the philosophy) that follows from that analysis, and a new method motivated by that thinking. An idea is usually a single point (as your supervisor may have told you before), but that point needs the support of everything above.

Everyone’s ideas start out simple and immature. Don’t worry, though—developing them usually takes a great deal of time and thought, even for relatively senior researchers. For me, I may spend months letting an idea take shape in my mind, and sometimes return to ideas I abandoned years ago. If you are just starting out in research, your ideas will certainly be far from perfect. But if you think along these lines and gradually build a complete structure in your mind, the process is useful both for the practical goal of getting a paper accepted and for the sincere desire to share your thinking.

Also, do not be a perfectionist. It is enough for an idea to feel good enough given where you are now. Thinking too much without ever taking action is not a good state to be in, either. What you should care about is whether your next idea or paper shows progress over your current one. As long as you are making progress, that is enough.

It is also important to remember that logic is at the heart of an idea. This is why, after hearing what a collaborator wants to do, I sometimes say “reasonable,” and sometimes hold my head in frustration because I cannot follow the logic of the story being told. When it comes to logic, what matters the most is that the reasoning flows clearly and that the claims can be checked. The connection between each claim and the next should be clear and easy to follow, and the claims on which the argument rests should be verifiable, whether you plan to test them directly through experiments or theory, or justify them by citing some previous papers.

Having several points is fine, too. But it is usually necessary to distinguish the main idea from the supporting ones and organize the writing around the main idea. Otherwise, the paper becomes difficult to write and difficult for readers to follow. (You are trying to tell them something interesting, not leave them feeling confused or frustrated, right?) From this perspective, there is still one central point, with the others serving as extensions or further discussion.

This also helps explain why you often see papers built around naive ideas get accepted. Although there is considerable randomness in the review process, these papers with high probability tell a complete and convincing story. In my view, the quality of this whole story determines whether a researcher is competent. Beyond that, the choice of research direction—what field to work in—is what distinguishes the very best researchers from ordinary ones like us. This is a matter of research taste, although luck also plays an important role.

The only thing we can do is read more papers across different fields. But it is a mistake to approach them like a high school student who takes everything as truth. After all, many papers have problems, not to mention those written for questionable reasons and published for reasons that are hard to understand. So we need to learn to distinguish good work from poor work. A paper from a leading research institute, or with a famous name among its authors, is not automatically a good paper. Nor do you have to consider a paper good just because everyone else does. Judge it for yourself and think about it.

To me, a good paper either offers a thorough and insightful discussion of the background and related work, helping me become familiar with a field I may not have known before and understand its ultimate aims or underlying philosophy, or has thoughtful ideas and fancy techniques. It is interesting to understand those ideas, and even more interesting to understand the insights behind them. Even a few words can inspire your own thinking or give you a new understanding of something.

Read more. Judge more. Think more. Reading and thinking more are always worthwhile.

How to Write a Paragraph

Before discussing how to start writing and organize a paper (once the idea is clear, writing the paper becomes much easier), let’s first consider a more basic question: how do we write a paragraph?

First, I have nothing against using GPT or Claude to help with writing. This post itself is being written with GPT. But at least for now, GPT cannot infer your reasoning from nothing or from a few scattered remarks. Even when you give it the whole story, it often produces writing that looks correct on the surface but is hollow underneath, and sometimes even misleads you.

I have tried taking a shortcut by asking GPT to write a first draft, intending to revise the logic and wording afterward. However, I was completely misled by it and spent even more time figuring out what was wrong with its wording. In my judgment, GPT cannot yet handle complex discussions at the level of an entire paper. It is better suited to helping you develop the text when you provide the specific reasoning, or discussing a concrete question with you. Even then, its writing is far from reliable. Many logical problems remain, and you need to read it repeatedly.

When I wrote my first paper many, many years ago, without GPT at that time, I probably read and revised each paragraph hundreds of times. With GPT, that much repetition may no longer be necessary, but it is still worth spending a day examining the reasoning: is the logic crystal clear, and do the transitions between paragraphs flow naturally? This requires repeated reading and thinking, both from the reader’s perspective and from that of a professional researcher striving for precision and rigor. Looking only at metrics such as AI-detection scores is a form of self-deception. It offers a false sense of assurance.

We want readers to understand what we mean without difficulty, while also expressing our own distinctive thinking, perhaps even our philosophy. Needlessly complicated explanations and deliberately making the text harder to read are forms of mystification that reflect a lack of confidence in ourselves and a lack of professionalism. The simplest and most common way to organize a paragraph is to begin with the main point and then elaborate on it. The first sentence usually connects to the previous paragraph and makes clear what this paragraph will do. The sentences that follow should serve that purpose, with reasoning that flows clearly and smoothly. As the paragraph draws to a close, try introducing a new point that leads into what follows.

There are also a few additional issues I would like to discuss here.

First, there is a difference between logical leaps (逻辑跳) and a lack of logical focus (逻辑散). Some people confuse the two, perhaps because of how they are expressed in Chinese, but they are fundamentally different. The distinction is a little easier to understand in English.

Logical leaps mean that the underlying logic is there, but its presentation is not quite smooth. This is like a mathematical proof in which someone extraordinarily clever or with strong intuition skips a few steps: the conclusion can still be rigorously justified, even though the reader needs those missing steps to follow the reasoning.

A lack of logical focus is a more serious problem. It means that you do not really know what the current discussion is meant to accomplish or which specific claim it is meant to support. This is like listing facts that appear correct in a mathematical proof, without knowing what conclusion you intend to establish or being able to derive a sound conclusion from them. If this is the case for you, remember that the first priority is to know exactly what you are doing.

Second, I want to discuss the difference between a lack of information and high information entropy. Both are serious problems, and neither seems to receive enough attention from many of my collaborators. Usually, I rely on intuition to judge whether each sentence is as good as it can be. But when I am really unsure or trying to explain a problem to a collaborator, apart from flawed reasoning or a lack of justification, my most common diagnosis is that the sentence provides no useful information or has too much entropy.

By a lack of information, I mean that the sentence does not support your argument and only makes things more complicated. Even if it contains something useful, it may still place an unnecessary burden on the reader. If you really want to keep it, the best approach is to find a separate place to discuss it later in the paper or in an appendix.

By high information entropy, I mean that a sentence is ambiguous, either in its wording or in its structure. For example, in a passage I read recently about agents, the word “update” was used without specifying what was being updated. But in the literature discussed in that passage, the harness, memory, and even the LLM can all be updated. “Update” needs further qualification, such as “LLM update,” to make its meaning clear. Eliminating this kind of ambiguity is an important goal when you repeatedly read through your writing to check it.

A final, relatively less important point concerns word choice. Beyond ambiguity in terminology, there are many other details to examine: adverbs that qualify a claim accurately, appropriate connectives, and even the implications of the verbs themselves. I used to look up synonyms for a verb and repeatedly consider the subtle differences between them before choosing the most suitable one. I remember that, a long time ago, I would often spend an entire afternoon thinking about which word to use in a single place. That habit has stayed with me. I still often challenge GPT about whether a particular word or sentence implies something I do not intend to claim.

Terminology should also be defined when it first appears. This is an important convention in scientific and technical writing. Your readers may come from different fields and may not know everything that is familiar to people in your area. A word may also mean different things in different fields. It is your responsibility to remove that ambiguity. From the perspective of making the paper self-contained, you should also define every term you use. Whether the work is theoretical or engineering-oriented, your goal should be to make the paper understandable with as little effort as possible, even for readers new to the field.

Consistency in wording also matters, especially for terms you have already defined. Do not keep switching between different names. Once something has been precisely defined or discussed, its terminology (whether nouns or verbs) should remain consistent throughout the rest of the paper. Otherwise, the writing falls short of the rigor it should maintain.

How to Write an Abstract

The most important principle in organizing a paper is to put everything where it belongs, or where readers expect to find it. A straightforward example is that your paper may contain many interesting findings, but clearly not all of them belong in the abstract. First, readers may no longer be able to read at the pace they find most comfortable and familiar. Second, your abstract might end up being several pages long. The same principle applies to the other parts of the paper.

So let us first discuss what a standard abstract should do. Let us assume you already have a story in mind (you need to have thought through your idea before you start writing, right? At least, you should have a complete story by now, even if it is still immature). The abstract’s purpose is then to tell that story as precisely as possible in as few words as possible.

Like the story we discussed above, a standard abstract consists of several parts. Let us take a typical paper that proposes a new algorithm as an example. The first sentence usually tells readers what problem we are studying. You can introduce the broader research area or, more directly, state that you are studying an important problem within that area. The second sentence usually explains why the problem matters, whether in practice or in research (we need to convince readers that the problem is important, right? If it is not, no matter how well you conduct the research, it will not be meaningful), and gives a broad view of how people in the field understand or approach it.

In the next one or two sentences, we usually acknowledge the contributions of previous work very briefly, but then turn to its limitations. You do not need to list every limitation you can think of in the field (we want a short abstract, right?). Focus on the limitation you aim to address. You may need to define or describe it briefly and emphasize how it seriously constrains progress in the field. Next comes your contribution. To keep the reasoning flowing smoothly, we usually first analyze why this limitation arises or what previous work has overlooked. This discussion should, in turn, motivate your new method for improving on previous work. You need to describe its core contribution briefly but insightfully and convincingly, and can try to connect it back to your analysis above by explaining how it addresses the limitation. The abstract usually ends with a brief statement about strong experimental results or the broader implications of the work.

What matters most in an abstract is telling your story and explaining the logic behind it in the simplest, most direct, and shortest way possible. You need to trim your story down to its essential reasoning. A good abstract is crucial. I only write mine when I feel mentally fresh or when my thinking is most active, and then spend several days revising it to make the logic rigorous and easy to follow. Write for the reader, not for your own satisfaction. The abstract matters because it provides the skeleton of the entire paper and guides its organization. In fact, the introduction is a more detailed account of the abstract, and the rest of the paper develops the introduction further.

Once the logic is sound, make sure that no sentence or part of a sentence in the abstract is wasted. If a sentence does not help readers understand what we are doing or what we are thinking, delete it. On the other hand, avoid empty descriptions. Everything should be informative, and readers should enjoy reading the abstract and find it easy to understand. If an additional description or definition would help them follow the story, add it. This often takes the form of a subordinate clause or a parenthetical remark, but a separate sentence is fine, too. You need to develop your own feel for these tradeoffs. We want an abstract that is easy to understand, informative, and short.

In general, I feel that the shorter the abstract, the more skillful the writing. Of course, this only holds if readability is not sacrificed. If your abstract is both easy to understand and short, that is really good. But if you find that your abstract is very complicated or struggle to write one of a typical length (look at a few papers and you will get a sense of what that means), it usually means that you either have a bad idea or have not really thought through its logic. At that point, take a step back and think carefully about where the problem lies.

Ideally, someone working in exactly the same area should understand what you have done after reading your abstract. It is even better if it sparks their interest enough to read the method section or study the paper carefully.

Researchers from other areas within the same broader field, such as other LLM researchers, should come away from your introduction with a sense of the area, an understanding of its current state, and a clear picture of your contribution. People just starting out in research and undergraduate students should be able to follow the more detailed preliminaries, methods, and analyses that come later. These parts should teach them something and offer guidance. Your goal should be to make what you write understandable to everyone with as little effort as possible.

How to Write an Introduction

We have wandered a little off topic. Let us get back to the question: how should we write the introduction? We said above that the most important task of the abstract is to make the logic clear as simply as possible. What matters most in the introduction? I think it is to tell that story again, this time in a self-contained way. By self-contained, I mean that readers should be able to understand your paper while consulting as little additional material as possible.

For example, simply saying “we follow the same idea as X to do this” is not self-contained. That does not mean you should stop citing X. Instead, say “we do A to achieve B, based on the insight that C, following the same idea as X.” Can you see the difference? Readers can keep following your reasoning without having to look up X, because you have already explained at a high level what X does. If this is important enough, it will need a more detailed explanation in the preliminaries or method section, right? That is also part of making the paper self-contained. You can then add a cross-reference, such as “(cf. Section …),” to direct readers to that explanation.

Being self-contained also means defining and describing the concepts you discuss. You cannot assume that readers know every concept you mention. These definitions and descriptions should appear where readers need them. Keep thinking about where each concept should be introduced and explained, and look for undefined terms that readers may not understand. This is also part of putting everything where it belongs.

As we discussed above, the introduction expands on the abstract. Each sentence, or group of sentences, in the abstract often corresponds to a whole paragraph in the introduction. Of course, the introduction may include more background, a fuller account of previous work, or a separate paragraph developing an insight or analysis. But its overall reasoning should still follow the abstract. A tightly reasoned abstract makes it easier to build a tightly reasoned introduction, though you still need to work carefully on the connections as you write.

The structure of an introduction is usually fairly standard. Let us take LLM unlearning, an area I know well, as an example. In the first paragraph, I would introduce the broader background that leads to unlearning. I might explain that LLMs are important and widely used, but that their deployment raises widespread privacy concerns that hinder further progress. This establishes a problem that lets the second paragraph introduce unlearning as a promising direction for addressing it. I would then define unlearning, explain at a high level how it addresses privacy concerns, outline the main approaches and representative papers, and perhaps offer my assessment of their relative strengths and weaknesses.

The first paragraph places your work in the context of a major problem in the broader field, such as privacy. If you do not know how to write it, do not worry. Read a few other papers: within a field, this paragraph often follows a fairly standard pattern. Of course, it is even better if you have your own perspective to offer.

The second paragraph, as you can see, serves part of the purpose of related work. This is now a common approach. Compared with placing that discussion entirely in a separate related work section, it saves readers unfamiliar with the area from having to interrupt their reading and look elsewhere in the paper for background. It is much smoother to give them an overview in the introduction: the main categories of previous work, how each broadly approaches the problem, and their respective strengths and weaknesses. You can still include a more detailed related work section later, often after the experiments or in the appendix. But the discussion in the introduction should be more selective. Keep it short while giving readers an overview, and organize it to help them understand your story or idea. The goal is to reduce the burden on the reader.

If necessary, you can devote a separate third paragraph to the limitations of previous methods. More importantly, analyze why the particular limitation you emphasize arises. As we said before, this analysis should be focused and lead naturally to your method. More precisely, the reasoning should run the other way: your analysis is what motivates the design of your method. The analysis needs justification. This can come from findings in previous work, an explicitly stated assumption, a theoretical proof, or direct or indirect experimental evidence.

Then comes the most important paragraph: introducing your method. Based on your analysis, explain intuitively why it can address the limitation of previous work. You can then discuss the framework, additional benefits, or anything else you think belongs in the introduction. Here, too, you need to make choices. What is most important? What is secondary but still worth mentioning in the introduction? What belongs in another section of the main text, and what should go in the appendix? The guiding principle is still whether readers can understand your work with reasonable effort. Never exhaust their patience. Keep the reasoning clear, avoid rambling, and keep the introduction within a typical length. As a general guideline, the abstract and introduction together should take no more than two pages.

After discussing the method, you can briefly report experimental results or refer to your theoretical analysis to support its rationale. Finally, you can use one or two sentences to outline the new thinking you hope your work will inspire.

Some authors then add another summary of their contributions, such as “Our contributions are threefold.” From what I have seen, this format was once popular in papers from US institutions and later became less common there, though it still appears in papers from Asian institutions. Personally, I do not favor it. If the writing is clear enough, readers should already understand the contributions. I also see it as a form of pandering to reviewers.

The introduction should be self-contained, simple, and easy to follow. Ideally, each paragraph should focus on one thing, with appropriate transitions into and out of it. If a paragraph needs to discuss several parallel points, organize them clearly. The simplest option is to number them 1, 2, and 3 so readers can follow your progression. Keep the reasoning tight throughout the introduction. Wherever possible, begin each paragraph with its main point and then elaborate on it, and pay close attention to the connection between the end of one paragraph and the beginning of the next.

How to Write the Rest of the Paper

Next, how should we write the rest of the paper? Let us begin with the preliminaries, which are relatively straightforward to write. This is the best place to define the problem precisely and describe it using mathematical notation. We usually define the data, the model, and the problem we are studying, and introduce an appropriate formulation.

I also like the preliminaries to serve part of the purpose of related work. I briefly discuss why the area matters and what makes it challenging, then introduce some classical methods that are closely connected to the new method we will present later. We did something similar in the introduction by outlining the main lines of existing work. Here, however, we can focus on the line our own work follows.

Within a reasonable length, we can include more detail: the specific formulations, the insights behind them (the insights matter most, right?), and their strengths and limitations. Their strengths explain why this line of research is worth pursuing, while their limitations help us transition smoothly to the discussion that follows. Everything serves the goal of making the reasoning flow more naturally.

What comes next can vary. If the limitation you discuss has not been examined in previous work, is relatively complicated, or deserves closer analysis, you can include a separate analysis section. I like this approach because it makes the reasoning more rigorous. Here, we can carefully examine what is wrong with previous methods, how much these problems matter in experiments, and why they arise, using intuitive reasoning, empirical evidence, or theoretical analysis.

Like every other part of the paper, this detailed analysis needs clear reasoning. Develop it around one central thread, or work through the problem layer by layer. Whatever part of the paper you are writing, readability and clear logic matter most. Highlight the conclusions of your analysis and use them to lead into the method section. Even without a separate analysis section, you should devote a paragraph to explaining the problems with existing methods, supported by appropriate references or cross-references to your own experiments or appendix, to make the discussion more convincing.

Next comes the method. Its central purpose is to explain our insight into how to address the problems with previous methods, based on our analysis. Presenting the full pipeline should come only after that insight is clear. Otherwise, readers will struggle to understand what you are trying to convey. We cannot expect them to work through the details paragraph by paragraph without knowing the underlying idea, and then somehow connect with us on a spiritual level and automatically grasp our profound thinking by the end, right?

Start with the main point and then elaborate on it. Explain the insight first so readers can see that it is reasonable and feasible. They will then have a reason to examine the details with a clear sense of what to look for. Those details should also help readers understand the insights behind the design, rather than simply list the steps. Again, aim for clear reasoning, readability, and a simple structure.

The experiments are relatively straightforward to organize. Divide them into sensible parts, such as simple settings, more complex settings, and real-world applications. Begin by introducing the datasets and baselines. Many baselines will be methods discussed in the preliminaries, supplemented by more advanced or recent approaches. Then describe the experimental setup, including the models, hyperparameters, and any additional tricks.

Discuss the results for each part: why does your method outperform the baselines, what explains the improvement, and why does it matter? This discussion should still follow your story, connecting back to your analysis of the limitations of previous methods and your insight into how to address them. Again, communicating insights matters more than listing details. Few readers can remember every detail of an algorithm, but its underlying insight is much easier to remember.

Finally, there is the conclusion. This is usually a short paragraph summarizing your contribution and its broader impact. But if you have more to say and want to better position your contribution, you can write a full page. What matters most is still communicating your insights and thinking: how might your work inspire further research, or potentially change existing systems or paradigms? The conclusion should convey these ideas rather than simply list what you have done.