Killing Two Birds with One Stone: Unifying Retrieval and Ranking with a Single Generative Recommendation Model

• Existing generative approaches often handle one stage or train each stage separately, while prior joint methods focus on data or loss transfer and leave bias and information loss unresolved.The paper frames retrieval and ranking as interdependent and identifies differing loss magnitudes and convergence rates as a challenge to simultaneous improvement.Source: §1 · p. 2
• UniGRF converts retrieval and ranking into sequence-generation tasks in one model, distinguishing their outputs by position.The framework is described as model-agnostic and plug-and-play, with information sharing between stages and no additional cost.Source: §1 · p. 2
• A ranking-driven enhancer uses ranking precision to generate high-quality retrieval samples, while a gradient-guided adaptive weighter balances stage losses.The enhancer forms a parameter-sharing loop; the weighter monitors optimization rates and adjusts stage weights.Source: §1 · p. 2
• Experiments report that UniGRF outperforms state-of-the-art baselines across various benchmark datasets.Ablations and further analysis are reported to support its scalability, multi-stage collaboration, and synchronized optimization.Source: §1 · p. 2

2.1 Retrieval Models for Recommendation

• Retrieval work includes sequential models such as GRU4Rec, BERT4Rec [43], and SASRec [23], LLM-enhanced representations, and generative approaches such as HSTU [67] and HLLM [3].HSTU still trains retrieval and ranking separately.Source: §2.1 · pp. 2–3

2.2 Ranking Models for Recommendation

• Ranking work spans feature-interaction models such as DCN [56] and DeepFM [10], sequential models including DIN [75] and DIEN [74], and LLM-based or generative approaches such as BAHE [9] and HSTU [67].The HSTU ranking adaptation appends candidate items to user sequences and adds MLP and prediction heads.Source: §2.2 · p. 3

2.3 Cascade Framework for Retrievers and Rankers

• Cascade methods connect stage objectives, incorporate unexposed data [41], jointly optimize stage models, distill knowledge [20], or estimate selection bias [73].These approaches use traditional recommendation models and do not address efficient collaboration and synchronized optimization in generative recommendation.Source: §2.3 · p. 3

• A user history records chronological item interactions and click status, retaining the most recent n interactions or padding shorter histories to length n.The interaction indicator is 1 for a click and 0 for no click.Source: §3 · p. 3
• The generative recommendation objective is to predict the next item the user will interact with from that history.Source: §3 · p. 3
• UniGRF performs retrieval and ranking simultaneously with one generative model: it selects candidates from the item pool, scores them, and recommends the next item.Source: §3 · p. 3

• UniGRF reformulates retrieval and ranking as sequence-generation tasks within one model-agnostic framework to support information transfer and scalability.Source: §4 · p. 3
• A ranking-driven enhancer generates accurate, high-quality samples with minimal computational overhead to create a mutually reinforcing loop between stages.Source: §4 · p. 3
• A gradient-guided adaptive weighter is designed to synchronize performance improvement across retrieval and ranking.Source: §4 · p. 3

4.2 Ranking-Driven Enhancer

• UniGRF unifies retrieval and ranking within one generative recommendation model.The overview also identifies a ranking-driven enhancer and a gradient-guided adaptive weighter.Source: Figure 1 · p. 4
• A ranking-driven enhancer is included to support collaboration between the stages.Source: Figure 1 · p. 4
• A gradient-guided adaptive weighter is included for synchronized optimization.Source: Figure 1 · p. 4

4.1 A Unified Generative Recommendation Framework

• UniGRF is plug-and-play with autoregressive generative architectures and halves time and space complexity.It combines retrieval and ranking in one generative recommendation framework.Source: §4.1 · p. 4

4.1.1 Constraint on the Retrieval Stage in UniGRF.

• Retrieval predicts the next item from prior interactions, adds the most similar item from the pool to the candidate set, and optimizes with sampled softmax.The loss pulls the predicted item embedding toward the actual next item and away from randomly sampled negatives.Source: §4.1.1 · p. 4

4.1.2 Constraint on the Ranking Stage in UniGRF.

• Ranking estimates click probabilities for candidate items with a small neural network and compares scores with user feedback using binary cross-entropy.The shared model parameters let ranking's more precise user-interest representations enhance retrieval, while full-pool negative sampling is described as reducing ranking data bias.Source: §4.1.2 · pp. 4–5

4.2.1 Hard-to-Detect Disliked Items Generator.

• Randomly sampled negatives can be easy to distinguish from positives, so the generator selects challenging dislikes using relative retrieval and ranking scores.It repeatedly trains on the selected items until they are distinguished effectively and removed from the hard-negative set.Source: §4.2.1 · p. 5

4.2.2 Potential Favorite Items Generator.

• Items in the negative sample set that receive ranking scores above threshold α are flagged as potential favorites and relabeled positive in the next iteration.This addresses the possibility that randomly sampled negatives are actually liked by the user.Source: §4.2.2 · p. 5

4.2 Ranking-Driven Enhancer

• Ranking precision supplies retrieval with more informative training data through hard-to-detect dislikes and potential favorites, while shared parameters support bidirectional enhancement.The framework maintains the original data distribution and adds minimal computational overhead.Source: §4.2.2 · p. 5

4.3 Gradient-Guided Adaptive Weighter

• Different convergence speeds in the unified framework complicate simultaneous optimization, motivating weights that adapt to each stage’s learning speed.The mechanism monitors gradient update rates over time to align convergence speeds.Source: §4.3 · pp. 5–6
• The method qualitatively estimates convergence speed from the magnitude of loss decrease, using relative optimization magnitudes for retrieval and ranking.A larger convergence-rate value indicates slower convergence and calls for a larger task weight.Source: §4.3 · p. 6
• The adaptively weighted retrieval and ranking losses form the final optimization loss, which the paper says enables synchronous optimization.Temperature adjusts the difference between stage weights, while λa and λb scale the losses to similar magnitudes.Source: §4.3 · p. 6

5.1.1 Datasets.

• Experiments use MovieLens-1M, MovieLens-20M, and Amazon-Books, with Table 1 identified as reporting dataset statistics.The datasets are described as varying in size and derived from real-world user interactions.Source: §5.1.1 · p. 6, Table 1 · p. 6

5.1.2 Compared Methods.

• Comparisons include classic retrieval and ranking methods, generative models trained independently for single tasks, and cascade frameworks instantiated with HSTU.The listed cascade baselines are RankFlow-HSTU and CoRR-HSTU.Source: §5.1.2 · pp. 6–7

5.1.3 Evaluation Protocols.

• All methods use the same chronological train-validation-test partition, and retrieval is evaluated against the entire item set using NDCG@K, HR@K, and MRR at K=10 and 50, while ranking uses AUC.The last interaction is test, the second-to-last is validation, and earlier interactions are training data.Source: §5.1.3 · p. 7

5.1.4 Parameter Settings.

• For each dataset, baselines share hyperparameter settings and a learning rate of 1e−3, while negative-sample counts and training epochs vary by task and dataset.Transformer depth also varies by dataset; all models are trained using eight computation cards.Source: §5.1.4 · p. 7

5.2 Overall Performance

• Generative HSTU and Llama outperform non-generative methods when retrieval and ranking use different models.The authors suggest richer feature representations and deeper user-interest relationships may explain this pattern.Source: §5.2 · p. 7
• UniGRF improves both stages, with a larger boost in ranking, which directly affects recommendation quality.The authors note that retrieval generates a candidate set and is less sensitive to metric changes.Source: §5.2 · p. 7
• UniGRF consistently outperforms all baselines in retrieval and ranking with both HSTU and Llama.The comparison spans three datasets.Source: §5.2 · p. 7, Table 2 · p. 8, Table 3 · p. 8

5.3 Ablation Study

• All tested ablation and fine-tuning variants perform worse than UniGRF-HSTU on MovieLens-1M.The ablations remove the enhancer, the weighter, or both; the fine-tuning variants transfer parameters between separately trained stages.Source: §5.3 · p. 8, Table 4 · p. 8
• Removing both modules performs worst and fails to surpass HSTU; removing the enhancer degrades retrieval, while removing the weighter degrades ranking.Fine-tuned variants improve on HSTU but do not surpass UniGRF-HSTU, even with more training epochs.Source: §5.3 · p. 8

5.4 Impact of Hyper-parameter

• On MovieLens-20M, retrieval and ranking performance continuously rises as hard-negative count m increases from 0 to 20.The total negative-sample count is fixed at 256 for this dataset.Source: §5.4 · p. 9, Figure 4 · p. 9
• On MovieLens-1M, both stages peak around m=5, while larger hard-negative counts may overfit and reduce sample diversity.The total negative-sample count is fixed at 128; the authors recommend balancing performance gains against overfitting on small datasets.Source: §5.4 · p. 9

5.5.1 Analysis of the Impact of Optimization on Unified Framework Performance.

• With simple loss addition on MovieLens-20M, ranking performance peaks after several epochs and rapidly declines while retrieval continues rising.The authors attribute this divergence to retrieval's harder negative-sampling task receiving disproportionate optimization attention.Source: §5.5.1 · p. 9
• Adaptive two-stage weighting keeps ranking performance improving and makes retrieval improvement smoother by adjusting loss weights to training rates and convergence speeds.The comparison uses performance curves over epochs on MovieLens-20M.Source: §5.5.1 · p. 9

5.5.2 Analysis on Scaling Law.

• On MovieLens-20M, increasing Transformer layers lowers loss and improves retrieval and ranking performance.The tested HSTU architectures use 2, 4, 8, 16, 24, or 32 layers; scalability is demonstrated on this large public dataset, with industrial datasets left for future study.Source: §5.5.2 · pp. 9–10, Figure 4 · p. 9

6 Conclusion

• UniGRF combines retrieval and ranking in one generative model to address information loss in multi-stage recommendation.Its ranking-driven enhancer and gradient-guided adaptive weighter support collaboration and synchronized optimization between stages.Source: §6 · p. 10
• Experiments on benchmark datasets show that UniGRF outperforms state-of-the-art baselines.The authors describe the framework as scalable and as amplifying generative models' strengths.Source: §6 · p. 10