Tencent proposes a discrete diffusion language model for recommendation

_reachsumit · x · 2026-07-24

Tencent presents a discrete diffusion language model for recommendation that replaces standard autoregressive generation.

Instead of predicting items left-to-right, the model uses bidirectional denoising plus iterative refinement. The result is a recommendation formulation that treats generation more like recovering a full sequence from noise than like classic next-token prediction.

Original post →

More from Research

Research channel →