Tec-Do Technology Leads ECCV 2026 MARS2 Workshop on Agentic Commerce

ECCV 2026 is organized by the European Computer Vision Association and lists Google, Meta, Apple, and Amazon among its main sponsors. The MARS2 Workshop, formally titled Multimodal Reasoning and Slow Thinking in the Large Model Era: Towards System 2 and Beyond, assembled an organizing committee with scholars from Tsinghua University, the University of Oxford, Nanyang Technological University, and Seoul National University. Its keynote speakers included Paul Pu Liang of MIT, Yarin Gal of the University of Oxford, Shanxin Yuan of Queen Mary University of London, and Fahad Shahbaz Khan of Linköping University. Their talks covered multimodal reasoning, long-chain reasoning, zero-shot generalization, and agentic systems.

The workshop focused on a shift from fast, perception-like AI to slower, reasoning-driven AI. In Agentic Commerce, where AI agents may increasingly act for users in commercial decisions, simply recognizing content is not enough; an agent must locate evidence and explain why a decision is justified. Multimodal reasoning is presented as the key to that transition.

The MARS2 multimodal reasoning challenge offered a total prize pool of $100,000 and was built around three tracks: MAC for overall semantic understanding, VTG for temporal evidence grounding, and MDC for marketing strategy attribution. Together they formed a chain from seeing to finding to reasoning. The challenge attracted 64 teams from around the world and more than 1,060 submissions, according to Leiphone. Participants included researchers from the University of Science and Technology of China, Nankai University, and Sun Yat-sen University, as well as engineers from ByteDance, JD.com, and Xiaohongshu.

QbitAI reported that the tracks used the M-CAR benchmark, built from Tec-Do's real commercial scenarios. It contains 3,108 advertising videos in more than 30 languages, totaling 36.5 hours and split into 18,198 semantic segments, with an average segment length of about seven seconds. Contestants ran models locally and uploaded results to EvalAI for unified scoring. The top MAC score was 86.67, while VTG and MDC topped out at 62.27 and 63.53. The Boys won MAC and MDC, and Ya PTers won VTG; The Boys, described as a ByteDance-related team, took two first places and one second place. MAC and MDC models were limited to 14 billion parameters, VTG to 8 billion, and only open-source weights were allowed.

The results showed that current multimodal models perform relatively well on System 1 perception tasks but drop sharply on System 2 commercial decision problems that require multi-hop reasoning and temporal causal attribution, according to Leiphone. An ablation found that adding a cross-modal, spatiotemporally aligned audio event timeline improved localization accuracy by 16.7 points, while scaling a model from 4 billion to 8 billion parameters produced a 0.2-point decline. The finding points to modality completion and input enhancement as more cost-effective than parameter expansion under compute constraints. Champion solutions used Proposer-Critic dual-model verification, coarse-to-fine two-stage localization, and duration-adaptive token allocation. Their common approach was described as evidence-chain engineering rather than parameter stacking, breaking reasoning into evidence acquisition, spatiotemporal alignment, and consistency verification. The technical report is titled 2026 Challenge on Multimodal Reasoning: From Multimodal Perception to Complex Reasoning, and the M-CAR benchmark and code were open-sourced on GitHub.

Tracy Chen, CTO of Tec-Do Technology, said the success of MARS2 validated a judgment that AI agents are moving from concept to real commercial scenarios. Marketing is one of the few fields that requires complex intelligence and can quickly verify results, she said, adding that Tec-Do will continue to use its technology and commercial closed-loop advantages to connect academic frontiers with industry needs in more real scenarios. Zhang Yuxue, vice president of brand public relations at Tec-Do, said in closing remarks that the workshop showed how frontier technology can solve real business problems, and that the goal was to open a channel between academic research and commercial deployment and explore reusable, scalable Agentic Commerce applications with global scholars.

As the initiator of MARS2, Tec-Do has built its own Tec-Chi professional large model. In January 2026, the Tec-Chi question-answering reasoning model ranked first globally in the SuperCLUE advertising and marketing professional large model evaluation with 85.82 points. In July 2026, the Tec-Chi content understanding model ranked second overall in the SuperCLUE overseas marketing video understanding list with 86.43 points. The company says it served more than 100,000 brands going global in 2025, covering more than 200 countries and regions, and its Tec-Chi model and Navos marketing multi-agent platform are co-evolving. QbitAI also reported that Tec-Do began cross-border marketing services in 2017, has managed more than 400 million advertising strategies and 14 million SPUs, and in the first three quarters of 2025 generated $116 million in AI marketing solution revenue, up 68.5 percent year on year and 89.5 percent of total revenue, with a net revenue retention rate of 132.5 percent. In July 2026, it became one of ChatGPT Ads' first official technology partners.