PaperlayerView exact revision on arXiv

Source-linked AI summary

When Diffusion Recommenders Stop Sampling: Preserving Generative Diversity

Louis Castres Saint Martin, Alain Rakotomamonjy, Patrick Gallinari

arXiv:2610.36954v1cs.IR

TL;DR

Sequential recommenders often produce deterministic rankings, and existing target-item diffusion recommenders can collapse to nearly identical samples. DARTS separates latent representation learning, diffusion training, and ranking; across three Amazon Reviews 2023 categories, it improves ranking metrics over several baseline families while preserving diversity.

  • Problem

    Sequential recommenders return a single ranking, while diffusion models offer sampling but existing target-item approaches can produce nearly identical rankings across draws.

    Source: §1 · pp. 1–2

  • Method

    DARTS freezes a pretrained latent space, trains a conditional denoiser with latent regression, and postpones ranking supervision to decoder fine-tuning.

    Source: §1 · p. 2

  • Results

    DARTS improves standard ranking metrics by 6.8–12.0% over discriminative, generative-retrieval, and diffusion baselines across three Amazon Reviews 2023 categories.

    Source: §1 · p. 2

  • Takeaways & Limitations

    Averaging multiple posterior samples further improves ranking metrics, turning sample diversity into additional accuracy gains.

    Source: §4.4 · p. 9, §5 · p. 9

    Leave-one-out evaluation observes only one next interaction, so ranking metrics cannot establish whether alternative samples are plausible futures.

    Source: §5 · pp. 9–10

Abstract

from arXiv · show

A user history can admit multiple plausible next interactions, yet conventional sequential recommenders produce a single deterministic ranking. While diffusion models offer a natural way to sample alternative continuations, we show that existing target-item diffusion recommenders often suffer from severe sampling collapse, producing nearly identical outputs across independent draws. We identify two ubiquitous design choices that promote this collapse: joint learning of the latent space and denoiser, and in-trajectory supervision via discrete losses. To address these failure modes, we introduce DARTS (Diverse and Accurate Recommendation Through Sampling), a three-stage latent diffusion framework that explicitly decouples representation learning, generative modeling, and discrete ranking by learning and freezing the latent space before diffusion training and postponing ranking supervision to the final decoding stage. We further characterize sampling diversity using complementary measures of variation before and after decoding, and use posterior-predictive aggregation to exploit this diversity at inference. Across three Amazon Reviews 2023 categories, DARTS preserves substantially greater sampling diversity while outperforming strong sequential, generative-retrieval, and diffusion baselines by 7-12% on standard ranking metrics. Aggregating multiple samples further improves ranking accuracy at a fixed slate size.

1 INTRODUCTION

Sequential recommenders often return deterministic rankings, while diffusion recommenders can sample alternatives but often produce nearly identical rankings. DARTS separates representation learning, diffusion training, and ranking, and reports stronger ranking performance alongside preserved diversity.

1 INTRODUCTION

  • Existing diffusion recommenders often yield nearly identical rankings across repeated draws, despite the potential to sample multiple continuations.The authors attribute this collapse to jointly learning item representations with the denoiser and applying catalogue losses during denoising.Source: §1 · pp. 1–2
  • DARTS freezes a pretrained latent space, trains its denoiser with regression alone, and applies ranking supervision only when fine-tuning the decoder.The paper also introduces diagnostics for diversity in latent and item space and reports that both sampling diversity and ranking performance improve against the tested baselines.Source: §1 · p. 2
  • Across three Amazon Reviews 2023 categories, DARTS improves standard ranking metrics by 6.8–12.0% over discriminative, generative-retrieval, and diffusion baselines, while multi-sample aggregation adds gains.The paper also reports that aggregating several posterior samples into a fixed-size slate further improves ranking accuracy.Source: §1 · p. 2

2 RELATED WORK

Prior sequential recommenders range from discriminative next-item models to generative retrieval, while diffusion recommenders model either interaction vectors or target-item representations. DARTS combines a fixed latent space with continuous sampling and final-stage catalogue decoding.

2 RELATED WORK

  • SASRec and BERT4Rec represent Transformer-based next-item approaches, while semantic-identifier models generate discrete item codes autoregressively; DARTS samples continuous next-item representations.The cited works include SASRec, BERT4Rec, and semantic-identifier approaches (Zheng et al., 2025).Source: §2 · p. 2
  • Diffusion recommenders either diffuse interaction vectors or target-item representations; target-item methods commonly learn item representations jointly with the denoiser or add item-level losses during denoising.DARTS instead freezes a pretrained latent space, uses latent regression for denoiser training, and applies ranking supervision to the final decoder.Source: §2 · pp. 2–3
  • Latent diffusion separates representation learning from generation, while ELF remains continuous until final discretization; DARTS brings both design ideas to recommendation.The paper tests whether a fixed latent space and a single final mapping to catalogue items preserve stochastic sampling.Source: §2 · p. 3

3 METHOD

DARTS separates sequential recommendation into representation learning, latent generation, and catalogue ranking. It learns a fixed latent space before conditional diffusion and fine-tunes the decoder on sampled latents for inference-time ranking.

3.1 OVERVIEW

  • DARTS trains an encoder and decoder, then a denoiser, then the decoder again, with each stage updating a different parameter set.The overview figure shows the masked history producing fixed prefix latents that condition reverse diffusion before the decoder ranks the catalogue.Source: §3.1 · p. 3, Figure 1 · p. 3

3.2 BACKGROUND: CONDITIONAL DENOISING DIFFUSION

  • Conditional diffusion adds Gaussian noise to a clean target over T steps and learns a reverse process that estimates the clean vector from a noisy vector, timestep, and history condition.The forward schedule approaches a standard normal at its final step; generation starts from Gaussian noise.Source: §3.2 · p. 4

3.3 ITEM REPRESENTATIONS AND MODEL BACKBONE

  • The model distinguishes catalogue items, item-representation vectors, and encoder latents; a bidirectional encoder and causal decoder connect these spaces.Hybrid item representations combine semantic and collaborative information, and the decoder outputs item representations for catalogue scoring.Source: §3.3 · pp. 4–5

3.4 STAGE 1: SEQUENCE AUTOENCODER PRETRAINING

  • Stage 1 pretrains the encoder and decoder with masked item modeling to reconstruct items at masked positions, always masking the final item.The final-item mask aligns pretraining with next-item prediction; the encoder and item representations are then frozen for later stages.Source: §3.4 · p. 5

3.5 STAGE 2: LATENT DIFFUSION

  • Stage 2 fits a conditional denoiser to the next-item latent in the fixed space, using the masked-prefix encoding as condition and the full-sequence final-slot encoding as target.The target is contextualized by the sequence encoder, and the stage uses latent regression without catalogue-related loss.Source: §3.5 · pp. 5–6

3.6 STAGE 3: DECODER FINE-TUNING

  • Stage 3 runs reverse diffusion, concatenates the sampled target latent with the prefix, and fine-tunes the decoder with a temperature-scaled BPR loss to rank the true item first.This adapts the decoder to inference inputs and addresses the shift from the masked final slot used in pretraining without changing the learned stochastic latent distribution.Source: §3.6 · p. 6

4 EXPERIMENTS

Across three Amazon Reviews 2023 categories, DARTS leads the evaluated baselines on ranking metrics while retaining variation across sampled recommendations. The experiments also examine which design choices affect sampling diversity and whether aggregating draws improves accuracy.

4.1 DATASETS

  • The evaluation uses MUSICAL INSTRUMENTS, INDUSTRIAL AND SCIENTIFIC, and VIDEO GAMES, retaining users and items with at least five interactions and each history’s 20 most recent items.Histories are ordered chronologically; evaluation uses leave-one-out Recall@K and NDCG@K for K ∈ {5, 10}.Source: §4.1 · p. 6, §4.2 · p. 7

4.2 OVERALL PERFORMANCE - ACCURACY

  • The experiments compare sequential recommenders, semantic-identifier generative retrieval models, and diffusion baselines, including models trained with DARTS’s hybrid item representations.Results for the first two baseline families are taken from prior work, except ActionPiece; the paper trains the +T5 comparisons itself.Source: §4.2 · p. 7
  • For recommendation performance across three Amazon Reviews 2023 categories, DARTS performs best on all twelve dataset–metric combinations, improving over the strongest baseline in each column by 6.8%–12.0%.Source: Table 1 · p. 7, §4.2 · p. 7

4.3 ASSESSING THE DIVERSITY-ACCURACY COOMPROMISE

  • DARTS combines R = 0.30 and PR = 4.7 with decoded-set dissimilarity J = 0.30, while the baselines produce nearly fixed slates.On SCIENTIFIC, DiffuRec has PR = 5.0 but R = 4.8 × 10^-5, preserving 98% of its top-10 between draws; DreamRec has R = 0.08 and PR = 0.005.Source: §4.3 · p. 8, Table 2 · p. 8

4.4 ABLATIONS

  • Adding catalogue cross-entropy inside diffusion reduces sampling diversity without improving ranking: at α = 1, PR falls to 0.65 and NDCG@5 falls 57% from α = 0.Even α = 0.1 reduces R by more than half and lowers PR from 4.7 to 1.5.Source: §4.4 · pp. 8–9
  • Jointly updating item representations during Stage 2 reduces R from 0.300 to 0.024, PR from 4.70 to 1.30, and NDCG@5 by 42%.The comparison uses latent regression loss.Source: §4.4 · p. 9
  • Averaging M = 4 draws improves DARTS NDCG@5 and NDCG@10 by 3.2% over M = 1 on SCIENTIFIC, while DiffuRec moves by at most 0.6%.Throughput falls 3.3× from M = 1 to M = 4; the M = 1 model is 3.6× faster than DiffuRec while achieving better ranking accuracy.Source: §4.4 · p. 9

5 DISCUSSION

Removing joint latent learning and in-trajectory catalogue supervision preserves stochastic variation through decoding; DARTS leads the reported comparisons, and aggregating samples improves ranking metrics. The discussion notes that leave-one-out evaluation cannot establish whether alternative samples are plausible futures.

5 DISCUSSION

  • On SCIENTIFIC, ranking by the arithmetic mean of independent draws improves DARTS NDCG@10, whereas DiffuRec remains nearly flat.Source: Figure 2 · p. 9
  • Removing both collapse-inducing choices preserves stochastic variation through decoding; DARTS leads every Table 1 comparison, and averaging draws improves all four ranking metrics.The discussion also reports that injecting noise into a discriminative model does not provide the same aggregation gain.Source: §5 · p. 9
  • Leave-one-out evaluation observes a single next interaction, so ranking metrics cannot determine whether alternative samples are plausible futures.Source: §5 · p. 9
  • Multi-future settings would permit direct measurement of calibration, controllable diversity, and coverage; structured outputs such as semantic identifiers are a next target.Source: §5 · p. 10

A.1 DATASET STATISTICS

The appendix documents dataset statistics, comparison models, module blocks, the three training stages, inference, and hyperparameter selection. It describes a staged pipeline in which later stages reuse frozen components and inference samples a target latent before catalogue ranking.

A.2 COMPARED MODELS

  • The comparison set spans sequential encoders, generative retrieval, and diffusion; controlled diffusion +T5 variants use DARTS’s hybrid item representations.Reported sources and models trained in a shared codebase are specified for the comparisons.Source: §A.2 · pp. 12–13

A.1 DATASET STATISTICS

A.3 MODULE BLOCK DIAGRAMS

  • The denoiser cross-attends from noisy target latents to prefix states, while the decoder uses causal attention over history and cross-attends to latent inputs before producing catalogue logits.Source: §A.3 · p. 13

A.4 TRAINING STAGES

  • Across the three stages, only designated parameters are updated; the decoder is pretrained in Stage 1 and fine-tuned in Stage 3, while Stage 2 has no catalogue-level objective.Source: Table 4 · p. 14

A.5 INFERENCE

  • At inference, the model encodes the prefix once, samples a final latent through the reverse diffusion steps, and uses decoder logits to rank the catalogue.The default evaluation uses one sample; repeated-sample evaluation varies diffusion noise while sharing the prefix.Source: §A.5 · p. 14

A.6 HYPERPARAMETERS

  • Checkpoint selection uses masked-item accuracy in Stage 1, latent RMSE in Stage 2, and NDCG@10 in Stage 3; Stage 3 uses fewer epochs because it runs the reverse chain on every training batch.Stage 2 is the longest stage in wall-clock time because validation runs the full reverse chain.Source: §A.6 · p. 14

B EFFECT OF ITEM REPRESENTATIONS

Semantic item features improve every evaluated model, but their gains vary substantially. DARTS is comparatively insensitive to representation choice and remains competitive against identifier-only baselines.

B EFFECT OF ITEM REPRESENTATIONS

C DISCRETE SUPERVISION INSIDE THE DIFFUSION PROCESS

A controlled sweep adds catalogue cross-entropy to diffusion training while holding other settings fixed. Even a small discrete-loss component sharply reduces latent diversity, and ranking accuracy worsens as its weight increases.

C DISCRETE SUPERVISION INSIDE THE DIFFUSION PROCESS

D EFFECT OF UPDATING ITEM REPRESENTATIONS DURING DIFFUSION TRAINING

The ablation finds that freezing item representations and using latent regression together yields the strongest diversity diagnostics and NDCG@5. The alternatives do not support a complete ordering.

Setup

  • The experiment compares frozen or updated representations with latent regression or decoded-logit cross-entropy, holding other components and hyperparameters unchanged.All four Stage-2 configurations start from the same Stage-1 checkpoint.Source: D EFFECT OF UPDATING ITEM… · p. 17

Results

  • Frozen representations with latent regression achieve the highest R, PR, and NDCG@5 among the tested configurations.Updating representations under latent regression reduces R from 0.300 to 0.024, PR from 4.70 to 1.30, and NDCG@5 by 42%.Source: D EFFECT OF UPDATING ITEM… · p. 17, Table 8 · p. 17
  • Updating representations under cross-entropy yields R = 0.050 and PR = 2.40, but NDCG@5 remains 48% below DARTS.These effects are not additive, so the three alternative configurations cannot be ranked against one another overall.Source: D EFFECT OF UPDATING ITEM… · p. 17
  • J does not track recommendation quality and is interpreted alongside latent diagnostics and NDCG.The updated-representation regression run has the lowest J (0.25) despite NDCG@5 being 42% below DARTS.Source: D EFFECT OF UPDATING ITEM… · p. 17

E STOCHASTIC DISCRIMINATIVE BASELINE

This section tests whether adding randomness to SASRec’s catalogue scores can reproduce the benefits of sampling learned predictive states. It finds that perturbation and aggregation alone do not close the gap to deterministic ranking.

Method

  • Gumbel noise is added independently to catalogue log-probabilities, and sorting each perturbed vector yields a Plackett–Luce ranking whose first item is an exact categorical draw.The baseline uses the checkpoint’s training temperature of 0.07.Source: E STOCHASTIC DISCRIMINATIVE BASELINE · pp. 17–18

Results

  • At M = 15, geometric-mean aggregation remains 11–14% below deterministic SASRec across all four metrics.The gains saturate before M = 15; the figure’s dashed line is deterministic SASRec and its insets report values and gaps at M = 15.Source: Figure 4 · p. 18, E STOCHASTIC DISCRIMINATIVE BASELINE · p. 19
  • A single Gumbel ranking loses about 80% NDCG@5 and 79% NDCG@10, while arithmetic aggregation remains about 78% and 75% below deterministic SASRec, respectively.The arithmetic-aggregation comparison is at M = 5.Source: E STOCHASTIC DISCRIMINATIVE BASELINE · p. 18, Table 9 · p. 18
  • Geometric aggregation averages away independent log-probability perturbations and converges back toward SASRec’s original ordering, rather than combining complementary predictions.At M = 5 it remains about 28% and 26% below deterministic SASRec on NDCG@5 and NDCG@10, respectively.Source: E STOCHASTIC DISCRIMINATIVE BASELINE · p. 19

F POOLING DRAWS: UNIONRECALL

Pooling ranked lists from multiple draws expands the slate but does not improve accuracy over a single longer list at a matched slate size. The section distinguishes this from probability-space aggregation, which improves performance while keeping the slate fixed.

Definition and measurement

  • UnionRecall@n(M) measures the fraction of contexts with the held-out item in at least one draw’s top-n; the mean pooled slate size records the accompanying item-count cost.Under leave-one-out evaluation, each context has one relevant item, and UnionRecall is non-decreasing in M by construction.Source: F POOLING DRAWS: UNIONRECALL · p. 19

Results

  • UnionRecall@5 rises from 0.0344 at M = 1 to 0.0547 at M = 30, a 59% relative increase, while the mean slate grows from 5 to 11.5 items.Thirty lists of five contain 11.5 distinct items rather than 150, indicating draws concentrated in a stable catalogue neighbourhood.Source: F POOLING DRAWS: UNIONRECALL · p. 19, Table 10 · p. 19
  • At M = 5, pooled top-5 lists show 8.0 items on average and recover the target for 4.47% of users; a single top-8 list shows the same number and recovers 4.62%.The single-draw curve dominates throughout, with the same direction at n = 10.Source: F POOLING DRAWS: UNIONRECALL · pp. 19–20

Why pooling does not improve accuracy

  • At matched slate size, a single longer list matches or beats pooling several draws, so sampler diversity alone is not an accuracy gain.Figure 5 compares UnionRecall@n(M) and single-draw hit rate@k against the number of items shown.Source: F POOLING DRAWS: UNIONRECALL · pp. 19–20, Figure 5 · p. 20

Probability-space aggregation

  • Probability-space aggregation averages per-draw distributions and truncates once, improving performance by roughly 3% against M = 1 at a fixed k-item slate.The comparison is budget-neutral because it keeps the number of shown items fixed.Source: F POOLING DRAWS: UNIONRECALL · p. 20

G INFERENCE COST

At one draw, DARTS is faster and more accurate than DiffuRec+T5, while additional draws improve ranking at a serving-cost premium. Against SASRec+T5, DARTS trades lower throughput for higher NDCG@10.

Comparison with DiffuRec

  • At M = 1, DARTS achieves 2,102 predictions per second versus 590 for DiffuRec+T5, with 11.1% higher NDCG@10.The comparison uses reported Table 1 ranking results.Source: G INFERENCE COST · p. 20
  • DARTS’s single-draw speed advantage over DiffuRec reflects lower per-step cost: a single latent query cross-attends to history rather than re-encoding the item sequence at each step.At the reported batch size, DARTS has 2.4 ms per-step latency versus DiffuRec’s 54 ms, offsetting DARTS’s 6.25× greater number of reverse steps.Source: G INFERENCE COST · p. 20

Cost of additional samples

  • Each added draw repeats the complete reverse diffusion chain, so increasing M reduces throughput approximately proportionally.The paper characterizes M as an explicit quality–compute trade-off; DARTS diversity is present even at M = 1.Source: G INFERENCE COST · pp. 20, 22
  • Increasing DARTS from M = 1 to M = 4 improves NDCG@10 by 3.2% but reduces throughput from 2,102 to 629 samples per second, a 3.3× reduction.At M = 4 throughput remains slightly above DiffuRec+T5’s 590 samples per second; at M = 5 it falls below that value.Source: Figure 6 · p. 21, Figure 7 · p. 21, G INFERENCE COST · p. 22

Comparison with a non-generative recommender

  • SASRec+T5 reaches 30,570 samples per second, 14.5× DARTS throughput at M = 1, but has 14.2% lower NDCG@10.Throughput is measured as complete predictions per second, including encoding, reverse diffusion, final-position decoding, and catalogue scoring.Source: G INFERENCE COST · pp. 20, 22

Related papers

Loading 2610.36954v1…