MIRAGE fixes Euclidean voids in flow-matching sequential recommendation
_reachsumit · x · 2026-07-28
This paper studies a failure mode in flow-matching sequential recommendation: straight paths in Euclidean embedding space can pass through regions with little valid item semantics, which the authors call the Euclidean void.
- They propose MIRAGE (Manifold-Informed Rectification for Accelerated Generation of Embeddings), which reshapes embedding geometry while keeping the original probability path unchanged.
- The method uses an item co-occurrence graph as a proxy for the semantic manifold during training only.
- That lets the model keep one-step inference while improving trajectory grounding in valid item support.
- Experiments on four real-world datasets show consistent gains over state-of-the-art baselines.
More from Research
- JIT-Agent: Improving LLMs via Just-in-Time Harness Evolution — NationalUniversityofSingapore · 2026-08-27
- D³-MOPD: Dynamic Scheduling for Multi-Teacher Distillation — Zechen Sun · 2026-08-27
- Frontier Models Complete Only ~20% of Scientific Workflows — apodex · 2026-08-27
- Agent-G²: Gaussian Guidance for Long-Horizon RL — ZhejiangUniversity · 2026-08-27
- AnTrap: Evaluating GUI Agent Robustness Against Anomalies — Guo Gan · 2026-08-27
- VGI-bench: Probing Visual Reasoning in Video Gen Models — Xuan He · 2026-08-27