From September 8 to 12, the premier computer vision conference ECCV 2026 convened in Malmo, Sweden, organized by the European Computer Vision Association and backed by major sponsors including Google, Meta, Apple, and Amazon.

On September 9, the MARS2 Workshop, themed around multimodal reasoning and slow thinking in the large model era, was initiated and successfully hosted by Tecdip, marking the sole Agentic Commerce-focused workshop at this year's ECCV led by a Chinese tech firm.

The MARS2 Workshop brought together elite academic forces, with an organizing committee featuring scholars from Tsinghua University, the University of Oxford, Nanyang Technological University, and Seoul National University. Keynote sessions featured four internationally renowned academics—Paul Pu Liang from MIT, Yarin Gal from Oxford, Shanxin Yuan from Queen Mary University of London, and Fahad Shahbaz Khan from Linkoping University—who delved into frontier topics including multimodal reasoning, long-chain inference, zero-shot generalization, and agentic systems.

Under the theme of advancing multimodal reasoning and slow thinking toward System 2 and beyond, the workshop tackled the core challenge of AI's evolution from quick perceptual processing to deliberate reasoning. Agentic Commerce scenarios, where AI agents make business decisions on behalf of users, demand such advanced reasoning capabilities—simply understanding visual input is insufficient; systems must locate evidence and articulate rationale, making multimodal reasoning the linchpin of this transition.

Beyond academic discourse, the workshop launched a multimodal reasoning challenge with a total prize pool of $100,000, structured around three core competency dimensions: overall semantic understanding, temporal evidence localization, and marketing strategy attribution. This framework formed a complete evaluation chain for AI's progression from perception to identification to comprehension.

The competition ultimately attracted 64 elite teams from around the globe, submitting over 1,060 solutions. Participants included scholars from institutions such as the University of Science and Technology of China, Nankai University, and Sun Yat-sen University, alongside technical representatives from companies like ByteDance, JD.com, and Xiaohongshu. Winning teams shared their technical approaches and key findings at the workshop.

The technical solutions have been compiled into a report titled "2026 Challenge on Multimodal Reasoning: From Multimodal Perception to Complex Reasoning," with the M-CAR benchmark dataset and complete codebase open-sourced on GitHub for global researchers to replicate and advance.

The challenge advanced global multimodal AI research and industrial application on three fronts, particularly in Agentic Commerce. First, it clarified the current boundaries of AI capabilities—results across the three tracks show that existing multimodal models perform well on perception tasks, but face a sharp performance cliff when confronted with commercial decision-making problems requiring multi-hop reasoning and temporal causal attribution, highlighting a clear shortfall in high-order reasoning and a gap from deployment-ready business agents.

Second, it revealed a cost-effective optimization pathway. Ablation experiments demonstrated that incorporating cross-modal spatiotemporal aligned audio event timelines improved localization accuracy by 16.7 points, while scaling model parameters from 4B to 8B yielded a marginal decline of 0.2 points, suggesting that under compute constraints, industry resources are better directed toward modality completion and input enhancement.

Third, it provided clear implementation strategies. Championship solutions adopted techniques such as Proposer-Critic dual-model validation, coarse-to-fine two-stage localization, and duration-adaptive token allocation, with the core philosophy of replacing parameter scaling with evidence chain engineering—decomposing reasoning into a closed loop of evidence acquisition, spatiotemporal alignment, and consistency verification. This points toward a more pragmatic direction for AI marketing products: the focus lies not on model size, but on reasoning pipeline design.

The competition outcomes indicate that the next leap in AI marketing capabilities hinges on designing traceable, verifiable reasoning processes rather than expanding model scale.

Tracy Chen, CTO of Tecdip, commented: "The successful hosting of MARS2 Workshop validates a key insight—AI agents are transitioning from concept to real business scenarios. Marketing stands out as one of the few domains that demands both complex intelligence and rapid validation of results. Tecdip will continue leveraging its dual strengths in technology and business integration, fostering deeper connections between academic frontiers and industry needs across more real-world settings, enabling academia to tackle practical business problems, industry to witness technology deployment, and global AI talent to engage in direct dialogue."

In her closing remarks, Zhang Yuxue, Tecdip's Brand PR VP, added: "The discussions at MARS2 Workshop showcased how cutting-edge technology addresses real business challenges, which is precisely why Tecdip organizes discussions and challenges at premier international academic conferences. Our goal is to bridge academic research and commercial application, working alongside global scholars to explore scalable Agentic Commerce technology pathways."

As the initiator of the MARS2 Workshop, Tecdip's proprietary "TecJi" specialized large model has delivered impressive results in global benchmark evaluations. In January 2026, the TecJi Q&A reasoning model ranked first globally with a score of 85.82 in the SuperCLUE advertising marketing specialized model evaluation. In July of the same year, the TecJi content understanding model secured second place overall with 86.43 points in the SuperCLUE overseas marketing video understanding category. From text reasoning to video comprehension, the TecJi model has earned authoritative recognition twice, and initiating MARS2 on this solid technical foundation represents a natural progression from proprietary development to ecosystem collaboration.

In 2025, Tecdip served over 100,000 brands in overseas expansion, reaching more than 200 countries and regions. The company's self-developed TecJi specialized large model and Navos marketing multi-agent platform evolve synergistically, continuously powering enterprise global marketing and growth. With this workshop as a new starting point, Tecdip will deepen its focus on Agentic Commerce, anchoring industrial-scale scenarios to propel AI innovations from the laboratory into the commercial arena. Through the MARS2 challenge series and the open-source M-CAR benchmark, the company aims to transform real-world business complexity into academic research testbeds, accelerating the paradigm shift of multimodal models from perception to reasoning.