Skip to content
Book Open access

Group-Relative RL Post-Training for Sequential Recommendation

Nov 2026 · Proceedings of the 35th ACM International Conference on Information and Knowledge Management · 1 citation · 45 references

Abstract

Sequential recommender systems are typically trained with next-item cross-entropy (CE), yet evaluated by top-N ranking metrics that reward relative item ordering in the recommendation slate. This mismatch remains unaddressed for strong collaborative sequential models. Recent group-relative reinforcement learning has shown promise for LLM alignment and LLM-based recommenders, but whether a critic-free post-training paradigm can reliably improve standard collaborative sequential recommenders remains unclear. We introduce GRFT (Group-Relative Fine-Tuning), a GRPO-style post-training framework that aligns a pretrained sequential recommender directly to slate-level ranking rewards. GRFT samples multiple top-K slates per user context via Plackett-Luce sampling without replacement, scores them with ranking rewards, and updates the model with a PPO-clipped group-relative objective anchored to the frozen supervised reference. We instantiate rankwise and listwise variants to support position-level and whole-slate credit assignment. Under leakage-resistant global temporal splits on four datasets, GRFT consistently improves a strong sequential backbone, achieving gains of up to 25.3% NDCG@20 and 19.5% Recall@20. The improvements remain complementary to stronger supervised initialization, showing that reward-aligned post-training adds value beyond better supervised training alone. GRFT thus establishes that GRPO-style, critic-free post-training can improve strong collaborative sequential recommenders.

Read PDF

We use cookies to run the site and, with your consent, for analytics and to show ads. See our Cookie Policy.