Driving the Development of the Agent Evaluation System, Mininglamp Technology (2718.HK) Achieves a Four-Dimensional Closed Loop of "On-Device Model—Inference Acceleration—Collaboration Platform—Evaluation Standard"

Recently, the achievements related to WebRetriever, a self-developed comprehensive evaluation benchmark for large-scale Web Agents by Mininglamp Technology (2718.HK), were officially accepted by the European Conference on Computer Vision (ECCV 2026), a top-tier international computer vision conference. The benchmark covers 800 real online websites and 1,550 tasks, spanning multiple industries including e-commerce, finance, healthcare, government affairs, and education.

It is understood that WebRetriever quantifies for the first time the real capability gap between an agent "reaching the correct page" and "truly completing the task," establishing an evaluation standard closer to real-world applications for this rapidly evolving field.

Quantified for the First Time: From "Building Agents" to "Defining Agent Standards"

According to data disclosed by Mininglamp Technology (2718.HK), WebRetriever revealed critical data for the first time in a real network environment: the current optimal model possesses a navigation success rate of approximately 45%, but its end-to-end task completion rate is only around 20%.

WebRetriever is precisely the first to separate "navigation success" from "task completion," allowing the industry to clearly see the true nature of this capability fracture for the first time. Addressing three major limitations of current mainstream evaluation methods—namely, limited data scale with a single domain, over-reliance of automatic evaluation on visual screenshots that fails to capture fine-grained interactive semantics, and equating "navigation success" with "task completion"—WebRetriever provides a systematic solution.

Its core technologies include the NavEval evaluation framework (LLM-as-Judge, achieving a 91.2% consistency rate with human expert judgments, significantly higher than the previous best level of around 81% among similar methods) and a three-level evaluation protocol (basic navigation, document-assisted navigation, and end-to-end full-link task execution) to accurately locate technical deficiencies.

For Mininglamp Technology, the release of WebRetriever holds multiple strategic values. ECCV is a top-tier international conference in the computer vision field. The acceptance of WebRetriever means that Mininglamp Technology's technical accumulation in the field of Web Agent evaluation has received authoritative recognition from the academic community. This is not merely a paper publication, but also an international endorsement of the company's technical capabilities.

In addition, it helps advance the Agentic AI strategy in depth. CSC Financial pointed out in its July 2026 research report that Mininglamp Technology is at a critical juncture in its strategic transformation from traditional data intelligence to Agentic AI. The release of WebRetriever represents an extension of this transformation at the technical infrastructure layer; the company is not only building Agents but also defining the evaluation standards for Agent capabilities.

In the Agent era, whoever defines the evaluation standards is highly likely to command a voice in the industrial ecosystem. WebRetriever serves exactly as Mininglamp Technology's strategic positioning in this direction.

The launch of WebRetriever forms a technical matrix alongside Mano-P, Cider, and Octo. In the first half of 2026, Mininglamp Technology open-sourced Mano-P, an on-device GUI model (ranking first globally on the dedicated OSWorld model leaderboard), Cider, an inference acceleration framework, and Octo, an Agent collaboration platform. The addition of WebRetriever allows the company to form a complete technical closed loop across four dimensions: "on-device model—inference acceleration—collaboration platform—evaluation standard."

AI Agents Hold Broad Prospects; WebRetriever Drives the Construction of Agent Industry Applications

The industrial background behind the release of WebRetriever is the explosive growth of the enterprise-grade AI agent market. According to IDC data, the market size of enterprise-grade AI agents in China reached RMB 21.2 billion in 2025, is expected to increase to RMB 44.9 billion in 2026, and is projected to surpass RMB 332 billion by 2029. IDC predicts that the annual growth rate of active agents will exceed 200% between 2026 and 2027.

Data from CCID Consulting shows that China's AI agent market size will reach RMB 13.53 billion in 2026, with a year-on-year growth rate exceeding 70%, where four major industries—government affairs, manufacturing, energy, and finance—will jointly account for over 70%.

However, the high growth in market size has not translated directly into extensive business penetration. A Gartner survey indicates that only 17% of enterprises have deployed AI agents, while more than half remain in the exploration or pilot phase. The reason is precisely what WebRetriever revealed: a significant gap exists between the demonstration capabilities and the production capabilities of current Agents.

Industry insiders pointed out: "The significance of this type of evaluation system is not merely to 'score' Agents, but more importantly, to help the industry find the direction for technical optimization, allowing enterprises to judge how far different Agent products are from actual application." WebRetriever is not just an academic achievement, but also an infrastructure construction that moves the Agent industry from "demonstration" to "application."

From a broader perspective, at a time when the user scale of AI search has surpassed 700 million with a penetration rate exceeding 68%, and when AI agents are beginning to undertake more and more work, the market requires not only "smarter models," but also "digital employees that can be verified, managed, and continuously optimized." The emergence of WebRetriever provides the first systematic evaluation framework for this exact demand.

It is understood that Mininglamp Technology will officially launch the WebRetriever Global Challenge in the near future. The competition will be open to global researchers, developers, and technical teams to test the real combat capabilities of their respective Web Agents in an evaluation environment closest to actual applications.

NewTimeSpace Disclaimer: All content herein is the original work of NewTimeSpace. Any reproduction, reprinting, or use of this content in any other manner must clearly indicate the source as "NewTimeSpace". NewTimeSpace and its authorized third-party information providers strive to ensure the accuracy and reliability of the data, but do not guarantee the absolute correctness thereof. This content is for reference only and does not constitute any investment advice. All transaction risks shall be borne by the user.