# VIDEO 5 ALIBABA CLOUD RESEARCH

https://www.youtube.com/watch?v=5Jjcj3yvDfI

[00:01] Go see some posters.
[00:08] Who are you?
[00:08] Uh, I'm Tan from Alibaba cloud, Alibaba token hub motor as a service team.
[00:16] And currently I'm focused on working on some researchers in the domain specific applications like prosper customer e-commerce applications.
[00:23] And we have one paper set from ACL 2026, uh, focus on this applications.
[00:30] That's a big deal to have a paper accepted at I mean I, I have a video of all the papers just shown this morning and there's another set tomorrow, right?
[00:38] There's hundreds or thousands, lots of people, lots of papers, and very smart people too.
[00:45] Having conversation with them is a really, is a really great thing.
[00:50] I think the most valuable thing in attending this conference, so trying to translate your because a lot of these papers are way over my head, either they're very detailed about technology or math that's going on in.
[01:02] The large language model.
[01:04] What what is your paper?
[01:07] What was the innovation that brought this paper?
[01:09] Uh actually our paper focus on a very uh valuable applications the cross body e-commerce.
[01:13] So uh let's begin with this questions.
[01:19] Um the global trade is a $26 trillion market.
[01:28] So it's a so if we applying the AI agents the frontier techni technologies to this applications we can create a huge valable a huge values.
[01:38] So our paper makes focus on a very challenging task uh expert if search tasks on this cross body e-commerce and we uh found that the current frontier agents falling massively behind the super expert level uh with with over 50 absolute gaps.
[01:56] So it's a very challenging and very realistic and real world challenges for current frontier agents.
[02:01] Give give me an idea of what what an
[02:03] Agent will be able to do because of your paper that I wasn't able to do before.
[02:11] I think it's not what an agent can do, it's what a previous or existing works do or do cannot do.
[02:19] Uh, in the previous works the model focus on the uh level one data, we call the raw web content.
[02:25] For example, the uh very challenging deep search tasks like the blocks comp or the web arana and or the level two data, the structural uh the understanding and the utilization of the structural resources.
[02:39] For example, the database or the knowledge graphs.
[02:44] But current frontier agents can easily solve this problems or it's not a change for them.
[02:49] Uh uh we focus on the level three data, the theoretical rule data.
[02:53] Uh, in this data which uh the agent must to ret to call tools to retrieve the expert knowledge and rules to deduce a unique 10-digit code for every single crossport product.
[03:06] Shipment.
[03:09] This code is valuable because it dictates the accurate tariffs and the street compliance in the crossport e-commerce.
[03:18] So this, the level three heretical root data, exist three core challenges.
[03:25] First, the agent must conduct a heretical reasoning on the rules, and one single early mistake these cascaded into the total failure.
[03:31] And second, this rules have the regular bure semantic boundaries.
[03:38] For example, the excluding the terms like excluding, including, or for example use this rules create huge uncertainty for current frontier agents.
[03:48] And third, the logic dependency the rules are deeply intertwined or coupled, so it agents cannot see them in isolation; they must consider the relationship among these rules to handle such complex tasks.
[04:02] Wow, this is makes a big deal for Alibaba because they're always shipping saying.
[04:08] Things cross border, right?
[04:10] And like you said, they have to worry about tariffs or compliance or some things can't go to one country to another.
[04:19] And so if we're going to build a genetic systems that manage all that, it has to get it all right.
[04:27] Yes.
[04:29] Creates huge value for our view world.
[04:30] Yeah.
[04:32] Great.
[04:32] From your point of view, what's happening in AI right now, you know, and particularly at ACL?
[04:39] We're here at hanging out with 3,000 researchers that are building core pieces of the large language model or the Gentech harnesses that are going to run that model.
[04:50] What's going on from your point of view?
[04:51] Because I'm not a researcher.
[04:54] How do you look at this world?
[04:55] I have lots of conversations with some junior researchers or the very senior researchers.
[04:59] I noticed or in my view I think that the topic of the auto research for optimization is a really...
[05:08] Hot topic uh is similar to the self-improving or self uh involving agent uh because uh the the concept of this technology represents as agent and help them to improve themselves without some survivations.
[05:26] So that's a really uh promising future directions for current large language models or agents.
[05:31] For example, cloud recently uh show their blogs to prove that the cloud model itself can help the researchers to uh train their models to achieve the superhuman level.
[05:47] So I think it's a really future works.
[05:49] Well, thank you so much for giving me a little chance and congrats on uh getting did you is there a contest?
[05:56] Tell me how papers work in this world because because I see all these papers I walked around and and you know captured some of them.
[06:06] Actually we have two papers acceptab well the paper is not nominated.
[06:09] By the rest of paper committee, but we don't know if what's the what kind of award award is that.
[06:16] But uh, this uh, the monetary paper is mainly for the just we talked about the cross-body e-commerce, and we defined a new challenging problem: the heretical replications that agent must follows the expert knowledge as rules to deduce the 10-digit HS code for the you know the cross-body commas.
[06:38] And the agent must to follow this at each branch, it must to apply this rules to at a very uh in every step, and one wrong step means the total failure, so it's a very challenging task.
[06:51] And another paper uh in this year, our number of our teams, we define a new robust uh agentic harness for deep search tasks uh using the structure, using the structure the table and the for and the formulation to make the long.
[07:11] Horizon deep search more robust.
[07:14] Uh, specifically we, we uh, we reformulate the deep search ts or the search ts as the table completion.
[07:27] Wow, a very is easy formulation but a new perspective of this deep search and our experiments show that it can largely effectively improve the performance and robustities of these models in their long horizon deep search.
[07:46] You mentioned long horizon uh working the the Gent harnesses and and the models now are building systems that can run for months on earth hours or months, right?
[07:57] What's the challenge of of building that right now?
[08:04] You know to get it an agent to work on something for months without causing problems without going down and and I think the context and the memory is
[08:14] The real bottleneck for the agent for as a human we can sleep we can distribute our um uh daily experiments into something we don't know there something parous in our head so I think the memory and context how called the subconscious Yes.
[08:32] Yes.
[08:32] Yes.
[08:32] Yes.
[08:32] Yes.
[08:36] The how to utilize this context and memories effectively is the core to the for the months of the uh exe execution of this agent.
[08:43] I think this sport we're building a robot eventually.
[08:46] It's going to walk up to us and have a conversation with us and it's going to remember us.
[08:51] Yes.
[08:51] The next time we see the robot, right?
[08:53] It's going to have a very powerful memory and very powerful AI thanks to more personalized.
[09:00] Yeah, very personalized, right?
[09:02] It'll know what kind of music I like versus your music or what kind of sports you watch versus the sports I watch or what kind of color shirts you buy versus, you know, I have different shirts.
[09:13] All
[09:15] Right.
[09:17] So, and it keeps track of all that and makes everything easier to deal with, particularly if you're having automatic shopping in your home and do some work for us.
[09:29] Be a good colleague to to to work with us.
[09:33] I think as a robotics is very impressive.
[09:36] Yeah.
[09:39] Well, thank you so much for giving me a little taste of what your life is like.
[09:43] Congrats on the pay for acceptance and hopefully you uh win and uh thanks for giving me a little taste.
[09:47] It's a really great honor to have a conversation with you.
[09:49] Thank you.