OpenOneRec Technical Report

TL;DR

OpenOneRec presents RecIF-Bench, the first holistic benchmark for instruction-following in recommendation, and releases a full-stack training framework plus OneRec-Foundation models (1.7B and 8B). By introducing Itemic Tokens and a two-stage pre-training pipeline that blends recommendation-domain data with broad general-domain reasoning data, the approach achieves state-of-the-art results on RecIF-Bench and strong cross-domain transfer on the Amazon benchmark. The framework includes on-policy distillation and GRPO-based reinforcement learning to balance general intelligence with domain-specific recommendation quality, and reveals data-driven scaling laws that favor data growth over model size in recommendation. Despite demonstrated gains, the work notes limitations in tokenizer transferability, data mixing strategies, and limited chain-of-thought improvements, inviting further community collaboration and exploration of scalable, intelligent recommender systems.

Abstract

While the OneRec series has successfully unified the fragmented recommendation pipeline into an end-to-end generative framework, a significant gap remains between recommendation systems and general intelligence. Constrained by isolated data, they operate as domain specialists-proficient in pattern matching but lacking world knowledge, reasoning capabilities, and instruction following. This limitation is further compounded by the lack of a holistic benchmark to evaluate such integrated capabilities. To address this, our contributions are: 1) RecIF Bench & Open Data: We propose RecIF-Bench, a holistic benchmark covering 8 diverse tasks that thoroughly evaluate capabilities from fundamental prediction to complex reasoning. Concurrently, we release a massive training dataset comprising 96 million interactions from 160,000 users to facilitate reproducible research. 2) Framework & Scaling: To ensure full reproducibility, we open-source our comprehensive training pipeline, encompassing data processing, co-pretraining, and post-training. Leveraging this framework, we demonstrate that recommendation capabilities can scale predictably while mitigating catastrophic forgetting of general knowledge. 3) OneRec-Foundation: We release OneRec Foundation (1.7B and 8B), a family of models establishing new state-of-the-art (SOTA) results across all tasks in RecIF-Bench. Furthermore, when transferred to the Amazon benchmark, our models surpass the strongest baselines with an average 26.8% improvement in Recall@10 across 10 diverse datasets (Figure 1). This work marks a step towards building truly intelligent recommender systems. Nonetheless, realizing this vision presents significant technical and theoretical challenges, highlighting the need for broader research engagement in this promising direction.

Figures (8)

• Figure 1: Holistic Performance Overview.Left: Evaluation on RecIF-Bench and general LLM benchmarks. Our model achieves SOTA performance on recommendation tasks while effectively retaining general knowledge. "Best Baseline" denotes the highest performance achieved by existing methods for each specific task. Right: Amazon Benchmark results. Our model demonstrates exceptional cross-domain transferability, consistently surpassing leading baselines across 10 diverse datasets.
• Figure 2: The Overall Framework of OneRec-Foundation. (1) Pre-Training: Integrates collaborative signals with general semantics via Itemic-Text Alignment and mixed-domain Co-Pretraining. (2) Post-Training: Unlocks diverse downstream capabilities via SFT, and balances general reasoning with recommendation performance through alternating General Distillation and Rec-RL. (3) Evaluation: Comprehensively assesses holistic capabilities on RecIF-Bench and validates cross-domain transferability on Amazon datasets.
• Figure 3: Data distribution analysis of RecIF-Bench. (a) Item popularity distribution (log-log scale) across domains. (b-d) Distribution of user history lengths for Short Video, Ad, and Product domains, respectively.
• Figure 4: Task Taxonomy of RecIF-Bench. We organize 8 tasks across 4 capability layers, specifying the instruction, context, and target.
• Figure 5: Scaling laws on recommendation-domain data.Top-left: training loss vs. FLOPs for different model sizes, and the smooth convex envelope shown in grey. Top-right: log-log plot of training loss vs. model size under fixed token budgets. Bottom-left: compute-optimal model size $N_{\text{opt}} \propto C^{0.44}$ as a function of compute budget $C$. Bottom-right: compute-optimal token budget $D_{\text{opt}} \propto C^{0.56}$ as a function of $C$.