Baidu Search Achieves Recursive Self-Evolution, Baidu App Upgrades Four Major AI Features

In 2026, competition among AI Agents delved into a question closer to users: what kind of AI assistant does a person actually need?

Personalization, the ability to act, and long-horizon capability—these three have gradually become recognized metrics. Once clearly defined, the competition returns to a more fundamental question: where does a person's context come from, and can that context be turned into results?

The Baidu App product upgrade briefing held on October 30 can be understood along this thread: centered on understanding needs, delivering tasks, and providing continuous service, Baidu is pushing search to participate in more complete user affairs.

At this briefing, Baidu App upgraded features across four major AI directions. Baidu Search achieved recursive self-evolution and fully upgraded the Search AI API, priced during the event at as low as 4.9 yuan per thousand calls—lower than the DeepSeek Search API.

For Baidu, the commercial logic of this path is clear: search has long played the role of connecting needs with information, and personalized Agents give this accumulated foundation a new destination—extending from suggesting solutions to participating in the process of solving problems itself.

User needs are the first stop

Do you remember how users used the internet? For the vast majority of people who encountered the internet in the mid-to-early 2000s or earlier, their core early experience was almost inseparable from search or portal website navigation—essentially, finding the user's "first stop for needs."

Whoever a user first tells their goal to has the opportunity to participate in the subsequent information selection, tool invocation, and service connection.

When planning a trip, a user needs to look up guides, compare hotels, confirm transportation, and then adjust arrangements based on price and weather. Various tools are already quite abundant, but the work of finding tools, passing information along, and connecting steps is often still borne by the user. A complete need gets split across different apps and pages.

When AI can understand goals, retain the necessary task context, and invoke the corresponding tools, users have the opportunity to hand more coordination work to an AI assistant. For users, the value shows up as less repeated explanation, shorter operation paths, and more timely follow-up help.

This also explains why personalization is so important. The same trip planning—traveling with parents versus traveling alone—requires considering different conditions; the same organizing of materials—a client-facing report versus an internal discussion draft—has different requirements for information and expression. Only by bringing these specific conditions into the task can general capabilities become a service usable by an individual.

Judging from current industry trends, personalization is the first puzzle piece for all Agent products.

Muse makes memory the product foundation, letting AI remember users' long-term preferences and history; phone assistant products rely on system-level permissions to retain context across app operations; browser and desktop Agents leverage logged-in states to directly use users' existing account information.

The three paths differ, but what they want to obtain is quite similar: a continuous record that belongs to a specific person.

People express needs in many ways, and search is the most proactive of them. It requires almost no education cost—when users open the search box, they usually already know what they want. Where they are, when they ask, what they searched before, whether they care more about price or time—this information naturally attaches to the same search.

Accumulated over more than twenty years, Baidu holds the longest record of needs in the Chinese-speaking world.

In the past, one search delivered ten links, and filtering, comparing, and judging were all done by users themselves. Now what is delivered is an organized answer, or possibly something that can be used directly. Between the question in the search box and the result obtained, there are no longer a dozen or so jumps—so the role it plays has changed too: what it takes on has shifted from questions to tasks.

Looking further, continuous service may also change how users choose apps. To observe this competition, one needs to examine where a product comes into contact with users, what problems it takes on, and how it makes help happen continuously. Together, these factors determine whether a personal assistant can truly enter daily life.

It is reported that Baidu Search has achieved recursive self-evolution. Through an online feedback loop enabling multi-Agent collaborative self-evolution, Baidu Search precipitates every real retrieval and user feedback as training signals, which continuously flow back through a "retrieval—feedback—retraining" loop, driving recursive self-evolution of the model and enabling continuous iterative upgrades without manual intervention, with overall iteration efficiency improved by more than 80%.

When users open search, they usually already have some clear purpose—looking up a location, comparing a category of products, understanding a policy—each corresponding to some problem that needs solving. Search naturally sits between the expression of a need and the next step of action.

Baidu App therefore has a relatively direct service path: starting from the question the user has already raised, understanding the constraints behind it, and then pushing the task forward.

Search's accumulation is first reflected in taking on intent: a group of queries around the same topic can reflect the process of a need gradually becoming clear. When a user goes from "where to go" to "how to get there" and then to "is it suitable to bring elderly people," they are in fact continuously supplementing decision conditions. Under user authorization and controllable conditions, this context can help subsequent services reduce repeated communication.

Second, search connects to a continuously changing external world. Personal affairs are often affected by new information—a destination adds an event, a hotel adjusts its price, and the original arrangement needs to be re-evaluated. For an assistant to provide continuous help, it needs to obtain these changes in a timely manner and judge whether they are relevant to the current task.

This gives retrieval, information updating, and source judgment more direct task value in the Agent era. The model is responsible for understanding and organizing, while search provides the ability to stay connected to the real world. Only by combining the two can suggestions remain continuously close to the actual circumstances users face.

Baidu App's upgrade this time provides a concrete footnote. Wenxin Assistant takes on tasks in three ways: complex questions are pursued to the end through multi-step retrieval and organization, high-frequency needs are directly delivered as usable results, and long-running matters are continuously followed up after user confirmation.

For example, a travel case links itinerary planning, hotel price-drop alerts, and destination updates along a single thread, while key adjustments are still decided by the user.

English text is being prepared.

What these capabilities do is very concrete: they connect the links originally scattered among querying, understanding, and operation into a single thread. Search organizes services around users' goals, and personal assistants thus gain a way to be implemented without requiring users to change their habits. Users do not need to first accept a new set of reasons for use; they can start from familiar search behavior and hand over subsequent work step by step.

Baidu's opportunity lies in letting this transition happen across a large number of everyday needs. Every query may be the starting point of a service, tasks can continue to advance, and the relationship between search and users will also move one layer deeper.

The needs users raise span travel, work, study, and family life, and behind them are different kinds of professional knowledge, tools, and services. To expand coverage, there must be a supply system that can continuously supplement capabilities.

Competition therefore extends from product capability to the platform level. Front-end interaction determines how users express needs, the underlying information and tool supply determines how far needs can be satisfied, and effective matching and collaboration are also needed in between.

Baidu's layout unfolds along this direction. At this product upgrade communication meeting, Baidu newly upgraded its own search AI API, with the lowest price even lower than the price of DeepSeek's search API.

As the largest player in search, Baidu's search AI API has an AI-native four-layer architecture. The content layer gathers more than 80 languages worldwide and a high-quality index at the hundred-billion level, with the data credibility rate increased by more than 50%; the engine layer realizes the recursive self-evolution described above; the architecture layer supports high concurrency and low latency and has high-quality enterprise-level service guarantee capabilities; the application layer supports full access via API, MCP, and Skill, and after Wenxin Assistant is connected, the effect improves by 20%.

Information, applications, and system capabilities—the three things Baidu App needs—each have corresponding support, converging into a service system that can grow with demand.

As more developers use retrieval capabilities and more professional tools connect to actual needs, Agents will have the opportunity to cover more granular scenarios. For users, the result of popularization is very plain: when they encounter specific problems, they can more easily find suitable help.

As services go deeper, the key issue of trust comes into play.

For an ordinary query, users can check elsewhere on their own. Once AI participates in continuous tasks, whether information is timely, whether suggestions are grounded, and whether key operations can be stopped will more directly affect whether users are willing to continue handing over subsequent work.

Authoritative information, source traceability, and user confirmation should all be built into the service process rather than remedied afterward. The more solidly trustworthiness is done, the greater the possibility of going from occasional use to long-term use. For platforms, this long-term relationship constitutes a more sustainable competitive foundation.

In addition, the ability to continuously evolve is also important. Real needs are changing, and tools and services are also updating. AI products must be able to improve information acquisition, result organization, and collaboration methods from task feedback, translating capability improvements into more stable and more personalized services.

Baidu App has reorganized several long-accumulated assets. Needs provide the starting point of services, information and tools support task delivery, the open ecosystem expands coverage, and the feedback mechanism drives capability improvement.

The value of the entry point will be re-measured. In the past, people looked at how much information was connected and how many times it was invoked; now they also need to look at how many affairs it has helped users advance and whether it can maintain a continuous service relationship.

For Baidu App, the space opened by search's recursive self-evolution is to let a single search box take on more complete needs. Once users' questions have follow-ups, the value of search extends into a longer service process. Therefore, the value of Baidu App in the AI era should be redefined. (This article was first published in TMTPost APP)