OpenOneRec Technical Report (2512.24762v1)

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.

• The paper introduces OpenOneRec, a unified framework that integrates collaborative filtering with general semantic reasoning using a three-phase training pipeline.
• It presents a modality-aligned tokenization strategy via hierarchical quantization to bridge item data with language models.
• Experimental results demonstrate up to 30% Recall@10 improvement and superior cross-domain transfer, establishing state-of-the-art performance.

OpenOneRec: Unifying Recommendation and General Intelligence via End-to-End Instruction-Following Foundation Models

Motivation and Framework Overview

OpenOneRec addresses the limitations of existing recommendation systems, most notably their inability to scale beyond siloed datasets and thus failing to capture broader reasoning and instruction-following capabilities that characterize modern LLMs. Previous recommendation paradigms, even those employing generative modeling, remain domain experts—lacking extensive world knowledge and robust cross-domain transferability. OpenOneRec's framework, built upon the Qwen3 backbone, strives for a tight integration of collaborative filtering with general semantics using a three-phase strategy: scalable mixed-domain pre-training, hybrid post-training (including SFT, general distillation, and RL), and holistic evaluation via RecIF-Bench.

Figure 1: The OneRec-Foundation pipeline incorporates Itemic-Text Alignment, co-pretraining, SFT, distillation, and Rec-RL, comprehensively aligning collaborative signals with general knowledge and enabling broad downstream applicability.

Modality Alignment and Generative Task Formulation

The principal technical advance is the modality-aligned tokenization that bridges the discrete item domain (IDs and metadata) and the linguistic/tokensequence regime of LLMs. Itemic tokens are constructed via hierarchical quantization (RQ-Kmeans), compressing metadata embeddings to hierarchical discrete codes with prefix-sharing properties, facilitating generative transfer learning. Recommendation is cast as a unified autoregressive next-token generation task, with user history and instructions forming the prompt, and the target sequence encompassing next-item tokens or explanatory text.

RecIF-Bench: Comprehensive Instruction-Following Benchmark

RecIF-Bench is introduced as the first benchmark to fully evaluate recommendation foundation models across multifaceted capabilities. It comprises 120M interactions from 200K users in short video, e-commerce, and online advertising domains, supporting long-context sequential modeling, multimodal metadata, and interleaved text–item input formats. Task taxonomy expands across four hierarchical layers:

• L0 Semantic Alignment: Itemic token → textual metadata mapping (tests alignment quality).
• L1 Fundamental Recommendation: Next-item prediction over multiple domains, binary label prediction for user-item pairs.
• L2 Instruction Following: Contextual recommendation conditioned on natural language queries or explicit behavior labels.
• L3 Reasoning: Natural language explanation generation for recommendations.

Evaluation employs both classic retrieval metrics (Recall@K, Pass@K) and LLM-as-Judge protocols for generation tasks, enabling robust and repeatable comparison.

Figure 2: RecIF-Bench's task taxonomy spans alignment, prediction, instruction-following, and reasoning, specifying input context, instruction, and outputs across 8 diverse task types.

Figure 3: RecIF-Bench data distribution reveals highly skewed popularity and varied history lengths, demonstrating the need for long-context and cross-domain modeling.

Pre-Training Strategies and Scaling Law Analysis

Pre-training mixes recommendation-domain interaction/meta-data with general-domain reasoning and coding corpora. Itemic-Text Alignment initializes token embeddings within the linguistic latent space, followed by full-parameter co-pretraining to encode collaborative signals while controlling for catastrophic forgetting.

Empirical scaling law analysis diverges from Chinchilla scaling observed in open-text LLMs: recommendation tasks exhibit a more aggressively data-hungry compute-optimal configuration ($N_{\text{opt}} \propto C^{0.44}$, $D_{\text{opt}} \propto C^{0.56}$), as deduced from convex-hull fits of training loss envelopes. The irreducible entropy floor for recommendation is markedly lower than for general text, indicating high determinism given structured features and collaborative signals.

Figure 4: Scaling law plots confirm predictable capability scaling, but with a data-intensive regime for recommendation compared to standard LLM scaling.

Post-Training: SFT, On-Policy Distillation, and RL Optimization

Post-training is structured into three stages:

1. Multi-task SFT: Blends recommendation and reasoning instruction datasets, yielding models that recover instruction-following and cross-task transfer. This tunes both general and domain-specific capacities.
2. On-Policy Distillation: Leveraging policy gradients and per-token Reverse KL divergence to align student output distributions with the original Qwen3 teachers, while penalizing spurious itemic token generations in general text contexts. Robust trajectory truncation and exploration strategies mitigate vocabulary drift.
3. Recommendation-Oriented Reinforcement Learning: Employs Group Relative Policy Optimization (GRPO), directly optimizing ranking metrics with non-differentiable “hit” reward functions within grouped response sampling, penalized by KL regularization to preserve general reasoning skills.

   Figure 5: OneRec series post-training pipeline, showing sequential application of SFT, distillation, and reinforcement learning stages.

   

   Figure 6: On-policy distillation pipeline utilizes policy gradient methods and reverse KL divergence to recover general reasoning and ensure robust general-domain alignment.

Experimental Results: State-of-the-Art and Transfer Superiority

Across RecIF-Bench tasks, OneRec-Foundation surpasses all discriminative and generative baselines (including LC-Rec and TIGER), with both scaling in parameters and training data producing monotonic performance lifts. Notable is the strong cross-domain generalization: on 10 Amazon datasets, the model achieves an average 26.8% improvement in Recall@10 over the best known baselines, with relative improvements exceeding 30% in several verticals. Text-Augmented Itemic Tokens, combining hierarchical codes and keywords, demonstrate minimized collision rates and state-of-the-art transfer—balancing collaborative filtering capacity with semantic expressivity.

Figure 7: Holistic performance overview across RecIF-Bench and Amazon benchmarks, with OneRec-Foundation achieving best-in-class metrics and robust cross-domain transfer.

Figure 8: Joint multi-domain training boosts Recall@10 in Amazon benchmarks for OneRec-Foundation but degrades performance for TIGER, demonstrating OneRec's superior knowledge fusion and generalization.

Few-shot learning analyses show that recommendation foundation models maintain a much larger fraction of their full-data performance than baselines, highlighting the sample efficiency gains from large-scale pre-training.

Ablation studies demonstrate the necessity of modality alignment and the contribution of each post-training phase: SFT provides initial recovery, on-policy distillation restores general intelligence, and Rec-RL directly optimizes recommendation metrics without sacrificing reasoning.

Theoretical and Practical Implications

Practically, OpenOneRec emerges as a reproducible, instruction-following foundation model for recommendation that can be reliably transferred across unseen domains and tasks, vastly improving sample efficiency and resilience under data scarcity. Theoretically, the observed scaling regime, modality alignment strategies, and asymmetric trade-off between recommendation and general reasoning capacities raise new questions on optimal data mixing, vocabulary transfer, and continual domain adaptation. The results emphasize the necessity for further explorations into codebook indexing, curriculum mixing, and test-time scaling strategies to exploit CoT reasoning for recommendation utility.

Conclusion

OpenOneRec pioneers the unification of large-scale generative modeling and recommendation, establishing both the methods and evaluation protocols necessary for foundation modeling in this domain. By releasing comprehensive benchmarks, open data, full-stack training code, and validated scaling analyses, the work provides a robust platform for advancing research in intelligent recommender systems. Challenges remain in further balancing general intelligence retention with item indexing transferability and optimizing Chain-of-Thought induction for robust reasoning improvements. Ongoing community-driven research informed by these findings will be vital to accelerate progress towards universally deployable, intelligent recommendation agents.