IJCAI-ECAI 2026 runs from 15 to 21 August in Bremen, Germany. As the conference opens, Ant Group’s four accepted papers point to a thread distinct from the parameter race: a systematic search for adaptation to dynamic environments. Model accuracy keeps setting records in the lab, but once deployed in the real world, performance decay is almost routine, because the real world is never static.
Why static models fail Ant
Ant’s business pushes ‘not static’ to an extreme: more than 1bn users, over 8,000 services, and payment peaks versus 3am troughs that can differ more than a hundredfold. Risk models must answer shifts in black-market tactics within hours, and global deployment has to fit each country’s different financial infrastructure. In such an environment, a ‘train once, deploy forever’ static model falls into the trap of carving a mark on a boat to find a lost sword. The real question the four papers answer is how to move AI from a static solver to a dynamic self-adapter.
Clustering that follows the data
MSRGC-Net, proposed with Chongqing University of Posts and Telecommunications, clusters time series without training. It uses several untrained echo-state networks to extract local fluctuation and long-term trend, then builds ‘granular balls’ anchored by local density: dense regions form small, many balls, sparse regions form large, few balls, compressing N points to M while suppressing noise. A consensus strategy fuses multi-scale graphs for robust results without a preset cluster count. On five multivariate benchmarks across 15 metrics it took 12 firsts and 2 seconds, with near-linear complexity. In Ant’s business this means the system auto-adjusts clustering granularity to traffic, fine-grained at peaks, coarse at troughs.
Reinforcement learning that mixes itself
ROAD, with Shanghai Jiao Tong University, attacks the offline-to-online ‘distribution shift’ problem, where a model trained on historical data forgets on launch. Existing fixes fix the mix at 50-50 or use heuristic rules, but needs change by stage, more historical data early to stabilise, more live data later to explore. ROAD turns the mix ratio from a preset parameter into a dynamic decision variable, so the model stops forgetting and keeps improving. Across the four papers the message is consistent: intelligence is not a fixed solver but a system that senses change and rewrites its own strategy.
Read the original report (LeiPhone)
Translated and adapted from LeiPhone (leiphone.com).