news

한어Русский языкEnglishFrançaisIndonesianSanskrit日本語DeutschPortuguêsΕλληνικάespañolItalianoSuomalainenLatina

Synced

Author: Zhang Qian

In March of this year, Nvidia CEO Jensen Huang held a very special event. He invited the authors of the groundbreaking paper "Attention Is All You Need" to gather at GTC to discuss the future development direction of generative AI.

"Everything we enjoy in our field today can be traced back to that moment... You changed the world..." Huang Renxun said on the spot.

For a research-oriented industry like AI, the next opportunity to change the world may be hidden in a paper.

therefore,We have seen an unusual phenomenon in this circle: some CEOs with non-technical backgrounds have also started to stay up late to read papers, hoping to reduce the trial and error costs of decision-making.。

If this is the case for CEOs, let alone other practitioners in the field. Some time ago, OpenAI, Google, and Meta released big moves, and some startups are also constantly launching new models and new methods. I believe that many practitioners will feel that there are too many papers to read.

Giving AI a summary of a paper is a common reading method nowadays, but many AI summaries lack a clear hierarchy and detailed description of innovations and limitations. We need to go through several rounds of questioning to form a complete understanding of the paper. Moreover, some key model architecture diagrams and experimental result diagrams have to be found in the paper by ourselves, which actually saves very little time.

In TencentIngotIn the latest update of ", we saw the solution to these problems. Their new "Deep reading mode"supportLong article reading，Ability to output modular, illustrated analysis, very suitable for reading papers.

In order to verify the effectiveness of this new function, Synced conducted a first-hand test.

What is the point of "intensive reading" of a paper?

What is it like to use AI to read papers? In many cases, you throw a PDF to it, and it returns a summary + several overviews (sometimes up to 10). This information is indeed helpful, but sometimes it is difficult to distinguish what are the highlights, what the paper solves and what it does not solve, and what core issues are worth a closer look.

We found through actual testing that "Yuanbao" solves these problems by providing a series of modular and structured information.

Take a SIGGRAPH paper we tested as an example. If you throw the paper directly into it, the summary it returns is not much different from other AI. However, as long as you patiently scroll down, you will see a "deep read this document" button, which is the switch for "one-click direct access" to the intensive reading of the paper.

Different from the interface of the previous summary paper,Reading the pages carefully will break down the paper into very hierarchical levels, research background, research methods, experimental design, result analysis, and overall conclusion are each organized into a module, which is very similar to the layout of the paper usually introduced by Synced. All of these can be quickly jumped through the outline on the left.

Although each module has only a few words, these words are actually very informative. For example, in the research background module, the "Research Difficulties" paragraph describes four difficulties in just three short sentences, and the "Related Work" is a highly condensed version of the "Related Work" in Chapter 2, which clearly introduces the main technical routes in this field in one paragraph. Therefore, after reading this module, we can basically understand what problem the paper is studying and what the current research status is.

In addition to these conventional structured information, Yuanbao's intensive reading also has an eye-catching design——It will list the advantages and disadvantages of the paper., which helps researchers quickly understand what they can learn from this paper and what other issues are worth further research.

Why is this function so important? Peng Minghui, a professor at National Tsing Hua University in Taiwan, once wrote in an article about paper reading that papers are different from textbooks. Textbooks provide systematic knowledge that has been sorted and organized by others, while papers require readers to retrieve, filter, and organize knowledge from unorganized knowledge.The ability to analyze the strengths and weaknesses of existing research is particularly important. This is a key part of critical thinking and an important way to improve oneself in academic research.Yuanbao can help researchers save a lot of time on screening and preliminary understanding by quickly analyzing and summarizing the advantages and disadvantages of papers, allowing them to focus more quickly on papers that are directly related to their own research.

However, if you think the above information is too complicated, you can also jump directly to the lastKey Questions and Answers ModuleHere are a few of the most critical questions, so you can quickly understand the value of the paper and then judge whether it is worth spending time reading the original text. Of course, many previous AI assistants will also present some key questions at the end of the answer, and you can get the answer by clicking, but if you are a beginner or a reader with an interdisciplinary background, it may not be easy for you to judge which questions are more critical. Yuanbao's direct presentation method feels more intuitive.

Original pictures and numbers, who says AI can’t add pictures when reading papers?

When reading a paper, many people have a habit of looking at the pictures while reading the paper description. This makes it faster and easier to understand. However, most AI applications on the market now return text results, and if you want to see the pictures, you need to look for them in the original text.

We found in our testing thatYuanbao is one of the few AIs that directly captures the paper images and puts them in the corresponding position of the paragraph.For example, if the architecture is discussed in a module, the corresponding architecture diagram will be put up:

If you talk about the experimental results in a certain module, it will put the corresponding chart:

As we all know, the hallucination problem of large models has not yet been completely solved. Therefore, this presentation of the original image and original numbers is actually a more reliable output method, which is convenient for readers to verify the answers of the model at any time and is safer when used as a reference.

In addition, we also found that if you need to write a blog or other external output, Yuanbao can also help you draw charts, and you don’t need to tell it where to find the data. It can locate the relevant tables in the paper and extract the data for drawing. This function can be called up through the "Ask a Question" button on the right side of the intensive reading page.

You can study anytime and anywhere you want. Who says that studying for a paper is full of obstacles?

In addition to structured information and graphic output, we also found in our tests that Yuanbao actually has some very practical small functions that can make reading papers more convenient.

First isWord translation and search, these are two practical functions of the "original text" reading interface. Word translation can help readers who are not good at English to clear language barriers anytime and anywhere, and word search goes a step further, like making Yuanbao's search function into a plug-in, which can search for relevant information at any time. Moreover, the explanation given by Yuanbao is not just a brief summary, but also a modular expansion, which really makes "structured" and "informative" into every detail.

followed by"Offline reading". The practicality of this feature is that it allows you to review the intensive reading content and the original text in "flight mode" without wasting any fragmented time. This allows airlines to win back a round in the competition with high-speed rail. Maybe the next inspiration of researchers will come from reviewing intensive reading on the plane.

The last small function iscalculator". Some time ago, AI caused a lot of discussion because it couldn't tell which was bigger, 9.9 or 9.11. In Yuanbao, we found that it has integrated a calculator function, which can ensure that the answer is generated based on accurate calculation results. This function is very useful when we read experimental data.

Behind the intensive reading of long articles: It turns out that there are expert guidance

According to official information, Tencent Yuanbao's upgrade focuses on "long article intensive reading", which can natively support input of up to nearly 500,000 words. The papers we used in the test were far from this length, and most of the papers we come into contact with in daily life do not reach this length. Therefore, when using Yuanbao to intensively read papers, the context window is sufficient in most cases. Its modularization, graphic output, and small functions such as word search and translation also make reading papers truly convenient and efficient, and take another step towards "practicality".

This evolution is inseparable from the upgrade of the model behind it, the Tencent Hunyuan model. It is reported that in order to improve the professionalism and practicality of the model in the professional field, the Tencent Hunyuan teamWe specially invited experts in the field to summarize the core skills of each professional field and formulated the answer standards for professional questions., so that the model can serve as a real domain expert. So after using it, we feel that Yuanbao knows what information the paper readers need and how the information should be presented.

In addition to papers, this new feature can also be used for intensive readingFinancial reports, research reportsIn these scenarios, it can sort out information from multiple dimensions and generate professional charts such as DuPont analysis charts based on the report content, so that people who do not understand these documents can also understand the company's financial status and other information.

However, in terms of reading papers, Yuanbao still has some room for improvement, such as the lack of complete original text-translation comparison in the original text reading interface, and the recognition of formulas is sometimes not accurate enough. We also hope that Yuanbao can improve these problems in future updates.

However, as an app that has only been launched for more than two months, Tencent Yuanbao's performance has exceeded expectations. Its evolutionary trajectory allows us to see how large models will gradually become new productivity. We also look forward to this app bringing us more surprises.