Don't Waste It: Guiding Generative Recommenders with Structured Human Priors via Multi-Head Decoding

1 Introduction

• Industrial systems use accumulated human priors in post-hoc ranking adjustments, leaving them disconnected from end-to-end model learning.Examples include category-based diversity, interaction-value weighting, temporal discounting, and minority-user personalization.Source: §1 · p. 2
• A backbone-agnostic framework integrates human priors into generative training with lightweight, prior-conditioned adapter heads that disentangle user intent.Source: §1 · p. 2
• The framework generalizes multi-interest modeling across semantic, behavioral, temporal, and graph priors.Source: §1 · p. 2
• A hierarchical composition strategy models complex interactions across different prior types.It provides a flexible inductive bias for compositional representations.Source: §1 · p. 2
• The method improves accuracy, diversity, personalization, and user interest discovery across three large-scale datasets.Source: §1 · p. 2

2.1 Generative Recommenders

• Generative recommenders such as HSTU frame recommendation as sequential transduction, while HLLM uses separate item and user language models.A prevailing single-state representation can struggle with heterogeneous interests that conflict or evolve.Source: §2.1 · p. 3

2.2 Modeling Multi-Faceted User Intent

• Unsupervised multi-interest and disentanglement methods can conflate temporal or co-engagement factors, rely on costly or unstable discovery, and lack explicit semantic meaning.This can limit controllability for business objectives.Source: §2.2 · p. 3

2.3 Integration of Human Priors and Structure

• Prior recommender approaches use fixed taxonomies or post-hoc heuristics, which do not readily support diverse orthogonal priors and remain decoupled from core learning.Source: §2.3 · p. 3
• The proposed encode-then-project paradigm uses lightweight prior-conditioned heads to integrate diverse priors, bypass history clustering, and produce interpretable, controllable representations.Source: §2.3 · p. 4

3.1 Problem Formulation

• A sequential encoder maps the interaction history to a user state used to predict the next τ items; a single state may be suboptimal for diverse, evolving interests.Source: §3.1 · p. 4

3.2 Incorporating Human Priors via Conditioned Query Heads

• Lightweight prior-conditioned adapters project the shared user state into specialized queries, with compatibility masking restricting each head to items aligned with its prior group.The backbone processes the full interaction history; the heads specialize through their training signals.Source: §3.2 · pp. 4–5

3.3 Hierarchical Composition of Priors

• A sequential adapter tree learns shared upper-level representations before refining prior combinations, providing a coarse-to-fine inductive bias for sparse combinations.The design draws on hierarchical shrinkage to pull group-level estimates toward a common mean and help prevent overfitting rare combinations.Source: §3.3 · pp. 5–6

3.4 Training Objective

• The unified objective balances heads by prior-combination frequency and discounts farther-future losses; it supports next-token or contrastive losses, with contrastive negatives sampled within-group.Heads with no compatible positives in a batch are excluded from that batch’s loss.Source: §3.4 · p. 6

3.5 Inference and Score Fusion

• At inference, the system scores items using eligible heads and takes the maximum head score, allowing the most relevant specialist to dominate.The authors describe maximum fusion as computationally simple and interpretable.Source: §3.5 · pp. 6–7

3.6 Implementation Details

• Adapters add O(K · d^2) parameters and same-depth queries run in parallel; cached compatibility sets and shared-memory and on-demand data loading support scalable execution.The data pipeline stores valid sample starting indices rather than pre-generating rolling windows.Source: §3.6 · pp. 7–8

4.1.1 Pixel8M (Video) / 4.1.2 MerRec (E-commerce)

• Pixel8M tests semantic-category priors for multimodal video; MerRec uses six event types, including sparse monetization-related events, as priors.Pixel8M's 111 tags were consolidated into eight categories, while MerRec's offer and purchase events each occur less than 1% of the time.Source: §4.1.1 · p. 8, §4.1.2 · p. 8

4.1.3 EB-NeRD (News) / 4.1.4 General

• EB-NeRD uses Leiden clusters of an item co-engagement graph as graph priors, while temporal priors divide the forecast horizon into segments with dedicated heads.EB-NeRD is a news dataset with high-quality text; temporal priors apply across all datasets.Source: §4.1.3 · p. 9, §4.1.4 · p. 9

4.2 Experiment Setup

• The experiments integrate priors with HSTU and HLLM and evaluate recommendation accuracy using Recall@K and NDCG@K.The setup also compares against multi-interest baselines including ComiRec, REMI, and DualVAE.Source: §4.2 · p. 9

4.3 Main Results

• Human priors consistently improve Recall and NDCG over their backbone models, and combining priors can add further gains.Table 2 cautions that HLLM and HSTU results are not directly comparable because their context lengths differ; other baselines should be compared only with HSTU.Source: §4.3 · p. 9, Table 2 · p. 10
• ComiRec and REMI underperform their prior-injected HSTU counterparts; the framework scales to 88 heads, with one HSTU head using 0.14% of model parameters.The graph-prior implementation uses practical algorithmic clustering, and the authors report benefits despite noisy or approximated priors.Source: §4.3 · pp. 9–10

4.4.1 A Better Accuracy-Diversity Trade-off

• Item-prior variants achieve higher NDCG while maintaining higher entropy, indicating a better accuracy-diversity balance than variants without item priors.Entropy is measured over eight binary item categories on Pixel8M; REMI only marginally improves diversity over HSTU.Source: §4.4.1 · p. 11, Figure 2 · p. 11

4.4.2 User Interest Exploration

• +15.76% relative boost for NDCG@10 with HLLM LT/ST1 plus Item Prior on New Interest users, versus +3.58% with LT/ST1 alone.Across prior configurations, relative improvements are consistently higher on the New Interest subset than on All Users.Source: §4.4.2 · pp. 11–12, Table 3 · p. 12

4.4.3 Personalization

• User Prior yields bigger recommendation-quality improvements for smaller user groups, making quality more balanced across groups.The method assigns a separate adapter head to each of nine user groups clustered from a user co-engagement graph.Source: §4.4.3 · pp. 12–13, Figure 3 · p. 12

4.5 Scalability

• With priors, HSTU continues gaining from longer contexts and larger sizes, while the size 4 and 1b base model struggles beyond 20 items; gains gradually plateau.The comparison uses a fixed dataset, where increasing context length reduces the number of available training windows.Source: Figure 4 · p. 13, §4.5 · p. 13

4.6 Representation Space Visualization

• The Pixel8M UMAP compares No Prior, Item Prior, and REMI for one user's 50 past views, 10 targets, and top 50 recommendations, coloring outputs by head or interest.The UMAP is fit on history and target embeddings; learned item embeddings make the topology differ by model.Source: §4.6 · pp. 13–14, Figure 5 · p. 13
• No Prior recommendations cluster and miss diverse targets; Item Prior heads cover the target spectrum, whereas REMI's second interest only partly covers the main target cluster.REMI is shown with two interests, its best-performing setting in this visualization.Source: §4.6 · p. 14

4.7.1 How to Structure the Priors?

• Hierarchical composition outperforms the baseline and other strategies across five HSTU sizes; independent multiplicative heads can fragment data and weaken generalization for rare combinations.The comparison uses Item and Temporal priors on Pixel8M; the authors report similar trends with HLLM.Source: Figure 6 · p. 14, §4.7.1 · p. 14

4.7.2 Why it works: priors or just more heads?

• Maximum fusion performs better for Item Prior on NDCG/recall and diversity, whereas averaging performs better for Random, All, or Temporal Prior.The authors suggest heads under Item Prior assess candidates using different criteria, while the other heads assess more similar criteria.Source: §4.7.2 · p. 15
• Adding heads alone is insufficient: Random underperforms the baseline, while All improves modestly but trails the human-prior variants LT/ST8 or Item.The baseline in Table 4 is LT/ST1.Source: Table 4 · p. 15, §4.7.2 · p. 15

4.7.3 Long-Term vs. Short-Term Interests

• LT/ST variants beat the baseline at τ = 4, gain more at τ = 8, and incur only a small τ = 1 drop that disappears with more segments and Item Prior.Adding Item Prior also consistently benefits performance over LT/ST alone.Source: §4.7.3 · p. 15
• More LT/ST segments improve results through seg = 2, with diminishing gains thereafter; the benefit remains significant at 1b despite a slight dip from size4.The experiment trains on τ = 8 using only LT/ST prior, across HSTU model sizes.Source: Figure 7 · p. 16, §4.7.3 · p. 16

4.7.4 Training Objective

• Removing in-group negative sampling drops R@10 from 2.000 to 1.642, and removing frequency balancing also sharply reduces Recall and NDCG.The ablation reports Recall (R@10), NDCG (N@10), and Entropy (H@10).Source: Table 7 · p. 17, §4.7.4 · p. 17
• Recall and NDCG peak at γ = 0.7, while entropy first falls from γ = 1.0 to 0.9 then rises alongside accuracy through that peak.The reported optimum γ = 0.7 differs from the 0.9–0.995 range typically used in reinforcement learning.Source: Table 7 · p. 17, §4.7.4 · p. 17

• The framework integrates post-hoc business-rule priors into generative recommenders' learning, with prior-conditioned adapter heads and hierarchical composition for complex intent.The authors frame accumulated market-validated expertise as an advantage for recommendation systems.Source: §5 · p. 17
• Across experiments, the adapter heads improve accuracy, diversity, novelty, and personalization, while the backbone-agnostic design supports flexible modeling of complex user intent.Source: §5 · p. 17
• The work uses pre-defined priors; future work includes formal criteria for good-prior curation and more dynamic, context-aware fusion.The authors mention human-in-the-loop curation or automated discovery from data as possible directions.Source: §5 · p. 17

• Table 8 reports dataset statistics after preprocessing, with filtering criteria varying slightly by backbone.Source: Table 8 · p. 21
• For Pixel8M, HSTU filters users below 50 interactions and uses T = 50, τ = 8; HLLM filters below 20 and uses T = 10, τ = 4.The shorter HLLM context and forecast horizon reflect its higher computation cost.Source: §A.1 · p. 21
• The Pixel8M Miscellaneous tag covers over 5% of items and is represented by setting all eight binary features to True.The authors describe this as demonstrating robustness to noise in the human prior.Source: §A.1 · p. 21
• MerRec uses users with over 2,000 interactions and HSTU T = 400; EB-NeRD uses 44,968 users above 512 interactions, with HSTU T = 50, τ = 8 and HLLM T = 24, τ = 4.MerRec HSTU predicts one item ahead; EB-NeRD settings are reported separately for each backbone.Source: §A.1 · p. 21
• The co-engagement graph caps each highly active user's contribution at their last 1,000 interactions; Leiden clustering optimizes modularity, with small clusters merged.The resolution parameter is tuned to the desired number of clusters.Source: §A.1 · p. 21

A.2 Model Architectures and Hyperparameters

• A smaller γ may improve accuracy at the cost of lower diversity; both HSTU and HLLM use γ = 0.99.Table 7 is cited for the observed trade-off.Source: §A.2 · p. 21
• HLLM uses Qwen2-VL-2B-Instruct and Qwen2.5-1.5B for Pixel8M, and TinyLlama variants for text-only MerRec and EB-NeRD.Pixel8M uses the first model as Item LLM and the second as User LLM; the text-only datasets use TinyLlama-1.1B-Chat-v1.0 and TinyLlama_v1.1, respectively.Source: §A.2 · pp. 21–22
• HSTU configurations vary across five model sizes, with results reported for size 3 unless noted.Product IDs are remapped using items still interacted with after filtering.Source: §A.2 · p. 22, Table 9 · p. 22
• ComiRec and REMI use self-attention, with the best performance selected across tested interest counts; REMI also tunes β and sets λ to 100.The tested interest counts are {1, 2, 4, 8, 16, 32}, and REMI β values are {0.1, 1, 4, 10}.Source: §A.2 · p. 22
• DualVAE is reimplemented for window-wise sequential recommendation, tuning aspect count, dropout, γ, and β with a latent dimension of 32.Aspect counts are {4, 5, 10}; dropout is {0.1, 0.15, 0.2}, and γ and β are each searched over {1e −3, 1e −2, 1e −1, 1}.Source: §A.2 · p. 22

B Case Studies

• The history lacks “Informational & Educational” videos but suggests latent military and geopolitics interest through WWII, nuclear-waste-game, and atomic-egg-explosion clips.Video descriptions are omitted from Figure 8 due to space constraints.Source: B Case Studies · p. 22
• Even when filtered to “Informational & Educational,” baseline HLLM predictions are relatively generic, whereas the specialized head captures the latent interest.The target items include content about the Soviet Union and the Second Sino-Japanese War.Source: B Case Studies · p. 22
• The proposed model's long-term “Informational & Educational” head predicts Soviet Union and Russian territorial-waters content, including the user's Soviet Union target interest.The comparison uses Pixel8M HLLM predictions against the baseline, with history in the top two figure rows, targets in the third, baseline in the fourth, and model predictions in the last row.Source: B Case Studies · p. 22, Figure 8 · p. 23