betaFeedback

WorksComputer Science

Beyond Semantic IDs: Atomic Creator Tokens for Large-Scale Generative Recommendation

Abstract

Recommending creators that a user does not yet follow is a cold-start problem: from hundreds of thousands of popular creators and only a short trail of prior engagements, the system must predict the next creator worth surfacing. The prevailing generative recommendation recipe tokenizes each creator into a Semantic ID (SID) by quantizing a frozen embedding and trains a language model to generate SIDs. We show that this SID layer, not the language model, is the binding constraint. The creator-to-SID map is many-to-one, so colliding creators must be resolved by a popularity tie-break that caps creator-level recall and favors already-popular creators, and the frozen quantizer stops the training gradient from reaching the creator representation. We propose Atomic Creator Token (ACT), which gives every creator its own vocabulary token trained end-to-end. The map is bijective, so collisions are impossible and inference is a direct top-kk over creator logits. In a controlled comparison on production data in which the token layer is the only variable, ACT improves creator-level Recall@15 by 19.8 points (1.84×1.84\times) over a matched SID baseline, still leads when the SID codebook is enlarged 64×64\times, and serves at ∼2.7×{\sim}2.7\times lower per-user latency. The advantage carries over to the public KuaiRand dataset, where ACT is the only method that beats a global-popularity baseline, provided each creator token sees enough training examples. In a live A/B test, ACT is directionally positive on every reported metric against both the SID-enhanced and the classic production rankers.

Introduction

Recommending creators that a user does not already follow, i.e., unconnected creator recommendation, is a cold-start problem at ecosystem scale. Out of a multi-million catalog, the system must pick from hundreds of thousands of popular creators the few a user will enjoy, with no follower connection and only a short, sparse trail of prior engagements (opens, dwells, follows) on previously suggested creators. Because this signal is so limited, how the model represents creators largely determines what it can recommend. We focus on the head of the catalog, the creators above a follower threshold that a platform can realistically promote and where engagement concentrates: as the head vocabulary grows it covers a rapidly rising share of engagement, so a larger head vocabulary delivers more effective recall (§ 4.4).

The dominant generative recipe, exemplified by OneRec [2] and the Transformer Index for Generative Recommenders (TIGER) [21], recasts recommendation as sequence generation over Semantic IDs (SIDs). A pretrained graph neural network (GNN) or two-tower model produces a frozen creator embedding, which residual quantization [3] compresses into a short sequence of codebook indices; a language model is then fine-tuned to generate the next SID, and the generated SIDs are decoded back to creators through a codebook join and a popularity rerank.

We argue that this SID layer, not the language model, is the binding constraint, for two compounding reasons. First, the creator-to-SID map is surjective: the used code space is far smaller than the catalog, so many creators share one SID. Even a model that predicts the right SID must still guess which of its creators the user wants, and the popularity tie-break that resolves this guess caps creator-level recall and favors already-famous creators, the opposite of what surfacing fresh creators requires. Second, the quantizer is frozen and non-differentiable: the next-token gradient never reaches the creator representation, so the tokenizer and the language model optimize mismatched objectives. The same design also forces a codebook rebuild for every new creator and a multi-step decode-and-join at serving time.

We therefore propose Atomic Creator Token (ACT), which removes the SID layer entirely (Figure 1). ACT expands the language model’s vocabulary with one dedicated token per candidate creator, up to 30Y30Y tokens in our study, and trains the model end-to-end to generate the next creator token. Because the creator–token map is bijective, collisions are impossible and inference reduces to a direct top-kk over creator logits with no codebook join or popularity rerank. Because the creator embeddings are model parameters, the next-token gradient trains them directly, which also makes on-policy alignment from engagement feasible. The added capacity is modest: a 10Y×307210Y \times3072 embedding table in bf16 adds ≈0.6\approx0.6 GB.

Figure 1. Overview of ACT. (01) Each candidate creator gets one dedicated vocabulary token through a bijection ϕ\phi, instead of a many-to-one Semantic ID. (02) The creator embeddings are warmed up with the backbone frozen and then fine-tuned end-to-end with Low-Rank Adaptation (LoRA), so the gradient reaches them. (03) At inference, a single decode step per creator yields a top-kk over creator tokens, decoded by an O(1)O(1) lookup without a codebook join or popularity rerank.

ACT is not a drop-in replacement for arbitrarily large catalogs. One token per creator needs enough engagement examples per token to learn a useful embedding, and the output softmax grows with the vocabulary, so we position ACT as the head component of a routed recommender: ACT covers every above-threshold creator, whereas a frozen SID codebook reaches only the 86.3% with a GNN embedding, and conventional retrieval serves the sparse long tail. Our public-data study makes this premise explicit (§ 4.5). Our contributions are:

  1. a diagnosis of the SID bottleneck in cold-start creator recommendation: a surjective creator-to-SID map imposes a collision-limited recall ceiling, and a frozen quantizer blocks end-to-end representation learning;

  2. ACT, a generative recommender with a bijective per-creator vocabulary trained end-to-end, which is collision-free and join-free at inference and fits in a single-host memory budget at 30Y30Y creators;

  3. a controlled evaluation on production data in which the token layer is the only variable: ACT improves creator-level Recall@15 by +19.8 points (1.84×) over a matched SID baseline at a 10Y10Y-creator vocabulary, still leads when the SID codebook is enlarged to (4X)3(4X)^3 at 30Y30Y creators, and is cheaper to serve;

  4. a replication on the public KuaiRand dataset, together with an analysis of when ACT helps: it wins once each creator token has enough training examples, and a live A/BA/B test that is directionally positive on every reported metric.

Method

We first define the task and the SID baseline that ACT replaces (§ 2.2), then describe ACT’s tokenization and objective (§ 2.3) and its training pipeline (§ 2.4).

Task and Semantic-ID Generation

Task and Semantic-ID Generation

Task. Let U\mathcal{U} be the set of users and A\mathcal{A} the set of candidate creators. Each user u∈Uu \in\mathcal{U} has profile features xux_u (country, language, gender, age) and a chronological engagement history Htu=((a1,τ1),…,(at,τt))H_t^u = ((a_1,\tau_1),\ldots,(a_t,\tau_t)), where ai∈Aa_i \in\mathcal{A} is the ii-th engaged creator and τi\tau_i its engagement type (open, like, share, follow, or long dwell). An engagement-value function v(τ)≥0v(\tau) \ge0 scores these types, increasing from a brief open to a follow; we use it to weight NDCG and to shape rewards. The task is next-creator prediction: a model Pθ(at+1∣Htu,xu)P_\theta(a_{t+1} \mid H_t^u,x_u) with parameters θ\theta returns the KK highest-scoring creators A^iK\hat{\mathcal{A}}_i^K for each test instance ii, and we report

Recall⁡@K=1N∑i=1N1[ai⋆∈A^iK].(1)\operatorname{Recall}@K = \frac{1}{N}\sum_{i=1}^{N}\mathbb{1}\left[a_i^\star\in\hat{\mathcal{A}}_i^K\right]. \tag*{(1)}

where NN is the number of test instances, ai⋆a_i^\star is the ground-truth next creator of instance i=1,…,Ni = 1,\ldots,N, and 1[⋅]\mathbb{1}[\cdot] is the indicator function.

Semantic IDs. The SID baseline maps each creator to an embedding with a pretrained, frozen encoder E:A→RdEE:\mathcal{A}\to\mathbb{R}^{d_E} (a GNN) and quantizes it with LL codebooks, each holding MM centroids {eℓ,k}k=1M\{e_{\ell,k}\}_{k=1}^{M}, by greedily quantizing successive residuals:

r(0)=E(a),sℓ(a)=arg⁡min⁡k∈[M]∥r(ℓ−1)−eℓ,k∥2,r(ℓ)=r(ℓ−1)−eℓ,sℓ(a).(2)r^{(0)}=E(a),\qquad s_\ell(a)=\underset{k\in[M]}{\arg\min}\left\|r^{(\ell-1)}-e_{\ell,k}\right\|_2,\qquad r^{(\ell)}=r^{(\ell-1)}-e_{\ell,s_\ell(a)}. \tag*{(2)}

where dEd_E is the embedding dimension, r(ℓ)r^{(\ell)} is the residual left after level ℓ=1,…,L\ell=1,\ldots,L, eℓ,ke_{\ell,k} is the kk-th centroid of codebook ℓ\ell, and [M]={1,…,M}[M]=\{1,\ldots,M\}. The SID of creator aa is s(a)=(s1(a),…,sL(a))∈[M]Ls(a)=(s_1(a),\ldots,s_L(a))\in[M]^L, and

ν(s)=∣{a∈A:s(a)=s}∣(3)\nu(s)=\left|\{a\in\mathcal{A}:s(a)=s\}\right| \tag*{(3)}

is the collision count of code ss, the number of creators that share it. Because the reachable code space is effectively far smaller than ∣A∣|\mathcal{A}|, the map ss is surjective and ν(s)>1\nu(s)>1 for many codes.

Generation and inference. A language model generates the LL code tokens of the next SID autoregressively, Pθ(s∣H,x)=∏ℓ=1LPθ(sℓ∣s<ℓ,H,x)P_\theta(s\mid H,x)=\prod_{\ell=1}^{L}P_\theta(s_\ell\mid s_{<\ell},H,x), and is trained with supervised fine-tuning (SFT). Its gradient stops at the SID-token embeddings, so neither EE nor the codebooks adapt to the recommendation objective. At inference, beam search produces candidate SIDs, each SID is expanded to its ν(s)\nu(s) creators by a codebook join, and a popularity score breaks the ties before the top-KK creators are returned.

Atomic Creator Tokens

ACT keeps this generative setup but removes the SID layer. It assigns each creator a unique token through a bijection ϕ:A→{0,1,…,NA−1}\phi:\mathcal{A}\to\{0,1,\ldots,N_{\mathcal{A}}-1\}, where NA=∣A∣N_{\mathcal{A}}=|\mathcal{A}|, and expands the base vocabulary V\mathcal{V} of the language model with the creator tokens {⟨aj⟩}j=0NA−1\{\langle a_j\rangle\}_{j=0}^{N_{\mathcal{A}}-1}. Whereas the SID map has ν(s)≥1\nu(s)\ge1, ACT has ν≡1\nu\equiv1: no two creators share a token, and decoding a token back to a creator is an O(1)O(1) lookup ϕ−1\phi^{-1} instead of a one-to-many codebook join.

Model. We use a Llama-3.2-3B backbone ([5]) with hidden size d=3072d=3072. Each creator token adds one ordinary vocabulary embedding wϕ(a)∈Rd\mathbf{w}_{\phi(a)}\in\mathbb{R}^{d}, no different in kind from the model’s word-piece embeddings. A context z=(xu,Htu)z=(x_u,H_t^u) is serialized into a prompt, with the profile as text and the history as the sequence of the engaged creators’ tokens, and the backbone fθf_\theta produces the hidden state h=fθ(z)∈Rd\mathbf{h}=f_\theta(z)\in\mathbb{R}^{d} at the position that predicts the next creator. With decoding restricted to creator tokens, the next-creator distribution is the usual next-token softmax over A\mathcal{A}:

Pθ(a∣z)=exp⁡(wϕ(a)Th)∑a′∈Aexp⁡(wϕ(a′)Th).(4)P_\theta(a\mid z)=\frac{\exp(\mathbf{w}_{\phi(a)}^{\mathsf{T}}\mathbf{h})}{\sum_{a'\in\mathcal{A}}\exp(\mathbf{w}_{\phi(a')}^{\mathsf{T}}\mathbf{h})}. \tag*{(4)}

where a′a' ranges over all candidate creators. A creator is thus emitted as a single token rather than an LL-token SID.

Objective. We train with the next-token cross-entropy on the target creator,

LSFT=−E(z,a⋆)[log⁡Pθ(a⋆∣z)],(5)\mathcal{L}_{\mathrm{SFT}} = -\mathbb{E}_{(z,a^\star)}\left[\log P_\theta(a^\star\mid z)\right], \tag*{(5)}

where the expectation E\mathbb{E} is over training pairs of a context zz and its ground-truth next creator a⋆a^\star. There is no separate creator encoder: assigning each creator a vocabulary token is the entire representational change. Because the creator embeddings live inside the model, the loss updates each of them directly,

∇wjLSFT=(pj−1[j=ϕ(a⋆)])h,pj=Pθ(ϕ−1(j)∣z),(6)\nabla_{\mathbf{w}_j}\mathcal{L}_{\mathrm{SFT}} = \left(p_j-\mathbb{1}[j=\phi(a^\star)]\right)\mathbf{h}, \qquad p_j=P_\theta(\phi^{-1}(j)\mid z), \tag*{(6)}

where j=0,…,NA−1j=0,\ldots,N_A-1 indexes the creator tokens and pjp_j is the predicted probability of creator ϕ−1(j)\phi^{-1}(j). The SID pipeline, in contrast, has a zero gradient with respect to its frozen encoder EE.

Why direct tokens dominate. An SID is a deterministic, lossy function of its creator, so by the data processing inequality an SID history carries no more information about the next SID than a creator history carries about the next creator. The best next-token loss ACT can reach is therefore no worse than that of any SID scheme, for any vocabulary size.

Contrastive auxiliary loss (optional). To sharpen the discriminability of creator tokens, an in-batch contrastive term can be added over a batch B\mathcal{B} of context–creator pairs, treating the other creators in the batch as negatives:

Lcon=−1∣B∣∑i∈Blog⁡exp⁡(hiTwϕ(ai⋆)/κ)∑j∈Bexp⁡(hiTwϕ(aj⋆)/κ),(7)\mathcal{L}_{\mathrm{con}}=-\frac{1}{|\mathcal{B}|}\sum_{i\in\mathcal{B}}\log\frac{\exp(\mathbf{h}_i^{\mathsf{T}}\mathbf{w}_{\phi(a_i^\star)}/\kappa)}{\sum_{j\in\mathcal{B}}\exp(\mathbf{h}_i^{\mathsf{T}}\mathbf{w}_{\phi(a_j^\star)}/\kappa)}, \tag*{(7)}

where hi\mathbf{h}_i and ai⋆a_i^\star are the hidden state and target creator of pair ii, the index jj runs over the pairs in B\mathcal{B}, and κ>0\kappa>0 is a temperature; training then minimizes LSFT+λLcon\mathcal{L}_{\mathrm{SFT}}+\lambda\mathcal{L}_{\mathrm{con}} with weight λ>0\lambda>0. Our ablation finds that this term does not help § 4.7.

Training Pipeline

Initialization. By default, the new creator rows are initialized randomly, wϕ(a)∼N(0,σ02I)\mathbf{w}_{\phi(a)}\sim\mathcal{N}(0,\sigma_0^2\mathbf{I}) with standard deviation σ0=0.02\sigma_0=0.02. Optionally, they can be warm-started from the GNN embedding through a learned linear projection P\mathbf{P}, wϕ(a)=PE(a)\mathbf{w}_{\phi(a)}=\mathbf{P}E(a); unlike the SID codebook, this serves only as an initialization and the embedding remains fully trainable. Creators added later receive new rows without a codebook rebuild.

Stage 1: creator-token warm-up. With the backbone frozen, we train only the new creator rows on creator descriptions and co-engagement sequences, so that the tokens are grounded before sequential training.

Stage 2: sequential SFT. We adapt the backbone with Low-Rank Adaptation (LoRA) Hu et al. [10] and train the creator rows jointly with LSFT\mathcal{L}_{\mathrm{SFT}} on (profile, history) →\to next-creator examples.

Stage 3: post-training (optional). Because the creator tokens are part of θ\theta, the policy πθ(a∣z)=Pθ(a∣z)\pi_\theta(a\mid z)=P_\theta(a\mid z) is differentiable over the true action space A\mathcal{A}, so it can be aligned from engagement with a value-weighted reward

r(a)=∑τv(τ)1[(a,τ)∈Fu],(8)r(a)=\sum_{\tau}v(\tau)\mathbb{1}[(a,\tau)\in\mathcal{F}_u], \tag*{(8)}

where Fu\mathcal{F}_u is the set of future (creator, engagement-type) pairs of user uu and the sum runs over engagement types τ\tau. We evaluate Direct Preference Optimization (DPO) Rafailov et al. [20], Group Relative Policy Optimization (GRPO) Shao et al. [24], and Self-Distillation Policy Optimization (SDPO) Hüb­otter et al. (2026) from the SFT model in Section C; none improves on it, so the served model is the Stage-2 SFT model.

Related Work

Generative recommendation. Generative recommendation [30, 21] has become a compelling alternative to multi-stage filtering pipelines. It builds on a long line of sequential recommenders that model user histories with recurrent, self-attentive, bidirectional, and self-supervised encoders [7, 14, 26, 35], and more recently, on large-scale generative transducers that recast ranking as sequence generation [31]. A foundational design choice is the item identifier, which must carry both semantic content and collaborative signal [21, 25]. The dominant recipe tokenizes each item into a Semantic ID by discretizing a pretrained embedding into a hierarchy of codes and training a language model to generate them [21, 8, 13, 34, 9]. MusicRec [33] and FusID [15] further fuse textual and acoustic signals into the embedding before quantization [36]. To make these continuous embeddings usable by a large language model (LLM), vector-quantization techniques [28, 22, 1] such as the residual quantized variational autoencoder (RQ-VAE) [17] and residual quantization with k-means (RQ-KMeans) [3] map each embedding to a tuple of codebook indices. All of these methods place a many-to-one quantizer, fixed while the language model trains, between the item and the language model; ACT removes it.

Preference optimization. Reinforcement learning (RL) is the standard tool for post-training alignment of LLMs [6, 27, 19, 12, 16]. Methods that learn from verifiable rewards, such as GRPO [24], optimize policy gradients on outcome-level rewards, but sparse scalar feedback complicates credit assignment, and process reward models [23] that supply step-level supervision require an auxiliary model. In recommendation, DPO [20] contrasts preferred and rejected items at the item level, which does not decompose over multi-token SIDs; self-distillation methods [32, 18, 29] (Hübotter et al. 2026) instead let the policy, conditioned on realized feedback, act as its own teacher. Because an ACT creator is a single token, all three apply to ACT directly over the true action space; we find, however, that none improves on SFT in our setting (Section C).

Experiments

We compare ACT with the SID pipeline on production unconnected-creator data and on the public KuaiRand dataset. We ask whether ACT improves creator-level recall (Section 4.2), why (Section 4.3), how vocabulary size trades conditional recall against coverage (Section 4.4), whether the result transfers to public data and when it holds (Section 4.5), what each token layer costs to serve (Section 4.6), which design choices matter (Section 4.7), and whether the offline gain carries over online (Section 4.8).

Experimental Setup

Production data. Next-creator examples come from engagements with suggested, non-followed creators (opens, likes, shares, follows, and dwell), with auxiliary cross-surface engagement as additional context and each history truncated to the most recent engaged creators. We split by day so that no test interaction is seen in training, and the target is the next unconnected creator on any surface. We evaluate three nested candidate sets by decreasing follower threshold: toptop (NA=YN_A=Y), midmid (NA=10YN_A=10Y), and broadbroad (NA=30YN_A=30Y), so a larger vocabulary adds coverage rather than replacing creators. Vocabulary and codebook sizes are reported in relative units: YY is the smallest vocabulary and XX the base per-level codebook size, so the SID codebooks are X3X^3, (2X)3(2X)^3, and (4X)3(4X)^3. The 10Y10Y setting trains on ∼10\sim10M engagement events and tests on ∼1\sim1M; the other settings use a test set of the same scale.

Public data. To test transfer, we use KuaiRand-27K [4]: the task is to predict the next clicked creator from a user’s 20 most recent creators. We restrict the candidate set to the top-50k creators by training clicks (90.4% of training clicks), giving 27,065 users, 16.52M training and 100k test examples. Both token schemes use identical samples in identical order and differ only in the creator token; both are trained with Stage 2 only (no warm-up) for one epoch at learning rate 10−410^{-4}. We add two heuristic baselines, global popularity (the most-clicked creators in training) and recent history (the user’s most recent creators). To probe how much data ACT needs, we also train ACT on the smaller KuaiRand-1K, both on all of its creators and on its top creators.

SID baseline. The SID baseline follows OneRec [2]: a frozen GNN embedding quantized by RQ-KMeans (L=3L=3, M=XM=X unless stated) on the same creators, generated by the same backbone, and decoded by codebook join and popularity rerank. Only the token layer differs from ACT.

Metrics. Because the served unit is a creator, we score both methods by creator-level Recall@KK (K∈{1,5,10,15,50,200}K \in\{1, 5, 10, 15, 50, 200\}), i.e., whether the true creator is in the top-KK, and by normalized discounted cumulative gain (NDCG@200) weighted by the engagement value v(τ)v(\tau). For ACT a token is a creator. For SID we expand each predicted SID into its member creators, ranked by beam order and then by follower count within a SID. We also report SID cluster-level recall, which counts a hit whenever the predicted SID equals the target’s SID; it overstates delivered recall because one SID stands for ν(s)\nu(s) creators, so only creator-level recall is comparable to ACT.

Implementation. Stage 1 trains the creator rows for one epoch at learning rate 10−410^{-4}; Stage 2 applies LoRA (rank r=64r=64, scaling α=128\alpha=128, dropout 0.05) with the creator rows trainable, at learning rate 2×10−52 \times10^{-5} with a cosine schedule. Training runs in bf16 with fully sharded data parallelism (FSDP) on 4 hosts ×\times 8 graphics processing units (GPUs); inference uses vLLM over the expanded vocabulary with beam width 15 for K≤15K \le15 and 200 for the deeper cutoffs.

Main Results

ACT outperforms the SID baseline on every reported production metric at the 10Y scale with the base X3X^3 codebook. Creator-level Recall@15 rises from 23.58% to 43.41%, Recall@50 from 34.33% to 60.27%, and Recall@200 from 47.62% to 79.30%. The corresponding absolute gains are 19.83, 25.94, and 31.68 percentage points: ACT’s lead is present for short candidate lists and remains substantial for deeper ones. NDCG@200 also increases from 0.187 to 0.343, a relative gain of 83.4%, indicating that the recall improvement is accompanied by better performance under the rank-sensitive, engagement-weighted metric. The advantage holds across the vocabulary and codebook settings in Figure 2, where ACT leads at every plotted cutoff (K∈{1,5,10,15}K \in\{1, 5, 10, 15\}) in all four production settings. With the SID codebook fixed at X3X^3, ACT achieves Recall@15 of 59.80% versus 40.81% for SID at YY, and 43.41% versus 23.58% at 10Y (Table 1). Both methods have lower conditional recall in the larger vocabulary, but the absolute advantage of ACT stays close to 19–20 points. Its relative gain consequently widens from 46.5% to 84.1%. Enlarging the SID codebook improves the baseline but does not close this gap: at 10Y, increasing the per-level size from XX to 2X2X expands the nominal code space eightfold and raises SID Recall@15 from 23.58% to 29.21%. ACT still reaches 43.41%, retaining a 14.20-point advantage. At the broadest 30Y tier, even the (4X)3(4X)^3 codebook, with 64 times the nominal capacity of X3X^3, yields 30.15% Recall@15 against ACT’s 38.87%. ACT also leads at every smaller plotted cutoff in this setting, so its advantage persists beyond the smallest, most popular creator tier and against codebooks with substantially more capacity. § 4.3 examines the collision losses behind this pattern, and Section A details the broadest-scale comparison and its differing embedding coverage. All these recalls are scored at the creator level, after expanding SID predictions into ranked creators, and thus measure retrieval of the actual target rather than its semantic cluster. Since recall is conditional on the target belonging to each vocabulary, the lower values at larger tiers do not by themselves establish which catalog size is best for all engagement; § 4.4 evaluates that tradeoff by combining conditional recall with engagement coverage.

Figure 2. Creator-level Recall@KK of ACT and SID on production data at each vocabulary scale and SID codebook size. ACT leads at every KK in every setting.

Why Atomic Creator Tokens Beat Semantic IDs

SID predictions must identify a creator after identifying its semantic code, and collisions introduce a loss at this second step. Creator-level recall factorizes as

RcreatorSID=RclusterSID×ρ,ρ∈(0,1],(9)R^{\mathrm{SID}}_{\mathrm{creator}} = R^{\mathrm{SID}}_{\mathrm{cluster}} \times\rho,\qquad\rho\in(0,1], \tag*{(9)}

Here, RcreatorSIDR^{\mathrm{SID}}_{\mathrm{creator}} is SID’s creator-level Recall@15, which counts a hit only when the true creator appears in the final top-15 list. RclusterSIDR^{\mathrm{SID}}_{\mathrm{cluster}} is its cluster-level Recall@15, which counts a hit when the target’s SID is retrieved. ρ=RcreatorSID/RclusterSID\rho= R^{\mathrm{SID}}_{\mathrm{creator}}/R^{\mathrm{SID}}_{\mathrm{cluster}} is the resolution efficiency: the fraction of cluster-level hits retained after expanding the predicted SIDs into creators and applying the popularity tie-break. A value of ρ=1\rho= 1 means that this resolution step loses no hits, while a smaller value indicates a larger loss.

Whenever ν>1\nu> 1 and the tie-break is not user-specific, ρ<1\rho< 1, roughly ρ≲1/νˉ\rho\lesssim1/\bar{\nu} for mean collision count νˉ\bar{\nu}; ACT has ν≡1\nu\equiv1 and hence ρ≡1\rho\equiv1. Thus, even with identical cluster-level accuracy, SID loses a (1−ρ)(1-\rho) fraction of its hits, the collision tax. This is also why SID cluster-level recall can exceed ACT’s creator-level recall while SID is the weaker system: hitting the code does not identify the creator. The measured recalls in Table 1 show that ρ\rho falls from YY to 10Y10Y as νˉ\bar{\nu} rises under a fixed X3X^3 codebook. SID is therefore doubly penalized at scale: the code is harder to predict and the tax is heavier, whereas ACT pays only the first cost, which helps explain why its relative advantage widens with the catalog. Increasing the codebook size partly recovers these lost hits. Holding the creators, data, and decoding fixed and varying only the per-level size, (2X)3(2X)^3 at 10Y10Y substantially raises ρ\rho and SID recall, supporting collisions as an important source of the gap, yet ACT still leads (Figure 3). Pushed to 30Y30Y creators with a 64×64\times larger (4X)3(4X)^3 codebook, collisions fall monotonically with codebook size but never reach zero, because the frozen GNN embeddings cluster intrinsically (Section A). ACT keeps a clear lead even though SID is scored only on the creators that have a GNN embedding, while ACT is scored on all of them. Enlarging the codebook buys diminishing returns; only ν≡1\nu\equiv1 removes the tax outright. The learned representations provide a complementary view of the remaining gap. The data processing inequality discussed in § 2.3 relates SID’s fixed quantization to information loss, while ACT learns its creator embeddings directly under the sequential objective. Consistent with this distinction, the learned creator embeddings are largely decoupled from the frozen GNN prior: global, local, and layout-level agreement measures indicate little shared structure (Section D). ACT thus learns a task-specific geometry rather than simply re-deriving the collaborative geometry encoded by the SID quantizer.

VocabularySID codebookνˉ\bar{\nu}ρ\rhoTaxSIDACTΔ\Delta
YY (top)X3X^32.350.6125.940.8159.80+18.99
10Y10Y (mid)X3X^33.790.4726.823.5843.41+19.83
10Y10Y (mid)(2X)3(2X)^31.750.6416.729.2143.41+14.20
30Y30Y (broad)(4X)3(4X)^31.370.7410.430.1538.87+8.72

Table 1. Recall@15 and collision statistics as the vocabulary and the SID codebook grow. νˉ\bar{\nu}: mean creators per SID; ρ=Rcreator/Rcluster\rho= R_{\mathrm{creator}}/R_{\mathrm{cluster}} for SID; tax =Rcluster−Rcreator= R_{\mathrm{cluster}} - R_{\mathrm{creator}} (points); Δ\Delta: gain of ACT over SID. ACT has νˉ=ρ=1\bar{\nu} = \rho= 1 and no tax at every scale.

Figure 3. Creator-level Recall@15 of SID and ACT per setting, with the fraction of creators that share their SID. A larger codebook reduces collisions and narrows the gap, but the gap never closes.

Vocabulary Size and Coverage

A larger vocabulary makes each prediction harder but covers more of what users engage with. As the ACT vocabulary grows from YY to 30Y30Y, recall conditional on the target being in the vocabulary falls, while the share of engagement the vocabulary covers rises faster, so the effective recall over all engagement increases at every step (Figure 4). Within the head regime, a larger ACT vocabulary is therefore strictly better in delivered recall, which motivates scaling ACT up to the point where per-token data runs out (§ 4.5).

Figure 4. ACT across vocabulary tiers: conditional Recall@15 falls as the vocabulary grows, but engagement coverage cc rises faster, so effective Recall@15 (c×c\timesconditional) rises monotonically.

Public Data: KuaiRand

ACT’s advantage over SID also appears on KuaiRand-27K, where both models use identical training and test samples, the same sample order, and one epoch of Stage 2 training without vocabulary warm-up. ACT leads SID at every evaluated cutoff from K=5K=5 (Figure 5): Recall@5 is 1.138% versus 0.903%, and Recall@15 is 2.763% versus 1.859%, corresponding to relative gains of 26.0% and 48.6%. The absolute gap also grows from 0.235 to 0.904 percentage points, so the benefit extends beyond selecting a single creator. The two heuristic baselines provide a useful check on what the models learn. Global popularity is competitive on this task, reaching 2.351% Recall@15, while recent history reaches 1.982%. ACT exceeds both and is the only method above global popularity at every evaluated cutoff; its relative lift over popularity is 33.7% at K=5K=5, 21.8% at K=10K=10, and 17.5% at K=15K=15. SID, in contrast, falls below global popularity from K=10K=10 and also below recent history at K=15K=15. Its only advantage over ACT is at K=1K=1, where recall is 0.332% versus 0.294%. This small top-1 lead is consistent with the popularity prior introduced by returning the most popular creator within each predicted SID, but it does not persist as the candidate list grows. The public-data codebook has 1.30 creators per SID and a 36.2% colliding fraction, close to the collision profile of the production (4X4X) codebook. The result therefore holds with a substantially less colliding quantizer than the base production codebook. The learned ACT vocabulary is also used broadly: top-1 predictions span 12,260 distinct creators, compared with 4,399 distinct codes predicted by SID on the same test set. However, the smaller KuaiRand-1K experiments show that atomic tokens do not automatically outperform simple retrieval heuristics. With all creators included, ACT achieves 0.75% Recall@15 against 1.76% for the best heuristic; restricting the vocabulary to top creators raises ACT recall to 1.86%, still below the heuristic’s 3.78%. These settings provide about 5 and 34 training samples per creator token on average, compared with about 331 in the KuaiRand-27K top-50k setting. The larger dataset also supplies more total training data, so this comparison does not isolate a universal samples-per-token threshold. It does show that the public-data gain depends on the available supervision, rather than following from the one-token representation alone. Together, these results support the routed design of § 1: use ACT for well-observed creators in the head and conventional retrieval where creator-level training data remains sparse.

(a) Creator-level Recall@K.

Relative lift over global popularity.

(b) Relative lift over global popularity.

Figure 5. KuaiRand-27K (top-50k creators), with identical samples for ACT and SID. ACT is the only method that beats global popularity at every cutoff; SID falls below it from K=10K=10. The SID codebook has 1.30 creators per SID and a 36.2% colliding fraction. Top-1 predictions cover 12,260 distinct creators for ACT versus 4,399 distinct SIDs for SID.

Inference Cost

ACT is also cheaper to serve at every scale (Figure 6). Under the same hardware (32×A100-80GB), beam width 15, and ∼1M-user test set, ACT needs less GPU memory per card, serves more concurrent requests at a fixed memory budget, and has lower per-user latency. The driver is one token versus five: SID emits a five-token block (⟨bos⟩s1s2s3⟨eos⟩\langle\mathrm{bos}\rangle s_1s_2s_3 \langle\mathrm{eos}\rangle), which multiplies the decode steps and the key–value (KV) cache held by every beam, whereas the only extra cost of ACT is a wider softmax over the expanded vocabulary, a single bandwidth-friendly matrix multiplication. The token-layer choice is therefore a Pareto improvement in both accuracy and serving cost.

Inference cost of ACT and SID across vocabulary tiers.

Figure 6. Inference cost of ACT and SID across vocabulary tiers (32×A100-80GB, beam width 15, ∼1M-user test). Throughput is concurrent requests at fixed memory, relative to SID.

Ablation Study

GNN warm-starting yields modest recall gains over random initialization (Table 2). Adding the in-batch contrastive auxiliary loss to the warm-started variant reduces recall at every reported cutoff. We use the simpler model with randomly initialized, trainable creator embeddings for serving. Although ACT makes on-policy alignment feasible, none of DPO, GRPO, or SDPO improves on the SFT model (Section C), which keeps reinforcement learning off the serving critical path.

Table 2 Ablations of ACT design choices (SFT, 10Y10Y; creator-level Recall@KK with beam width 200; best per column in bold).

VariantR@15R@50R@200
Random init, trainable43.9160.9680.01
GNN warm-start, trainable44.6661.6280.44
+ contrastive loss43.3959.5477.22

Table 2.

Online Test

We evaluate ACT in a live four-arm A/B test against the SID-enhanced production arm (prod+SID) and the classic production ranker (prod). The ACT arm has a positive directional lift for every reported metric against both controls (Table 3). The largest change is in product engagement: +0.29% over prod+SID and +0.66% over prod, with the latter approaching statistical significance (p≈0.05p \approx0.05). The other changes are small and are best interpreted as directional evidence that the offline recall gain transfers online.

Table 3 Online relative change (%) of ACT against the SID-enhanced production arm (prod+SID) and the classic production ranker (prod). The product-engagement change against prod approaches statistical significance (p≈0.05p \approx0.05).

Metricvs. prod+SIDvs. prod
Overall consumption metric 1+0.11+0.05
Overall consumption metric 2+0.01+0.01
Product consumption metric+0.02+0.03
Product engagement metric+0.29+0.66

Table 3.

Conclusion

In cold-start unconnected creator recommendation, the binding constraint is the SID layer, not the language model: a surjective creator-to-SID map imposes a collision-limited recall ceiling, and a frozen quantizer blocks end-to-end learning. ACT removes this layer with a bijective per-creator vocabulary trained end-to-end. Under a controlled comparison, ACT improves creator-level Recall@15 by +19.8 points (1.84×1.84 \times) over a matched SID baseline, keeps its lead as the catalog and the codebook grow, transfers to the public KuaiRand dataset, and is cheaper to serve. Its premise is that every creator token sees enough training data, so ACT is best deployed as the head of a routed system, with the long tail served by conventional retrieval. Multi-interest decoding, hierarchical tokens for latency-critical settings, and on-policy alignment for objectives that SFT cannot express are natural next steps.

Largest-Scale Codebook Stress Test

Table 4 details the stress test of § 4.3: a 30YY-creator catalog with the SID codebook grown from X3X^3 to (4X)3(4X)^3. Collisions shrink monotonically with codebook size but never vanish: even the largest codebook leaves a large share of creators sharing a code, because the frozen GNN embeddings cluster intrinsically. Against the strongest SID, ACT still leads at every cutoff, and the comparison favors SID, which is scored only on the ∼76%\sim76\% of the 30YY creators that have a GNN embedding, whereas ACT is scored on all of them.

CodebookReachable codesUsed / reachableνˉ\bar{\nu}Colliding creators
SID X3X^3X3X^388.7%8.0698.4%
SID (2X)3(2X)^3(2X)3(2X)^343.5%2.0673.2%
SID (4X)3(4X)^3(4X)3(4X)^38.2%1.3743.3%
ACT (bijective)––1.000%
ModelR@1R@5R@10R@15
SID (4X)3(4X)^312.5721.8026.8330.15
ACT15.1027.3934.2738.87

Table 4. Codebook stress test (30YY catalog, L=3L=3). Top: collisions against codebook size. Bottom: creator-level Recall@K of the largest SID codebook and ACT (beam width 15, ∼1\sim1M test examples). Collisions against codebook size. Creator-level Recall@K (%): largest SID codebook against ACT.

Non-Generative Retrieval Baselines

The main text compares ACT with a matched SID model, the representation the generative recipe directly replaces. For completeness, we also place ACT against the non-generative retrieval it would sit next to in a routed system: the deployed creator-to-creator (C2C) retrieval stack. We report two of its live sources (A and B, identities masked) and the reciprocal-rank-fusion (RRF) ensemble C that unions them and is the one actually served. All methods are scored on the same ∼1\sim1M-user test set by creator-level Recall@15, conditional on the target falling inside the vocabulary at each scale; the generative models decode with beam width 15, and the retrieval sources return their native ranked lists.

Table 5 shows a clean ladder at every scale. The C2C sources retrieve few in-vocabulary candidates for these unconnected creators, so their recall is coverage-limited by construction. They are general-purpose retrievers tuned for the broad catalog rather than the popularity-scoped head that ACT targets, so the comparison is directional rather than matched. Still, the matched SID model clearly outperforms the best retriever, and ACT leads throughout, which supports reserving conventional retrieval for the long tail.

MethodY (top)10Y (mid)30Y (broad)
C2C source A6.705.585.04
C2C source B4.125.305.68
C2C ensemble C9.459.539.41
SID (creator-level)40.8123.5830.15
ACT59.8043.4138.87

Table 5. ACT against non-generative retrieval baselines: creator-level Recall@15 (%) on the in-vocabulary population, ∼1M test users; best per column in bold.

Post-Training

Because ACT’s creator embeddings are policy parameters, the policy πθ(a∣z)\pi_\theta(a \mid z) is differentiable over the true action space, so we ask whether on-policy alignment also helps. We apply three alignment methods to a shared SFT reference policy πref=πθSFT\pi_{\mathrm{ref}} = \pi_{\theta_{\mathrm{SFT}}}, with the value-weighted reward r(a)r(a) of § 2.4.

DPO (Rafailov et al., [20]) uses item-level pairs of a preferred creator a+a^+ and a rejected creator a−a^- and minimizes

LDPO=−E(z,a+,a−)[log⁡σ(βlog⁡πθ(a+∣z)πref(a+∣z)−βlog⁡πθ(a−∣z)πref(a−∣z))](10)\mathcal{L}_{\mathrm{DPO}} = -\mathbb{E}_{(z,a^+,a^-)}\left[\log\sigma\left(\beta\log\frac{\pi_\theta(a^+ \mid z)}{\pi_{\mathrm{ref}}(a^+ \mid z)} - \beta\log\frac{\pi_\theta(a^- \mid z)}{\pi_{\mathrm{ref}}(a^- \mid z)}\right)\right] \tag*{(10)}

where σ\sigma is the logistic function, β>0\beta> 0 controls the strength of the implicit KL constraint to πref\pi_{\mathrm{ref}}, and the expectation is over contexts with a preference pair.

GRPO ([24]) samples a group of GG on-policy creators {ai}i=1G∼πθold(⋅∣z)\{a_i\}_{i=1}^{G} \sim\pi_{\theta_{\mathrm{old}}}(\cdot\mid z) from the previous policy πθold\pi_{\theta_{\mathrm{old}}}, scores each with reward ri=r(ai)r_i = r(a_i), and forms group-normalized advantages A^i=(ri−mean⁡(r))/std⁡(r)\hat{A}_i = (r_i - \operatorname{mean}(\mathbf{r}))/\operatorname{std}(\mathbf{r}) over the group rewards r=(r1,…,rG)\mathbf{r} = (r_1,\ldots,r_G), updating θ\theta with the clipped policy-gradient surrogate and a KL penalty to πref\pi_{\mathrm{ref}}. Because an ACT creator is a single token, the diversity of a group is exactly the entropy of πθ(⋅∣z)\pi_\theta(\cdot\mid z): a low-entropy SFT model yields near-duplicate rollouts, std⁡(r)→0\operatorname{std}(\mathbf{r}) \to0, and the advantages, hence the gradient, vanish.

SDPO ([32, 18, 29] (Hübotter et al. 2026)) needs no external reward model: it conditions the policy on its own realized engagement feedback ff to form a self-teacher πθ(⋅∣z,f)\pi_\theta(\cdot\mid z,f) that supplies dense targets, and trains the student πθ(⋅∣z)\pi_\theta(\cdot\mid z) to match it. Its signal is proportional to the teacher–student gap, which collapses when conditioning on ff barely changes the distribution.

Results. No recipe beats SFT (Table 6). DPO is compared against the main SFT model; GRPO and SDPO run through the on-policy evaluation flow and are compared against its own reference. The best DPO run only matches SFT, and GRPO and SDPO fall below their reference. The cause is structural rather than a tuning artifact: a single-token, low-entropy policy over a 10Y-way vocabulary yields near-duplicate rollouts and a sparse, low-variance reward, whereas ACT already casts recommendation as dense, exactly labeled token prediction that SFT optimizes directly. For a production system this outcome is welcome: reinforcement learning need not sit on the serving critical path.

MethodR@1R@5R@10R@15Δ@15
SFT (reference)17.0430.5738.2943.41–
+ DPO (best)17.0730.4938.2043.26−0.15
RL-flow reference13.8823.8329.5633.47–
+ GRPO (best)14.4423.3728.6132.20−1.27
+ SDPO (best)11.1918.9823.5926.71−6.77

Table 6. Post-training the ACT SFT model with DPO, GRPO, and SDPO (10Y; creator-level Recall@KK in %, beam width 15; best configuration per method). Δ\Delta@15 is relative to the applicable reference.

Representation Analysis: ACT against the GNN Prior

Section 4.3 argued that ACT learns a task-specific creator geometry rather than re-deriving the frozen GNN prior. We test this by comparing the 3072-dimensional creator-token rows of the 10Y model with the 128-dimensional production GNN embeddings of the same 10,000 creators, using independent t-SNE projections (Figure 7) and three measures of structural agreement (Table 7).

Independent t-SNE layouts of the ACT creator embeddings and the production GNN prior

Figure 7. Independent t-SNE of the ACT creator embeddings (3072-d) and the production GNN prior (128-d) for the same 10,000 creators. The similar look is a t-SNE artifact; Table 7 shows the two spaces are largely decoupled.

MeasureValueReading
Linear CKA0.10near-zero shared global structure (1 = identical)
kNN@10 overlap0.07∼70× the random baseline of 0.001; 93% of neighbors change
Procrustes disparity (t-SNE)0.887layouts nearly uncorrelated (1 = uncorrelated)

Table 7. Structural agreement between the ACT creator embeddings and the frozen GNN prior (nn=10,000 creators).

We use three complementary measures because no single one is decisive. Linear centered kernel alignment (CKA) captures global alignment and is invariant to orthogonal transforms and isotropic scaling, so the dimensionality gap does not bias it. The kk-nearest-neighbor overlap (kNN@10) probes local structure: the fraction of each creator’s ten nearest GNN neighbors that stay among its ten nearest in ACT space. Procrustes disparity, computed on the two t-SNE layouts after the optimal rotation, reflection, and scaling, measures how far the layouts diverge once that freedom is removed. All three would sit near their “identical” end if ACT merely re-derived the GNN prior.

Instead, all three show that the two representations are largely decoupled (Table 7). Global agreement is close to zero; local overlap is well above chance, so a thin residue of collaborative locality survives, yet most of each creator’s GNN neighborhood is reorganized; and after optimal alignment the same creator lands far apart in the two layouts. The two t-SNE panels look superficially similar, a known t-SNE artifact; the quantitative measures, not the pictures, carry the conclusion. The cause is architectural: the creator rows are randomly initialized and learn their own geometry under the sequential objective, whereas an SID model is confined to the geometry the frozen quantizer already encodes.

References

  1. [1]Alexandre Défossez, Jade Copet, Gabriel Synnaeve, and Yossi Adi. High fidelity neural audio compression. Transactions on Machine Learning Research (TMLR), 2023.arxiv.org/abs/2210.13438
  2. [2]Jiaxin Deng, Shiyao Wang, Kuo Cai, Lejian Ren, Qigen Hu, Weifeng Ding, Qiang Luo, and Guorui Zhou. Onerec: Unifying retrieval and rank with generative recommender and iterative preference alignment. arXiv preprint arXiv:2502.18965, 2025.arxiv.org/abs/2502.18965
  3. [3]Matthijs Douze, Alexandr Guzhva, Chengqi Deng, Jeff Johnson, Gergely Szilvasy, Pierre-Emmanuel Mazaré, Maria Lomeli, Lucas Hosseni, and Hervé Jégou. The faiss library. IEEE Transactions on Big Data, 2025.
  4. [4]Chongming Gao, Shijun Li, Yuan Zhang, Jiawei Chen, Biao Li, Wenqiang Lei, Peng Jiang, and Xiangnan He. Kuairand: An unbiased sequential recommendation dataset with randomly exposed videos. In Proceedings of the 31st ACM International Conference on Information and Knowledge Management (CIKM), pages 3953–3957, 2022.DOI
  5. [5]Aaron Grattafiori, Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Alex Vaughan, et al. The llama 3 herd of models. arXiv preprint arXiv:2407.21783, 2024.arxiv.org/abs/2407.21783
  6. [6]Daya Guo, Dejian Yang, Haowei Zhang, Junxiao Song, Peiyi Wang, Qihao Zhu, Runxin Xu, Ruoyu Zhang, Shirong Ma, Xiao Bi, et al. Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning. arXiv preprint arXiv:2501.12948, 2025.DOI
  7. [7]Balázs Hidasi, Alexandros Karatzoglou, Linas Baltrunas, and Domonkos Tikk. Session-based recommendations with recurrent neural networks. In International Conference on Learning Representations (ICLR), 2016.arxiv.org/abs/1511.06939
  8. [8]Yupeng Hou, Zhankui He, Julian McAuley, and Wayne Xin Zhao. Learning vector-quantized item representation for transferable sequential recommenders. In Proceedings of the ACM Web Conference (WWW), 2023.arxiv.org/abs/2210.12316
  9. [9]Yupeng Hou, Jiacheng Li, Ashley Shin, Jinsung Jeon, Abhishek Santhanam, Wei Shao, Kaveh Hassani, Ning Yao, and Julian McAuley. Generating long semantic ids in parallel for recommendation. In Proceedings of the 31st ACM SIGKDD Conference on Knowledge Discovery and Data Mining (KDD), 2025.arxiv.org/abs/2506.05781
  10. [10]Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Liang Wang, Weizhu Chen, et al. Lora: Low-rank adaptation of large language models. Iclr, 1(2):3, 2022.arxiv.org/abs/2106.09685
  11. [12]Aaron Jaech, Adam Kalai, Adam Lerer, Adam Richardson, Ahmed El-Kishky, Aiden Low, Alec Helyar, Aleksander Madry, Alex Beutel, Alex Carney, et al. Openai o1 system card. arXiv preprint arXiv:2412.16720, 2024.arxiv.org/abs/2412.16720
  12. [13]Bowen Jin, Hansi Zeng, Guoyin Wang, Xiusi Chen, Tianxin Wei, Ruirui Li, Zhengyang Wang, Zheng Li, Yang Li, Hanqing Lu, Suhang Wang, Jiawei Han, and Xianfeng Chen. Language models as semantic indexers. In Proceedings of the 41st International Conference on Machine Learning (ICML), 2024.arxiv.org/abs/2310.07815
  13. [14]Wang-Cheng Kang and Julian McAuley. Self-attentive sequential recommendation. In IEEE International Conference on Data Mining (ICDM), 2018.
  14. [15]Haven Kim, Yupeng Hou, and Julian McAuley. Fusid: Modality-fused semantic ids for generative music recommendation. arXiv preprint, 2026.arxiv.org/abs/2601.08764
  15. [16]Nathan Lambert, Jacob Morrison, Valentina Pyatkin, Shengyi Huang, Hamish Ivison, Faeze Brahman, Lester James V Miranda, Alisa Liu, Nouha Dziri, Shane Lyu, et al. Tulu 3: Pushing frontiers in open language model post-training. arXiv preprint arXiv:2411.15124, 2024.arxiv.org/abs/2411.15124
  16. [17]Doyup Lee, Chiheon Kim, Saeh oon Kim, Minsu Cho, and Wook-Shin Han. Autoregressive image generation using residual quantization. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11523–11532, 2022.arxiv.org/abs/2203.01941
  17. [18]Yaxuan Li, Yuxin Luo, Bingxiang He, Jinqian Zhang, Chaojun Xiao, Cheng Qian, Tianyu Yu, Huan-ang Gao, Wenkai Yang, Zhiyuan Liu, et al. Rethinking on-policy distillation of large language models: Phenomenology, mechanism, and recipe. arXiv preprint arXiv:2604.13016, 2026.arxiv.org/abs/2604.13016
  18. [19]Team Olmo, Allyson Ettinger, Amanda Bertsch, Bailey Kuehl, David Graham, David Heineman, Dirk Groeneveld, Faeze Brahman, Finbarr Timbers, Hamish Ivison, et al. Olmo 3. arXiv preprint arXiv:2512.13961, 2025.
  19. [20]Rafael Rafailov, Archit Sharma, Eric Mitchell, Christopher D Manning, Stefano Ermon, and Chelsea Finn. Direct preference optimization: Your language model is secretly a reward model. Advances in neural information processing systems, 36:53728–53741, 2023.arxiv.org/abs/2305.18290
  20. [21]Shashank Rajput, Nikhil Mehta, Anima Singh, et al. Recommender systems with generative retrieval. arXiv preprint arXiv:2305.05065, 2023.arxiv.org/abs/2305.05065
  21. [22]Ali Razavi, Aäron van den Oord, and Oriol Vinyals. Generating diverse high-fidelity images with vq-vae-2. In Advances in Neural Information Processing Systems (NeurIPS), 2019.
  22. [23]Amrith Setlur, Chirag Nagpal, Adam Fisch, Xinyang Geng, Jacob Eisenstein, Rishabh Agarwal, Alekh Agarwal, Jonathan Berant, and Aviral Kumar. Rewarding progress: Scaling automated process verifiers for llm reasoning. arXiv preprint arXiv:2410.08146, 2024.
  23. [24]Zhihong Shao, Peiyi Wang, Qihao Zhu, Runxin Xu, Junxiao Song, Xiao Bi, Haowei Zhang, Mingchuan Zhang, YK Li, Yang Wu, et al. Deepseekmath: Pushing the limits of mathematical reasoning in open language models. arXiv preprint arXiv:2402.03300, 2024.
  24. [25]Anima Singh, Trung Vu, Nikhil Mehta, Raghunandan Keshavan, Maheswaran Sathiamoorthy, Yilin Zheng, Lichan Hong, Lukasz Heldt, Li Wei, Devansh Tandon, Ed H. Chi, and Xinyang Yi. Better generalization with semantic ids: A case study in ranking for recommendations. In Proceedings of the 18th ACM Conference on Recommender Systems (RecSys), 2024.
  25. [26]Fei Sun, Jun Liu, Jian Wu, Changhua Pei, Xiao Lin, Wenwu Ou, and Peng Jiang. Bert4rec: Sequential recommendation with bidirectional encoder representations from transformers. In Proceedings of the 28th ACM International Conference on Information and Knowledge Management (CIKM), 2019.
  26. [27]Kimi Team, Angang Du, Bofei Gao, Bowei Xing, Changjiu Jiang, Cheng Chen, Cheng Li, Chenjun Xiao, Chenzhuang Du, Chonghua Liao, et al. Kimi k1. 5: Scaling reinforcement learning with llms. arXiv preprint arXiv:2501.12599, 2025.arxiv.org/abs/2501.12599
  27. [28]Aäron van den Oord, Oriol Vinyals, and Koray Kavukcuoglu. Neural discrete representation learning. In Advances in Neural Information Processing Systems (NeurIPS), 2017.arxiv.org/abs/1711.00937
  28. [29]Chenxu Yang, Chuanyu Qin, Qingyi Si, Minghui Chen, Naibin Gu, Dingyu Yao, Zheng Lin, Weiping Wang, Jiaqi Wang, and Nan Duan. Self-distilled rlvr. arXiv preprint arXiv:2604.03128, 2026.
  29. [30]Longfei Yun, Yihan Wu, Haoran Liu, Xiaoxuan Liu, Ziyun Xu, Yi Wang, Yang Xia, Pengfei Wang, Mingze Gao, Yunxiang Wang, et al. Decoding ml decision: An agentic reasoning framework for large-scale ranking system. arXiv preprint arXiv:2602.18640, 2026.
  30. [31]Jiaqi Zhai, Lucy Liao, Xing Liu, Yu Wang, Rui Li, Xuan Cao, Leon Gao, Zhaojie Gong, Fangda Gu, Jiayuan He, Yinghai Lu, and Yue Shi. Actions speak louder than words: Trillion-parameter sequential transducers for generative recommendations. In Proceedings of the 41st International Conference on Machine Learning (ICML), 2024.
  31. [32]Siyan Zhao, Zhihui Xie, Mengchen Liu, Jing Huang, Guan Pang, Feiyu Chen, and Aditya Grover. Self-distilled reasoner: On-policy self-distillation for large language models. arXiv preprint arXiv:2601.18734, 2026a.
  32. [33]Yuqiu Zhao, Lei Shi, Yan Zhong, Feifei Kou, Pengfei Zhang, Jiwei Zhang, Mingying Xu, and Yanchao Liu. Musicrec: Multi-modal semantic-enhanced identifier with collaborative signals for generative recommendation. In AAAI Conference on Artificial Intelligence, 2026b. https://api.semanticscholar.org/CorpusID:286700432.
  33. [34]Bowen Zheng, Yupeng Hou, Hongyu Lu, Yu Chen, Wayne Xin Zhao, Ming Chen, and Ji-Rong Wen. Adapting large language models by integrating collaborative semantics for recommendation. In IEEE 40th International Conference on Data Engineering (ICDE), 2024.
  34. [35]Kun Zhou, Hui Wang, Wayne Xin Zhao, Yutao Zhu, Sirui Wang, Fuzheng Zhang, Zhongyuan Wang, and Ji-Rong Wen. S3-rec: Self-supervised learning for sequential recommendation with mutual information maximization. In Proceedings of the 29th ACM International Conference on Information and Knowledge Management (CIKM), 2020.
  35. [36]Yiming Zhu, Xu Liu, Ziyun Xu, Zheng Wu, Joena Zhang, Sirius Chen, Chenheli Hua, Silvester Yao, Qichao Que, Wentao Shi, Junfeng Pan, and Linhong Zhu. A general framework for multimodal llm-based multimedia understanding in large-scale recommendation systems. In Proceedings of the 49th International ACM SIGIR Conference on Research and Development in Information Retrieval (SIGIR), Short Paper, 2026.

Paper details

vlnc:W_wkexnz4u
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 1

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Computer Science

Not recorded for this paper

View statement
Looking for related papers…
Contents