EvoReason framework boosts latent reasoning for generative recommender systems

Generative recommender systems increasingly use reasoning-enhanced inference, but this often means deploying explicit chain-of-thought (CoT) traces that are verbose and unstable. Latent reasoning compresses CoT into continuous representations for low-latency serving, yet existing methods blindly distill raw textual CoT, which is suboptimal because recommendation reasoning traces contain redundant expressions and inconsistent paths. Enter EvoReason, a new framework from Zhuang Zhuang and six co-authors at arXiv (cs.IR:2607.29010).

EvoReason first mines reusable reasoning primitives from high-quality agentic recommendation trajectories. These primitives act as pseudo-tools that structure how the teacher produces CoT supervision, dramatically reducing redundancy and improving consistency. Then, it introduces a self-evolving on-policy distillation mechanism: during optimization, the teacher's primitive-guided reasoning is adjusted based on the student's current latent reasoning outcomes. This closed-loop co-evolution ensures the supervision stays aligned with the student's representational space, leading to progressively better reasoning transfer. The result is a more efficient and accurate latent reasoning approach for latency-sensitive generative recommender systems, addressing a key bottleneck in deploying reasoning-aware AI at scale.