USTC and Alibaba's CEDAR cuts demand forecast error 57% by decoupling decisions from external shocks
量子位 · wechat · 2026-09-17
CEDAR (KDD'26, USTC + Alibaba) models "decision-conditioned simulation": an Action-Interleaved Transformer treats states and actions as first-class tokens, while a residual module uses LLM-extracted social-media signals with cross-attention to capture external shocks. On 32M product trajectories from 1688 it cuts forecast error 57%+ vs top baselines; online A/B tests lifted merchant LTV 13% and ROI 15%.
More from Research
- François Fleuret: AI math is to math what numerical integration is to ODEs — francoisfleuret · 2026-09-17
- White paper reframes OpenAI-HuggingFace incident via semantic field mechanics — LopsidedLevel9009 · 2026-09-17
- UIUC Launches Center for Orchestrated Agentic Biology to Study Brain Aging with Multi-Agent AI — dotey · 2026-09-17
- ActionPiece rethinks action tokenization for VLA models, hits 94.8% on LIBERO — DeepCybo · 2026-09-17
- Berkeley's ComPO: zeroth-order preference alignment to mitigate likelihood displacement — Berkeley · 2026-09-17
- SpectralShift extends Gated DeltaNet context windows via spectral reparameterization — RUC-AIBOX · 2026-09-17