ECCV 2026 hosts first Agentic Commerce workshop: 64 teams compete on making AI understand ads

量子位 · wechat · 2026-09-10

ECCV 2026 featured MARS 2, the conference's only Agentic Commerce workshop led by a Chinese tech company (Tec-do/钛动科技), with speakers including Oxford's Yarin Gal and MIT's Paul Pu Liang. Its challenge offered a $100K prize pool and drew 64 teams with 1,060+ submissions.

The challenge is built on M-CAR, a dataset of 3,108 real ad videos across 30+ languages, split into 18,198 semantic segments, with three tracks: overall understanding (MAC), key-moment grounding (VTG), and marketing intent reasoning (MDC). Only open-weight models ≤14B (MAC/MDC) or ≤8B (VTG) were allowed. Results: MAC topped at 86.67, while VTG/MDC peaked around 62-63 — a 20+ point gap showing temporal grounding and business reasoning remain much harder. Team TheBoys (ByteDance background) took two championships.

A counterintuitive finding: adding a cross-modal "audio evidence timeline" boosted grounding accuracy by 16.7 points, while scaling from 4B to 8B parameters actually cost 0.2 points. Winning solutions used Proposer-Critic verification and an "evidence-chain engineering" approach. Benchmarks and code will be open-sourced.

The article also covers Tec-do's business: its Tec-Chi marketing multimodal models rank near the top on SuperCLUE marketing benchmarks, its Navos agent platform, and $116M AI marketing revenue (+68.5% YoY) in the first three quarters of 2025, plus the broader trend of tech companies becoming agenda-setters at top conferences.

Original post →

More from Companies & People

Companies & People channel →