acceptodds
Under review as a conference paper at ICLR 2027

Unified Framework for Generative Sequential Recommendation

Abstract

How much variation a generative recommender’s sampler introduces on the way to a sample, its stochasticity level, is a property of the user rather than of the task. A long history already determines the next item, so exploration can only cost; a short history, or interests spanning several catalog categories, leaves the target under-determined. No existing generative recommender can act on this, because none can vary the level at all: it is determined jointly by the interpolant and the training objective. We remove that constraint. Denoising diffusion (DDPM), its deterministic counterpart (DDIM), Flow Matching and Schrödinger Bridges are one SDE at different values of that scalar; the theory unifying them holds it fixed across the population, and extending that theory frees it to be any measurable function of the user, every conditional marginal preserved. The level appears in neither training objective, so a trained checkpoint reproduces these processes at inference rather than committing to one of them before training. Serving it per user pays in diversity at no cost in accuracy, on corpora whose users genuinely differ, and when the next several items are predicted at once.

Primary area:
generative models
Keywords:
Sequential Recommendation, Generative Recommender

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_PqHWOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes STIR, a framework for controlling sampling stochasticity in generative sequential recommendation. For a fixed training interpolant, it fits velocity and score fields and uses them in an SDE whose stochasticity level ϵ(t,h)\epsilon(t,h) can be selected at inference from the user's history. With exact fields and appropriate regularity, the score correction compensates for the injected noise and preserves the conditional interpolant marginals. Setting epsilon to zero yields an ODE, which is deterministic conditional on its initial noise draw.

The paper relates specific interpolant and sampler configurations to flow matching, Schrodinger bridges, DDPM, and DDIM; the empirical reconstruction study uses three interpolant-specific checkpoints. It further analyzes path choice and user-dependent stochasticity using a Gaussian-mixture model of the target. The practical personalization scheme clusters history representations into five groups and selects a stochasticity level for each group on validation data. Experiments on seven public benchmarks examine recommendation accuracy, sampling dispersion, catalog coverage under accuracy constraints, integration with DreamRec and PreferDiff, and prediction of multiple future items.

Critical strengths
  1. Treating the sampling procedure as a configurable component of a generative recommender is practically useful. The separation of velocity and score estimation provides a convenient way to explore stochasticity choices after training a fixed interpolant. The conditional marginal-preservation argument explains why the score correction is necessary and why conditioning the level on a fixed user history differs from making it depend on the evolving sample state.

  2. The empirical investigation is broader than a single accuracy comparison, it includes seven public benchmarks, multiple seeds, post-hoc diversification baselines, synthetic analyses, and two existing recommendation backbones. The group-based policy also improves the reported multi-target NDCG in all six settings of Table 3.

Critical weaknesses
  1. The novelty argument and baseline comparison need stronger support. DDIM already allows inference-time stochasticity control without retraining, and selecting its noise level based on a fixed user history does not inherently require changing the training schedule. A comparison with existing samplers using the same user-conditioned tuning would help clarify the benefit of STIR's learned fields and interpolant beyond sampler tuning itself. The unification claim also needs clarification, Section 5.1 uses three checkpoints, so the experiment does not demonstrate exact recovery of all four processes from a single fixed checkpoint.

  2. Some theoretical conclusions go beyond what the proofs establish. The uniqueness claim in Theorem 5(a) does not hold, since a quadratic time schedule can also give finite irreducible velocity loss. The history-length argument omits a nonzero within-mode loss, so it does not establish that the loss vanishes or justify the proposed switch-off rule. Lemma 12 only addresses Top-K set membership, not the ranking within the set, and therefore does not establish invariance of NDCG or MRR or the claimed limitation of re-ranking. These gaps weaken the theoretical support for the recommendation-specific conclusions.

Questions for authors
  1. Can the authors compare STIR against a fixed diffusion checkpoint using a user- or group-conditioned DDIM η\eta within its admissible range, with matched history groups, validation budgets, sampling budgets, and evaluation protocols? This would help separate the benefit of personalized sampler selection from that of STIR's learned fields and interpolant.

  2. Could the authors address the concerns above about Theorem 5(a), Proposition 7(b), and Lemma 12, and clarify the scope of these claims? Please also point to or provide the proof of the optimal noise envelope in Theorem 5(b); Appendix H.1 proves Theorem 3, while H.2 covers only 5(a).

  3. Eqs. (12)-(14) imply a per-sample score-regression target of −ξt/ς(t)-\xi_t/\varsigma(t), whereas Eq. (15), as written, uses +ξt/ς(t)+\xi_t/\varsigma(t). Is this a typographical error, and which sign was used in the implementation? Please also clarify the scope of the claimed equivalence to native training objectives after adding the retrieval loss.

  4. What does "recovery" mean in Section 5.1: equivalence of sampling equations, equality of population minimizers, or agreement of empirical results? On ML-1M, Table 2 reports NDCG@20 of 0.0442±0.00270.0442 \pm 0.0027 for native DDPM and 0.0637±0.00430.0637 \pm 0.0043 for the corresponding VP-STIR configuration. How does this difference relate to the recovery claim, and what criterion is used to judge agreement?

  5. How sensitive is the group-based rule to the number of clusters and the validation sample size? For the configurations selected under each accuracy budget, please report test NDCG@20 alongside coverage, together with variability across validation subsets and results for short-history users.

2 comments

Reviewer PqHW

We thank the reviewer for the careful technical review.

W1, Q1: DDIM η baseline

We ran this ablation on a frozen VP checkpoint, with STIR's groups (C=5), 512 validation users and protocol.

ML-1M NDCG@20ML-1M Cov@10Beauty NDCG@20Beauty Cov@10
native DDIM (η=0)0.03390.74640.01370.8890
best global η (η=1)0.04530.69050.01810.7745
per-cluster η0.04250.71340.01740.7982

Per-cluster η loses 4–6% NDCG@20 against the best global η; its +0.02–0.03 coverage only moves along the same trade-off. On VP, raising η even reduces coverage (0.746 → 0.690), whereas STIR's ε increases it (Fig. 2a, 3).

Why η cannot be personalized. By Thm. 6, the optimal level depends only on r(h)=δb(h)/δs(h)r(h)=\delta_b(h)/\delta_s(h) (Eq. 10), which users' modes and history move (Prop. 7). On VP, b=ωvp(x+s)b=\omega_{vp}(x+s) (App. D.3): both fields come from one noise head, so δb\delta_b and δs\delta_s are tied, r(h)r(h) barely varies, and all groups get the same level. STIR fits two separate heads (Eq. 7), so ϵ⋆(h)\epsilon^\star(h) can vary with the user.

W1, Q4: Meaning of recovery

Since the path is fixed at training, recovery is meant per path: one checkpoint recovers every process defined on its path: (I_lin, γ_σ) gives FM (ε=0) and SB (ε=σ²/2), (I_vp, 0) gives DDIM and DDPM; (I_lin, 0) is only an envelope ablation. It holds at three levels: same SDE (all four), same population minimizer (exact for FM, DDPM; at the optimum for SB, DDIM; App. D). The ML-1M DDPM gap comes from the schedule, not from ε. Native DDPM's discrete schedule (T=50, β up to 0.02) gives ᾱ_T ≈ 0.60 (x_T ≈ 0.78 x₁ + 0.63 x₀): it never trains on pure noise yet samples from it. VP-STIR's continuous schedule gives ᾱ(1) ≈ 5e-5, so training matches sampling (0.0637 vs 0.0442). We clarify this in §5.1.

W2, Q2: Theorem 5(a)

The uniqueness in Theorem 5(a) refers to the transport cost: the linear path is the only affine path minimizing it. The finiteness of the irreducible velocity loss was stated in contrast to the variance-preserving path, on which it diverges (App. H.2), not as a second uniqueness claim; we agree the wording suggested otherwise. The quadratic schedule α=1−t2\alpha=1-t^2, β=t2\beta=t^2 is indeed our straight line with a different clock, xt=(1−s)x0+s x1x_t=(1-s)x_0+s\,x_1, s=t2s=t^2, with finite loss ∫01s˙ V(s) ds≤2∫01V(s) ds\int_0^1\dot s\,V(s)\,ds\le 2\int_0^1V(s)\,ds. It is excluded by the transport cost: ∫01E∥x˙t∥2dt=∫01α˙2dt E∥x0∥2+∫01β˙2dt E∥x1∥2 ≥ E∥x0∥2+E∥x1∥2,\int_0^1\mathbb E\|\dot x_t\|^2dt=\int_0^1\dot\alpha^2dt\,\mathbb E\|x_0\|^2+\int_0^1\dot\beta^2dt\,\mathbb E\|x_1\|^2\ \ge\ \mathbb E\|x_0\|^2+\mathbb E\|x_1\|^2, by Cauchy–Schwarz, with equality iff α=1−t\alpha=1-t, β=t\beta=t; the quadratic clock costs 4/34/3 per term. We rephrase 5(a): the linear path uniquely minimizes the transport cost, and its irreducible velocity loss is finite, whereas it diverges on the VP path.

Q2: Proof of Theorem 5(b)

It was mistakenly left out and is now App. H.3.

W2, Q2: Proposition 7(b)

We agree. By App. H.2, Lb∗(h)=∫01Q dt+∫01Υ2 E[Varm∣xt(μm)] dt\mathcal L_b^*(h)=\int_0^1 Q\,dt+\int_0^1\Upsilon^2\,\mathbb E[\mathrm{Var}_{m|x_t}(\mu_m)]\,dt. History removes only the mode-confusion term, at rate O(e−cT)O(e^{-cT}); the within-mode term ∫01Q dt>0\int_0^1 Q\,dt>0 remains. We restate 7(b): ϵ⋆(h)\epsilon^\star(h) decreases with TT toward the level this term sets.

W2, Q2: Lemma 12

Indeed, ΔK\Delta_K fixes the Top-K set, not its order. As standard (SASRec, DreamRec), we evaluate on one held-out item (App. B), so NDCG and MRR depend only on its rank rr. We add: if ∥x^′−x^∥<Δr/2\|\hat x'-\hat x\|<\Delta_r/2, with Δr\Delta_r the smaller score gap between the target and its two neighbours, rr is unchanged, hence NDCG and MRR. HR@K, used in Fig. 5, needs only ΔK\Delta_K.

Q3: Sign in Eq. (15)

A typo, now fixed: the target is −ξt/ς(t)-\xi_t/\varsigma(t). The code samples with s^=−e^/ς(t)\hat s=-\hat e/\varsigma(t), so results are unaffected. The equivalence covers the regression part; natives carry the same retrieval term (App. C).

Q5: Sensitivity of the group rule

We ran all four analyses (new appendix). ML-1M, τ=1.00: each group gets the ε with the highest coverage keeping validation NDCG@20 at the deterministic level, chosen on 512 validation users. Values are relative to C=5 (Cov@10 67.6%, NDCG@20 0.0814); ± is the SD over 5 validation subsets.

C520100
Δ Cov@100+20.9% ± 2.1+27.5% ± 0.9
Δ test NDCG@200−8.9%−36.7%
short history: Δ Cov@100+49.8% ± 2.5
short history: Δ NDCG0−35.3%

Larger C buys coverage at the cost of test NDCG, as each choice rests on fewer users (§5.3): C=20 gains 21% coverage for 9% NDCG, C=100 only 28% for 37%. More validation users make the choice more reliable:

validation users1282565121024
Δ Cov@10+16.9% ± 6.2+4.4% ± 13.50 ± 16.0−4.9% ± 14.9
Δ test NDCG@20−18.6% ± 21.2+0.7%0+5.1%

We recommend C ≤ 5 and ≥1,024 validation users; all budgets are in the appendix.

Thank you for your response. My core concern about novelty remains unresolved, and I will maintain my score.

Reviewer_tQBLOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper proposes STIR, a unified framework that allows generative recommendation models to adjust their sampling stochasticity at inference time without retraining.
It argues that users with diverse interests or shorter histories benefit from more stochastic sampling, while users with clear preferences need less.

Critical strengths
  1. Clear motivation
  • The motivation for personalized stochasticity is sound to me.
Critical weaknesses
  1. Truly personalized stochasticity?
  • The paper does not sufficiently establish how to estimate the optimal stochasticity level for each user at inference time. The group-wise validation strategy demonstrates the potential benefit of personalization, but does not fully validate a practical per-user policy.
  1. Scope of the paper.
  • The title "UNIFIED FRAMEWORK FOR GENERATIVE SEQUENTIAL RECOMMENDATION" is somewhat too broad.
    This paper focuses on the personalized stochasticity of diffusion-based recommenders.
  • Also, "Generative Recommendation" generally means LLM-based multi-token item representations.
    ex) Recommender systems with generative retrieval, NeurIPS'23
Questions for authors

Please refer to weaknesses.

2 comments

We are grateful to the reviewer for the thoughtful review and the helpful suggestions, which we address point by point below.

W1: Per-user stochasticity

The cohort rule is a deliberate design choice: it is cheap, label-free and directly deployable. At inference. This is already a per-user policy in the sense of Thm. 3 (ε may be any function of h), fitted on validation users and evaluated on held-out test users (Fig. 3, Table 5). Pooling users also makes the choice statistically reliable: a level chosen from a single validation item per user can be dominated by noise, whereas a cohort's level rests on many users.

To further show the benefit of personalization, as the reviewer suggested, we now choose ε per individual user from the user's own last interaction, in two ways: the naive per-user argmax, and an empirical-Bayes estimator that centers each user's utility curve over ε and shrinks it toward the cohort's curve with weight λ = σ²_between / (σ²_between + σ²_within/n), estimated from the data and cross-fitted on two halves of the users (details in App. O). Test Coverage@10 on ML-1M, same decodes as Fig. 3:

τGlobalCohortPer-user, naivePer-user, shrunk
1.000.5520.568 (+0.016)0.807 (+0.255)0.803 (+0.251)
0.980.5520.596 (+0.044)0.807 (+0.255)0.803 (+0.251)
0.950.5520.601 (+0.050)0.807 (+0.255)0.804 (+0.252)
0.900.5760.640 (+0.065)0.807 (+0.231)0.804 (+0.228)
0.800.6250.725 (+0.100)0.807 (+0.182)0.804 (+0.179)

Choosing ε per user gains up to +0.25 coverage over the global level, far beyond cohorts, at every budget including τ = 1.00. The estimated λ (0.82 to 0.86) shows the per-user signal is real, so shrinkage costs almost nothing here; when the signal is weak, as for accuracy (λ ≈ 0.1), it falls back to the cohort instead of fitting noise.

W2: Scope of the paper

We agree that the title is broader than the paper. Our scope, stated in the introduction, is the iterative-refinement family: denoising diffusion (DDPM, DDIM), Flow Matching and Schrödinger Bridges, which transport a source distribution to the data. We will retitle the paper "Unified Framework for Transport-Based Sequential Recommendation" and restrict the claims in the abstract, introduction, Table 1 and conclusion to this family. The revised App. A already positions SID-based generative retrieval (Rajput et al., NeurIPS 2023) as a distinct paradigm.

Thanks for the response. I will maintain my score.

Reviewer_ZsqNOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper investigates user-adaptive stochasticity in generative recommendation, arguing that different users require different levels of exploration depending on their interaction histories and interest diversity. The authors propose a unified stochastic differential equation (SDE) framework that connects DDPM, DDIM, Flow Matching, and Schrödinger Bridges, enabling user-specific stochasticity adjustment at inference time without modifying the training objective. Experiments demonstrate that the proposed approach improves recommendation diversity without sacrificing accuracy, particularly for users with heterogeneous preferences and multi-item prediction scenarios.

Critical strengths
  1. The topic of generative sequential recommendation is of great value to study, especially its wide application in industry.
  2. The motivation of studying the user-adaptive stochasticity in terms of long and short-term history makes sense.
Critical weaknesses
  1. It lacks a comprehensive reproduction and comparison with existing generative recommendation models. It tries to build the recommender models from scratch based on the generative models, which just discard the prior works/efforts in recommender system community.
  2. Regarding exploration and exploitation, many works have been studied in either reinforcement learning or recommender system, this work should make reference to these prior works, make extensive comparisons, and justify the motivation and novelty.
Questions for authors

please see the above weaknesses.

1 comment

We thank the reviewer for the constructive comments.

W1: Comparison with generative recommenders

We believe there is a misunderstanding: STIR is not a new recommender built from scratch, but a plug-in for the generative part of an existing diffusion or flow recommender, agnostic to its input, architecture, loss and training scheme. Existing generative recommenders use diffusion and flow models, originally developed for computer vision, as their generative backbone. Their contributions lie around this backbone, in the input, output, loss and training scheme, while the backbone itself is adopted as is: the only design question is which one to use. Its stochasticity level ε is inherited unchanged. Yet the uses differ: in vision the model is a sampler (many diverse outputs per prompt), in recommendation it is a predictor (one embedding per user, ranked by accuracy) whose uncertainty depends on the user's history, so the right level differs across users. STIR addresses this inherited part, orthogonal to prior work.

A generative recommender has four parts: an encoder that reads the user's history, a network trained to denoise, a sampler that turns noise into an item embedding, and a readout that ranks items against it. STIR keeps the encoder and readout and generalizes the generative part:

ComponentExisting recommenderSTIR
Inputfixed by the methodany (Thm. 3)
History encoderfixed by the methodany
Readoutfixed by the methodany
Pathfixed: variance-preserving (VP) or linearany path and any envelope γ(t)
Learned fieldsone (noise or velocity)velocity and score (one extra head, backbone frozen)
Objectivedenoising loss + extra terms (ranking, preference, retrieval)any; ε is in no loss (Eq. 7)
Stochasticityfixed by the objectiveε(t, h) per user, marginals preserved (Thm. 3)

The paper already includes this comparison in Table 2: plugged into DreamRec and PreferDiff, STIR improves normalized discounted cumulative gain:

NDCG@20ToysYelpGowallaGoodreads
DreamRec0.01340.01060.01200.0579
+ STIR0.02240.02100.03150.1235
PreferDiff0.02010.00540.01410.0679
+ STIR0.02820.02100.03110.1224

STIR is complementary to existing work: it can only add to it. FM, SB, DDIM and DDPM are all special cases of STIR, each one a particular level. A method's own sampler is therefore one of the levels STIR selects from, so STIR either keeps it or finds a better one. Comparing complete systems would mix architecture, objective and training; the meaningful test is the same model with and without STIR.

W2: Exploration vs. exploitation

Exploration in bandits and RL and the stochasticity STIR controls are different things. Bandit and RL methods (e.g., LinUCB, SlateQ) optimize long-term reward through interaction, departing from the current estimate to gather information. STIR controls the stochasticity of an offline sampler drawing from a fixed learned conditional p(·|h), with every conditional marginal preserved (Thm. 3), and involves no reward, feedback loop or policy update. The two are orthogonal and composable: ε could serve as the action of a bandit policy, and STIR as its candidate generator. We have (i) added a related-work paragraph positioning STIR against diversity re-ranking, bandits and RL (App. A), (ii) reserved "exploration" for its bandit/RL sense and "stochasticity" for the sampler throughout, and (iii) added an ε-greedy baseline to Table 5.

In existing diffusion and flow recommenders ε is fixed by the training objective (App. E), so no policy can set it per user without retraining; STIR removes this constraint. In our submission, we already compare against MMR and DDP re-ranking, and temperature sampling, i.e., Boltzmann exploration (Table 5, Fig. 3). At the strictest budget (τ = 1.00) all three stay within 0.005 of the deterministic sampler's coverage, while per-user ε gains on all seven corpora (+0.3 to +10.4 points). Per your request, we also trained bandit and RL policies that choose ε per user from six label-free features, rewarded by validation NDCG@20. Test NDCG@20, ML-1M:

Policy choosing ε8 levelsΔ
ε=00.0885
p-greedy, non-contextual, p=0.050.0875−1.1%
p-greedy, non-contextual, p=0.50.0882−0.3%
p-greedy, contextual, p=0.050.0848−4.1%
p-greedy, contextual, p=0.50.0843−4.7%
Q-regression (full info)0.0860−2.8%
Group Relative Policy Optimization (GRPO)0.0872−1.5%

Every reward-driven policy underperforms ε=0 on accuracy: departing from the learned conditional to gather information costs NDCG offline, whereas STIR's level changes dispersion while preserving that conditional. This is why we treat the two as complementary rather than competing.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.