LIGE-GR: A Smooth Leap from Ranking to Generative Recommendation in the LLM Era

# LIGE-GR: A Smooth Leap from Ranking to Generative Recommendation in the LLM Era
Source: [https://arxiv.org/html/2609.18148](https://arxiv.org/html/2609.18148)
Chenzhang He⋆Affiliation:Meta Platforms, Inc\., Menlo Park, CA, USASam Woodmansee⋆Affiliation:Meta Platforms, Inc\., Menlo Park, CA, USAShawn Lian⋆Affiliation:Meta Platforms, Inc\., Menlo Park, CA, USAWenjie Hu⋆Affiliation:Meta Platforms, Inc\., Menlo Park, CA, USARenjie Jiang⋆Affiliation:Meta Platforms, Inc\., Menlo Park, CA, USAZiheng HuangAffiliation:Meta Platforms, Inc\., Menlo Park, CA, USAXinyuan ZhangAffiliation:Meta Platforms, Inc\., Menlo Park, CA, USAZhihao ZhengAffiliation:Meta Platforms, Inc\., Menlo Park, CA, USAZhuoran YuAffiliation:Meta Platforms, Inc\., Menlo Park, CA, USARui LiAffiliation:Meta Platforms, Inc\., Menlo Park, CA, USALei YuanAffiliation:Meta Platforms, Inc\., Menlo Park, CA, USAZiwei LiAffiliation:Meta Platforms, Inc\., Menlo Park, CA, USAJimmy JiaAffiliation:Meta Platforms, Inc\., Menlo Park, CA, USAMert TerzihanAffiliation:Meta Platforms, Inc\., Menlo Park, CA, USAEkrem KocaguneliAffiliation:Meta Platforms, Inc\., Menlo Park, CA, USAYiming LiaoAffiliation:Meta Platforms, Inc\., Menlo Park, CA, USAZhichen ZhaoAffiliation:Meta Platforms, Inc\., Menlo Park, CA, USAYue YinAffiliation:Meta Platforms, Inc\., Menlo Park, CA, USAYue WengAffiliation:Meta Platforms, Inc\., Menlo Park, CA, USAWanlin MaAffiliation:Meta Platforms, Inc\., Menlo Park, CA, USAXufeng CaiAffiliation:Meta Platforms, Inc\., Menlo Park, CA, USAWeimiao WuAffiliation:Meta Platforms, Inc\., Menlo Park, CA, USAYezhou HuangAffiliation:Meta Platforms, Inc\., Menlo Park, CA, USADu ZhangAffiliation:Meta Platforms, Inc\., Menlo Park, CA, USAYukun DingAffiliation:Meta Platforms, Inc\., Menlo Park, CA, USAAaron JohnstonAffiliation:Meta Platforms, Inc\., Menlo Park, CA, USAYueming WangAffiliation:Meta Platforms, Inc\., Menlo Park, CA, USAZhaojie GongAffiliation:Meta Platforms, Inc\., Menlo Park, CA, USAYuting ZhangAffiliation:Meta Platforms, Inc\., Menlo Park, CA, USASerena LiAffiliation:Meta Platforms, Inc\., Menlo Park, CA, USAAdithya GaneshAffiliation:Meta Platforms, Inc\., Menlo Park, CA, USABoying LiuAffiliation:Meta Platforms, Inc\., Menlo Park, CA, USAHaichuan YangAffiliation:Meta Platforms, Inc\., Menlo Park, CA, USAXialu LiAffiliation:Meta Platforms, Inc\., Menlo Park, CA, USAMatt MaAffiliation:Meta Platforms, Inc\., Menlo Park, CA, USAQunshu ZhangAffiliation:Meta Platforms, Inc\., Menlo Park, CA, USAJohn Joshua MillerAffiliation:Meta Platforms, Inc\., Menlo Park, CA, USAPraveen RathinaveluAffiliation:Meta Platforms, Inc\., Menlo Park, CA, USACheng HuangAffiliation:Meta Platforms, Inc\., Menlo Park, CA, USAAadhar SachdevaAffiliation:Meta Platforms, Inc\., Menlo Park, CA, USAJosh KarnsAffiliation:Meta Platforms, Inc\., Menlo Park, CA, USAAndres Aaron GutierrezAffiliation:Meta Platforms, Inc\., Menlo Park, CA, USANeil AgarwalAffiliation:Meta Platforms, Inc\., Menlo Park, CA, USAGustas PladisAffiliation:Meta Platforms, Inc\., Menlo Park, CA, USAVladimir BatyginAffiliation:Meta Platforms, Inc\., Menlo Park, CA, USAGopal RayAffiliation:Meta Platforms, Inc\., Menlo Park, CA, USAAditya PriyadarshiAffiliation:Meta Platforms, Inc\., Menlo Park, CA, USAShantanu PatilAffiliation:Meta Platforms, Inc\., Menlo Park, CA, USAZhe WangAffiliation:Meta Platforms, Inc\., Menlo Park, CA, USAPenny PanAffiliation:Meta Platforms, Inc\., Menlo Park, CA, USAYiping HanAffiliation:Meta Platforms, Inc\., Menlo Park, CA, USAArun SinghAffiliation:Meta Platforms, Inc\., Menlo Park, CA, USAGuangdeng LiaoAffiliation:Meta Platforms, Inc\., Menlo Park, CA, USABi XueAffiliation:Meta Platforms, Inc\., Menlo Park, CA, USAXinyao HuAffiliation:Meta Platforms, Inc\., Menlo Park, CA, USAYang SongAffiliation:Meta Platforms, Inc\., Menlo Park, CA, USAYisong SongAffiliation:Meta Platforms, Inc\., Menlo Park, CA, USAMeihong WangAffiliation:Meta Platforms, Inc\., Menlo Park, CA, USAHaotian WuAffiliation:Meta Platforms, Inc\., Menlo Park, CA, USADeepak AgarwalAffiliation:Meta Platforms, Inc\., Menlo Park, CA, USAJi LiuAffiliation:Meta Platforms, Inc\., Menlo Park, CA, USA

###### Abstract

The remarkable success of large language models \(LLMs\) has provided important inspiration for the next generation of recommender systems\. Structurally, recommendation and language generation share a similarity: both aim to produce an ordered sequence that optimizes the user’s experience\. However, how to precisely absorb the essence of the LLM paradigm into mature industrial recommender systems remains an open problem\.

There are two challenges\. First, it is unclear how to incorporate the LLM paradigm — sequence\-level generation and optimization — into recommendation\. Second, real\-world recommender systems are mature systems that have been iteratively customized for years around specific products, business constraints, serving infrastructure, and organizational ownership\. Replacing such systems wholesale is often technically risky and organizationally disruptive\.

In this paper, we proposeLIGE\-GR, a listwise generation and evaluation recommendation framework that upgrades from a traditional ranking system \(itemwise recommendation\) toward a generative recommendation paradigm\. Instead of rebuilding the entire recommendation stack from scratch,LIGE\-GRgeneralizes the existing pointwise recommendation system into a listwise generation system\. This allows mature recommender systems to benefit from listwise optimization while preserving compatibility with existing models, value functions, and serving infrastructure\.

We validateLIGE\-GRin short\-video recommendation on Instagram Reels and Facebook Video\. On these recommendation surfaces,LIGE\-GRimproves time spent by1\.14%1\.14\\%on Instagram Reels and0\.72%0\.72\\%on Facebook Video, while requiring only modest additional inference resources\.

††correspondence:Ji Liu at[madisonliu@meta\.com](mailto:[email protected]);⋆equal contribution\.## 1Introduction

Modern recommender systems return an ordered list of content items for each user request, which are then sequentially exposed to the user\. This formulation spans short\-video and feed recommendation\. Despite their sophistication, many industrial recommender systems are still fundamentally built around itemwise optimization\. A ranking model predicts user engagement signals for each candidate item independently, and a prespecified value model \(function\) maps these predicted signals into a scalar score\. Items are then sorted by their individual scores to form a recommendation list\. Additional product constraints, such as diversity or integrity adjustments, are usually introduced through heuristic or rule\-based score modifications\. This is the general framework that state\-of\-the\-art products employ today\. Although this paradigm has been successful, it has a fundamental limitation: it scores each item independently rather than evaluating the recommendation sequence jointly\.

![Refer to caption](https://arxiv.org/html/2609.18148v1/figures/pointwise_listwise.png)Figure 1:Itemwise vs\. listwise optimization\. Given the user’s interest history, itemwise optimization independently scores candidates and returns the top ones, whereas listwise optimization returns the best combination\.In contrast, LLMs fundamentally address the same problem – generating the best sequence conditioned on a user’s request—but follow a completely different technical paradigm\. Rather than independently selecting the best token at each position, an LLM generates each token conditioned on the previously generated context, with the quality of the final output determined by the sequence as a whole\. This contrast exposes the gap in paradigm we address between traditional itemwise and generative listwise recommendation:

Itemwise\(Independent Optimization\) vs\.Listwise\(Joint Optimization\)

![Refer to caption](https://arxiv.org/html/2609.18148v1/figures/framework.png)Figure 2:Itemwise versus listwise recommendation across the three componentsLIGE\-GRupgrades\. Ranking model: a context\-free rankerCFscores each item from the user alone, whereas the context\-aware modelCAalso conditions on the items already placed in the list \(red\)\. Objective \(defined by the value model\): from a sum of itemwise scores to a listwise value over the whole sequence\. Decoder: incumbent itemwise greedy selection commits to a single path through the position\-by\-candidate lattice, whereas the RL\-based decoder explores alternative paths and returns the best list found\.As Figure[1](https://arxiv.org/html/2609.18148#S1.F1)illustrates, itemwise recommendation scores items independently, whereas listwise recommendation evaluates them as a sequence\. Realizing this sequential optimization requires three capabilities: predicting each candidate’s value in the context of items already selected, evaluating the sequence as a whole, and searching for a feasible sequence under product and latency constraints\.

Motivated by the success of LLMs, a recent trend explores achieving these capabilities by rebuilding the recommender from scratch as a fully generative system, e\.g\., in[Deng et al\. \(2025\)](https://arxiv.org/html/2609.18148#bib.bib8)\. However, a successful paradigm to fully leverage listwise optimization for mature industrial applications has yet to emerge, leaving this an open problem\. Specifically, attempting a wholesale replacement of an existing stack faces two additional critical barriers:

- •System challenge\.Mature recommender systems encode years of model improvements, product logic, serving optimizations, and business constraints\. A replacement can be disadvantaged in early comparisons because it must first recover much of this accumulated baseline value, therefore fair evaluation often requires careful, extended validation\. Once deeply integrated, the new system can also raise rollback and reliability risks, because reverting may no longer mean disabling an isolated component\.
- •Organization challenge\.Recommendation, advertising, and search teams are typically organized around existing recommender\-system components, including retrieval, ranking, value modeling, serving infrastructure, and product policy\. Moving to a different paradigm can therefore disrupt not only the technical stack but also team boundaries, ownership, and long\-term planning\.

LIGE\-GRaddresses both the listwise technical challenges and the replacement challenge through anadditive,revertible, andlow\-resource\-requirementframework that generalizes the existing itemwise recommendation system\.LIGE\-GRpreserves the structure of the mature system while adding three components: a listwise module in the ranking model, an extension from itemwise to listwise value modeling, and an upgrade from the incumbent itemwise greedy decoder—which selects the highest\-scoring remaining candidate at each list position—to an RL\-based sequence decoder, as shown in Figure[2](https://arxiv.org/html/2609.18148#S1.F2)\.LIGE\-GRis thus not a replacement of traditional recommendation but a generalization and an upgrade\.

We validatedLIGE\-GRin short\-video recommendation on Instagram Reels and Facebook Video, both against strong, optimized baselines\. On Instagram Reels,LIGE\-GRincreases time spent by 1\.14%, while requiring additional inference resources equivalent to roughly 10% of those used by the context\-free ranking component and increasing end\-to\-end per\-request latency by approximately 7% relative to the incumbent baseline\. On Facebook Video,LIGE\-GRincreases time spent by 0\.72%\.

## 2LIGE\-GR: the Generative Paradigm

The section introduces the proposed generative paradigm\. We start with the problem definition and the reformulation of itemwise recommendation into the listwise or generative framework, which upgrades the existing itemwise system to the listwise system\. The section ends with the introduction to the listwise recommendation\.

### 2\.1Problem Statement

For each user request, the recommender system returns an ordered list of content items:

VT=\[v1,v2,…,vT\],vt∈𝒞\.\\displaystyle\{V\_\{T\}=\[v\_\{1\},v\_\{2\},\\ldots,v\_\{T\}\],\\qquad v\_\{t\}\\in\\mathcal\{C\}\.\}\(1\)Here𝒞\\mathcal\{C\}is the candidate set for the request, andVt=\[v1,…,vt\]V\_\{t\}=\[v\_\{1\},\\ldots,v\_\{t\}\]denotes the selected prefix—the partial list afterttpositions, withV0=∅V\_\{0\}=\\emptyset; the items in a feasibleVTV\_\{T\}are distinct\. In our setting,TTis around 10\. The items are displayed sequentially to the user\. The objective is to construct a personalized list that maximizes user experience and product quality\.

### 2\.2Reformulating Itemwise Recommendation into a Generative Framework

A traditional industrial recommender system is essentially an itemwise optimization system\. For each candidate item, a ranking model predicts a set of user engagement signals, such as:

plike,pfollow,pshare,pwatchtime ¿ 10s,…\\displaystyle p\_\{\\text\{like\}\},\\quad p\_\{\\text\{follow\}\},\\quad p\_\{\\text\{share\}\},\\quad p\_\{\\text\{watchtime \> 10s\}\},\\ldots\(2\)An itemwise value model \(itemVM\) then combines these predicted engagement signalsppinto a scalar item score\. A typical item\-level value model can be written as:

itemVM​\(p\)=w1⋅plike\+w2⋅pfollow\+w3⋅pshare\+⋯\.\\displaystyle\\textbf\{itemVM\}\(p\)=w\_\{1\}\\cdot p\_\{\\text\{like\}\}\+w\_\{2\}\\cdot p\_\{\\text\{follow\}\}\+w\_\{3\}\\cdot p\_\{\\text\{share\}\}\+\\cdots\.\(3\)The system then selects the topTTitems according to their item\-level VM scores and returns them as the final recommendation list\. In mature systems, the raw VM score is often further adjusted by diversity penalties, integrity rules, and other product constraints\. These adjustments introduce some sequence\-level awareness, but they are usually implemented as rule\-based heuristics rather than learned listwise optimization\. The itemwise recommendation objective can be abstracted as:

arg​maxVT∑t=1TitemVM​\(CF​\(u,vt\)\),\\displaystyle\\argmax\_\{V\_\{T\}\}\\quad\\sum\_\{t=1\}^\{T\}\\textbf\{itemVM\}\\left\(\\textbf\{CF\}\(u,v\_\{t\}\)\\right\),\(4\)whereCFdenotes a*context\-free*predictor \(the item\-wise ranking model prediction of various useruu’s engagement signals such asplikep\_\{\\text\{like\}\}for itemvtv\_\{t\}\) that scores each item independently given the user featureuuand the item featurevtv\_\{t\}\. Under this objective, the optimal list consists of the topTTitems among the candidate set ranked by theiritemVMscores in Eq\. \([3](https://arxiv.org/html/2609.18148#S2.E3)\)\.

As mentioned previously, many mature recommendation systems include an additional component—a control layer, although its name may vary—to enforce diversity in the recommended list\. Its purpose is to prevent similar content, such as items from the same category, from appearing too close together\. Popular approaches include gap demotion rule\([Gong et al\., 2021](https://arxiv.org/html/2609.18148#bib.bib12);[Pei et al\., 2019](https://arxiv.org/html/2609.18148#bib.bib30)\), determinant point process \(DPP\)\([Pan et al\., 2020](https://arxiv.org/html/2609.18148#bib.bib27);[Meng et al\., 2019](https://arxiv.org/html/2609.18148#bib.bib25);[Wang et al\., 2021a](https://arxiv.org/html/2609.18148#bib.bib42);[Li et al\., 2018](https://arxiv.org/html/2609.18148#bib.bib18)\), and hard\-coded business restrictions\. We denote the additive control\-layer adjustment for candidatevtv\_\{t\}given prefixVt−1V\_\{t\-1\}byCL​\(vt∣Vt−1\)∈ℝ∪\{−∞\}\\textbf\{CL\}\(v\_\{t\}\\mid V\_\{t\-1\}\)\\in\\mathbb\{R\}\\cup\\\{\-\\infty\\\}\. Finite values modify the candidate’s value\-model score, while−∞\-\\inftymasks an infeasible candidate\. The following examples illustrate these cases\.

- •\(hard business rule\) if the category ofvtv\_\{t\}and some category inVt−1V\_\{t\-1\}are not allowed to appear within the same list because of some hard restrictions, then CL​\(vt∣Vt−1\)=−∞;\\textbf\{CL\}\(v\_\{t\}\\mid V\_\{t\-1\}\)=\-\\infty;
- •\(example gap demotion rule\) if the closest item inVt−1V\_\{t\-1\}tovtv\_\{t\}belonging to the same category is at positiont′∈\{1,⋯,t−1\}t^\{\\prime\}\\in\\\{1,\\cdots,t\-1\\\}, then CL\(vt∣Vt−1\)=−exp\(t′−t\+1\)⋅constant;\\textbf\{CL\}\(v\_\{t\}\\mid V\_\{t\-1\}\)=\-\\exp\(t^\{\\prime\}\-t\+1\)\\cdot\\text\{constant\};
- •\(example DPP diversity score\) measure the incremental diversity ofvtv\_\{t\}on the top ofVt−1V\_\{t\-1\} CL\(vt∣Vt−1\)=logdet\(Φt⊤Φt\)−logdet\(Φt−1⊤Φt−1\),\\textbf\{CL\}\(v\_\{t\}\\mid V\_\{t\-1\}\)=\\log\\det\(\\Phi\_\{t\}^\{\\top\}\\Phi\_\{t\}\)\-\\log\\det\(\\Phi\_\{t\-1\}^\{\\top\}\\Phi\_\{t\-1\}\),whereΦt:=\[ϕt,ϕt−1,⋯,ϕ1\]\\Phi\_\{t\}:=\[\\phi\_\{t\},\\phi\_\{t\-1\},\\cdots,\\phi\_\{1\}\]andϕt′\\phi\_\{t^\{\\prime\}\}corresponds to the embedding ofvt′v\_\{t^\{\\prime\}\}normalized by‖ϕt′‖=1\\\|\\phi\_\{t^\{\\prime\}\}\\\|=1\.

Adding the control\-layer term to Eq\. \([4](https://arxiv.org/html/2609.18148#S2.E4)\) gives the overall objective:

arg​maxVT∑t=1T\[itemVM​\(CF​\(u,vt\)\)\+CL​\(vt∣Vt−1\)\]\.\\displaystyle\\argmax\_\{V\_\{T\}\}\\quad\\sum\_\{t=1\}^\{T\}\\left\[\\textbf\{itemVM\}\\left\(\\textbf\{CF\}\(u,v\_\{t\}\)\\right\)\+\\textbf\{CL\}\(v\_\{t\}\\mid V\_\{t\-1\}\)\\right\]\.\(5\)The incumbent itemwise system obtains the returned list with an itemwise greedy decoder over the candidate set𝒞\\mathcal\{C\}:

vt=arg​maxc∈𝒞\\Vt−1itemVM​\(CF​\(u,c\)\)\+CL​\(c∣Vt−1\),t=1,⋯,T\.\\displaystyle v\_\{t\}=\\argmax\_\{c\\in\\mathcal\{C\}\\backslash V\_\{t\-1\}\}\\quad\\textbf\{itemVM\}\\left\(\\textbf\{CF\}\(u,c\)\\right\)\+\\textbf\{CL\}\(c\\mid V\_\{t\-1\}\),\\;t=1,\\cdots,T\.

### 2\.3LIGE\-GR: Listwise Recommendation as Generative Recommendation

LIGE\-GRupgrades the itemwise recommendation system into a listwise generation system\. In its basic vanilla form,LIGE\-GRtargets the following sequence\-level objective using contextual model predictions:

arg​maxVT∑t=1T\[itemVM​\(CA​\(u,vt∣Vt−1\)\)\+CL​\(vt∣Vt−1\)\],\\displaystyle\\argmax\_\{V\_\{T\}\}\\quad\\sum\_\{t=1\}^\{T\}\\left\[\\textbf\{itemVM\}\\left\(\\textbf\{CA\}\(u,v\_\{t\}\\mid V\_\{t\-1\}\)\\right\)\+\\textbf\{CL\}\(v\_\{t\}\\mid V\_\{t\-1\}\)\\right\],\(6\)whereCAis a context\-aware predictor that estimates the engagement values of itemvtv\_\{t\}conditioned on the preceding itemsVt−1V\_\{t\-1\}within the same list\.

The overall upgrade from itemwise recommendation toLIGE\-GRcomprises three component upgrades:

- •Ranking model:from a context\-free predictor to a context\-aware predictor\.
- •Value model:from an itemwise VM to a listwise VM\.
- •Decoder:from incumbent itemwise greedy selection to RL\-based sequence decoding\.

Importantly,LIGE\-GRstrictly generalizes the incumbent itemwise recommender\. Reverting its three upgraded components to the context\-free predictor, itemwise VM\+CL evaluator, and itemwise greedy decoder recovers the incumbent system described above\. The corresponding decoder settings are given in Section[3\.3](https://arxiv.org/html/2609.18148#S3.SS3)\. This property matters in practice\. It meansLIGE\-GRcan be introduced as a smooth upgrade to the existing recommender system, rather than as a disruptive replacement\.

![Refer to caption](https://arxiv.org/html/2609.18148v1/figures/model.png)Figure 3:The listwise model\. The context\-free module \(left\) is the itemwise model: each candidatevtv\_\{t\}is scored together with the user featuresuuby an interaction network and per\-task heads\. Its intermediate representationsvt′v^\{\\prime\}\_\{t\}are handed to the lightweight context\-aware module \(right\), which refines all per\-task predictions with a four\-layer causal Transformer that attends only to preceding items\. Within each module, all task heads share the same network weights, applied at every item or position\.This summed objective already benefits from context\-aware prediction; Section[3\.2](https://arxiv.org/html/2609.18148#S3.SS2)upgrades it into a true listwise value model\. In its most general form,LIGE\-GRtargets

arg​maxVTListVM​\(\{CA​\(u,vt∣Vt−1\)\}t=1T\)\.\\displaystyle\\argmax\_\{V\_\{T\}\}\\quad\\textbf\{ListVM\}\\left\(\\left\\\{\\textbf\{CA\}\(u,v\_\{t\}\\mid V\_\{t\-1\}\)\\right\\\}\_\{t=1\}^\{T\}\\right\)\.\(7\)HereListVMmay use item metadata and includes the control\-layer adjustments applied to the selected sequence\.

## 3LIGE\-GRDesign

The section introduces the detailed design inLIGE\-GRin each upgraded component: the listwise model, the listwise VM, thePalettedecoder, and serving optimization \(see Appendix[4](https://arxiv.org/html/2609.18148#S4)\)\.

### 3\.1Listwise Model: Context\-Aware and Context\-Free Predictors

LIGE\-GRupgrades the existing itemwise ranking model into a listwise model\. It contains two components, as illustrated in Figure[3](https://arxiv.org/html/2609.18148#S2.F3):

- •A context\-free predictor, which corresponds to the existing item\-wise ranking model’s predictions\.
- •A context\-aware predictor, which refines predictions based on the previously selected items\.

The context\-free predictor can be written as:

\(u,vt\)→yt,∀t=1,⋯,T,\(u,v\_\{t\}\)\\rightarrow y\_\{t\},\\quad\\forall t=1,\\cdots,T,whereuurepresents user features,vtv\_\{t\}represents the candidate item at positiontt, andyty\_\{t\}denotes predicted engagement signals, such as like probability, follow probability, or watch\-time prediction\.

The logical form of the context\-aware predictor is:

\(u,vt∣v1,v2,…,vt−1\)→yt,∀t=1,⋯,T\.\(u,v\_\{t\}\\mid v\_\{1\},v\_\{2\},\\ldots,v\_\{t\-1\}\)\\rightarrow y\_\{t\},\\quad\\forall t=1,\\cdots,T\.That is, the prediction for itemvtv\_\{t\}is conditioned not only on the user and the item itself, but also on the previously selected sequence within the same requested list\. This allows the model to capture listwise effects such as repetition, saturation, complementarity, diversity, and user fatigue\. Because the context provides additional information, a well\-designed context\-aware predictor should provide more accurate predictions than a context\-free predictor\.

In practice, for efficiency and maintainability,LIGE\-GRdoes not rebuild the entire ranking model from scratch\. Instead, the context\-aware predictor is implemented as a lightweight refinement module on top of the existing context\-free model leveraging a GPT\-style decoder\-only causal transformer\. Specifically, letvt′v^\{\\prime\}\_\{t\}denote an intermediate representation produced by the context\-free model for itemvtv\_\{t\}\. The context\-aware module takes the sequence of intermediate representations as input \(similar to token embeddings in the LLM\):

\(vt′∣v1′,v2′,…,vt−1′\)→yt,∀t=1,⋯,T\.\(v^\{\\prime\}\_\{t\}\\mid v^\{\\prime\}\_\{1\},v^\{\\prime\}\_\{2\},\\ldots,v^\{\\prime\}\_\{t\-1\}\)\\rightarrow y\_\{t\},\\quad\\forall t=1,\\cdots,T\.
This design has two practical advantages:

- •In the evaluated Instagram Reels configuration, the additional model remains lightweight, requiring additional inference resources equivalent to roughly 10% of those used by theCFcomponent\.
- •The original context\-free model is fully preserved, which allows the system to reuse existing training and serving infrastructure with minimal disruption\.

In our implementation, the context\-aware module uses a lightweight causal GPT\-like decoder\-only transformer with four heads and four layers\. The architecture itself is not the primary focus of this paper; future work can further optimize model design and serving efficiency\.

### 3\.2Listwise VM: From Itemwise Value to Listwise Value

Traditional recommender systems define value at the item level\. The value model estimates the utility generated by showing a single item to a user\. However, user experience is inherently listwise\. The value of an item depends not only on the item itself, but also on its position and on the previously consumed items\.

A simple listwise VM can be constructed by summing the context\-aware item values across positions:

ListVMvanilla​\(VT\)=\\displaystyle\\textbf\{ListVM\}\_\{\\text\{vanilla\}\}\(V\_\{T\}\)=\\;∑t=1T\[itemVM​\(CA​\(u,vt∣Vt−1\)\)\+CL​\(vt∣Vt−1\)\]\.\\displaystyle\\sum\_\{t=1\}^\{T\}\\left\[\\,\\textbf\{itemVM\}\\left\(\\textbf\{CA\}\(u,v\_\{t\}\\mid V\_\{t\-1\}\)\\right\)\+\\textbf\{CL\}\(v\_\{t\}\\mid V\_\{t\-1\}\)\\,\\right\]\.\(8\)This formulation is simple and already incorporates sequential context through the context\-aware predictor\. However, it ignores a fundamental factor: not every item in the list is necessarily reached by the user\. Items at later positions should be weighted by the probability that the user continues watching until that position\.

Therefore,LIGE\-GRintroduces a more principled listwise VM based on continuation probability\. Letpcontinue​\(Vt−1\)p\_\{\\text\{continue\}\}\(V\_\{t\-1\}\)denote the probability that the user continues after consuming the prefixVt−1V\_\{t\-1\}and reaches itemvtv\_\{t\}\. The listwise VM is defined as follows:

ListVMgolden​\(VT\)=\\displaystyle\\textbf\{ListVM\}\_\{\\text\{golden\}\}\(V\_\{T\}\)=\\;∑t=1Tpcontinue​\(Vt−1\)⋅\[itemVM​\(CA​\(u,vt∣Vt−1\)\)\+CL​\(vt∣Vt−1\)\]\.\\displaystyle\\sum\_\{t=1\}^\{T\}p\_\{\\text\{continue\}\}\(V\_\{t\-1\}\)\\cdot\\left\[\\textbf\{itemVM\}\\left\(\\textbf\{CA\}\(u,v\_\{t\}\\mid V\_\{t\-1\}\)\\right\)\+\\textbf\{CL\}\(v\_\{t\}\\mid V\_\{t\-1\}\)\\right\]\.\(9\)
The continuation probability can be recursively estimated as:

pcontinue​\(Vt\)=pcontinue​\(Vt−1\)⋅CAcontinue​\(u,vt∣Vt−1\),p\_\{\\text\{continue\}\}\(V\_\{t\}\)=p\_\{\\text\{continue\}\}\(V\_\{t\-1\}\)\\cdot\\textbf\{CA\}\_\{\\text\{continue\}\}\(u,v\_\{t\}\\mid V\_\{t\-1\}\),whereCAcontinue​\(u,vt∣Vt−1\)\\textbf\{CA\}\_\{\\text\{continue\}\}\(u,v\_\{t\}\\mid V\_\{t\-1\}\)is predicted by the context\-aware model andpcontinue​\(V\)p\_\{\\text\{continue\}\}\(V\)denotes the cumulative survival of prefixVV\.

This formulation \([9](https://arxiv.org/html/2609.18148#S3.E9)\) provides a statistically more faithful estimate of the expected value of the whole list, as each item’s contribution is weighted by the probability that the user actually reaches it\.

### 3\.3PaletteDecoder: An RL\-Based Decoder

Next we introduce how to optimize the definedListVMgolden\\textbf\{ListVM\}\_\{\\text\{golden\}\}objective in \([9](https://arxiv.org/html/2609.18148#S3.E9)\), given the context aware modelCA\. It also covers optimizing theListVMvanilla\\textbf\{ListVM\}\_\{\\text\{vanilla\}\}objective \([8](https://arxiv.org/html/2609.18148#S3.E8)\) by settingpcontinue≡1p\_\{\\text\{continue\}\}\\equiv 1\.

Mathematically, this is an optimal sequential decision making problem with absorbing states\. The proposed RL\-basedPalettedecoder \(as stated in Algorithm[1](https://arxiv.org/html/2609.18148#alg1)\) generates the output list one by one by iteratively expanding the global subsequence setℰ\\mathcal\{E\}and only keeping top\-bbinℰ\\mathcal\{E\}\.

Specifically,PaletteusesListVMgolden\\textbf\{ListVM\}\_\{\\text\{golden\}\}while constructing the recommendation list one position at a time\. It starts fromV0=∅V\_\{0\}=\\emptyset\. At positiont\+1t\{\+\}1,Paletteexpands each retained sequenceVtV\_\{t\}separately by filling the next position to obtainVt\+cV\_\{t\}\{\+\}cwherec∈𝒞∖Vtc\\in\\mathcal\{C\}\\setminus V\_\{t\}allowed by the control layer\. To avoid exponential computational complexity, it keeps only the top\-bbextensions with the highest search scores\. Repeating this step untilt=Tt=Tavoids evaluating every possible full list\.

In this process, the key is how to define “top” in expansion\. InPalettewe use the following evaluation criteria\. LetV\+cV\{\+\}cbe the target subsequence, itsQQ\-value is calculated by

Q⁡\(V\+c\)=ListVMgolden​\(V\+c\)\+F^​\(V\+c\)\.\\displaystyle Q\(V\{\+\}c\)=\\textbf\{ListVM\}\_\{\\text\{golden\}\}\(V\{\+\}c\)\+\\widehat\{F\}\(V\{\+\}c\)\.\(10\)whereListVMgolden​\(V\)\\textbf\{ListVM\}\_\{\\text\{golden\}\}\(V\)evaluates a subsequenceVV, withListVMgolden​\(∅\)=0\\textbf\{ListVM\}\_\{\\text\{golden\}\}\(\\emptyset\)=0\. Algorithm[1](https://arxiv.org/html/2609.18148#alg1)retains thebbextensions with largestQQ\. The beam widthbbdetermines how many sequences are retained, whileQQdetermines which sequences are retained\.

This construction admits a value\-based RL interpretation: subsequenceVVis the state, next candidateccis the action, andV\+cV\{\+\}cis the successor state\. The accumulatedListVMgolden​\(V\+c\)\\textbf\{ListVM\}\_\{\\text\{golden\}\}\(V\{\+\}c\)is the realized return, whileF^​\(V\+c\)\\widehat\{F\}\(V\{\+\}c\)is the estimated value\-to\-go from the successor state\. Their sumQ⁡\(V\+c\)Q\(V\{\+\}c\)is used for action selection\. On the other hand, replacingCAwithCF, using the itemwise VM\+CL evaluator, and settingb=1b=1,pcontinue≡1p\_\{\\mathrm\{continue\}\}\\equiv 1, andF^=0\\widehat\{F\}=0recovers the incumbent itemwise greedy decoder\.

Algorithm 1PaletteDecoding1:user context

uu, candidates

𝒞\\mathcal\{C\}, beam width

bb, list length

TT, context\-aware modelCAand continuation predictor

CAcontinue\\textbf\{CA\}\_\{\\mathrm\{continue\}\}, item value modelitemVM, control layerCL, future\-value estimator

F^\\widehat\{F\}\(e\.g\., Eq\. \([14](https://arxiv.org/html/2609.18148#S3.E14)\)\)

2:

ℬ←\{∅\}\\mathcal\{B\}\\leftarrow\\\{\\,\\emptyset\\,\\\};

ListVMgolden​\(∅\)←0\\textbf\{ListVM\}\_\{\\text\{golden\}\}\(\\emptyset\)\\leftarrow 0;

pcontinue​\(∅\)←1p\_\{\\text\{continue\}\}\(\\emptyset\)\\leftarrow 1
3:Notation:

V∈ℬV\\in\\mathcal\{B\}is a retained prefix;

c∈𝒞∖Vc\\in\\mathcal\{C\}\\setminus Vis an unselected candidate;

V\+cV\{\+\}cappends

ccto

VV\(Eq\. \([1](https://arxiv.org/html/2609.18148#S2.E1)\)\)\.

4:\(To recover the incumbent itemwise decoding, useCFin place ofCA, set

CAcontinue≡1\\textbf\{CA\}\_\{\\mathrm\{continue\}\}\\equiv 1,

b=1b=1, and

F^=0\\widehat\{F\}=0\.\)

5:for

t=1,…,Tt=1,\\ldots,Tdo

6:

ℰ←\{V\+c:V∈ℬ,c∈𝒞∖V,CL\(c∣V\)\>−∞\}\\mathcal\{E\}\\leftarrow\\\{\\,V\{\+\}c\\;:\\;V\\in\\mathcal\{B\},\\;c\\in\\mathcal\{C\}\\setminus V,\\;\\textbf\{CL\}\(c\\mid V\)\>\-\\infty\\,\\\}⊳\\trianglerightadmissible extensions; assume≥1\\geq 1per position

7:foreach

V\+c∈ℰV\{\+\}c\\in\\mathcal\{E\}do

8:

ListVMgolden​\(V\+c\)←ListVMgolden​\(V\)\+pcontinue​\(V\)⋅\[itemVM​\(CA​\(u,c∣V\)\)\+CL​\(c∣V\)\]\\textbf\{ListVM\}\_\{\\text\{golden\}\}\(V\{\+\}c\)\\leftarrow\\textbf\{ListVM\}\_\{\\text\{golden\}\}\(V\)\+p\_\{\\text\{continue\}\}\(V\)\\cdot\\left\[\\textbf\{itemVM\}\(\\textbf\{CA\}\(u,c\\mid V\)\)\+\\textbf\{CL\}\(c\\mid V\)\\right\]
9:

pcontinue​\(V\+c\)←pcontinue​\(V\)⋅CAcontinue​\(u,c∣V\)p\_\{\\text\{continue\}\}\(V\{\+\}c\)\\leftarrow p\_\{\\text\{continue\}\}\(V\)\\cdot\\textbf\{CA\}\_\{\\text\{continue\}\}\(u,c\\mid V\)
10:

Q⁡\(V\+c\)←ListVMgolden​\(V\+c\)\+F^​\(V\+c\)Q\(V\{\+\}c\)\\leftarrow\\textbf\{ListVM\}\_\{\\text\{golden\}\}\(V\{\+\}c\)\+\\widehat\{F\}\(V\{\+\}c\)
11:

ℬ←arg​top​\-​bV′∈ℰ⁡Q​\(V′\)\\mathcal\{B\}\\leftarrow\\operatorname\*\{arg\\,top\\text\{\-\}b\}\_\{V^\{\\prime\}\\in\\mathcal\{E\}\}\\;Q\(V^\{\\prime\}\)⊳\\trianglerightthebbhighest\-QQadmissible extensions \(all ofℰ\\mathcal\{E\}if\|ℰ\|≤b\|\\mathcal\{E\}\|\\leq b\)

12:return

arg⁡maxV∈ℬ​ListVMgolden​\(V\)\\arg\\max\_\{V\\in\\mathcal\{B\}\}\\textbf\{ListVM\}\_\{\\text\{golden\}\}\(V\)⊳\\trianglerightF^​\(V\)=0\\widehat\{F\}\(V\)=0for full\-length lists

#### 3\.3\.1Future\-Value Estimation

The future value componentF^​\(⋅\)\\hat\{F\}\(\\cdot\)in \([10](https://arxiv.org/html/2609.18148#S3.E10)\) is the key differentiation between RL based generator and the standard beam search generator which does not have such component\. However, it is also different from the standard RL based approach:Paletteuses an estimate to drive a closed form estimate and the standard RL based approach acquires a model to estimate it\.

The remaining design question is how to estimateF^​\(Vt\)\\widehat\{F\}\(V\_\{t\}\)cheaply enough for serving\. We derive a lightweightF^\\widehat\{F\}from quantities already available during decoding\. Richer alternatives, such as a learned value model or a deeper planner based on Monte Carlo tree search, would require training an additional model or performing repeated model evaluations; we leave them to future work\.

##### Step\-based estimate\.

A first approximation uses the average per\-position VM\+CL score observed in the selected prefix before continuation weighting\. Define

s¯​\(Vt\)=1t​∑τ=1t\[itemVM​\(CA​\(u,vi∣Vτ−1\)\)\+CL​\(vτ∣Vτ−1\)\]\.\\displaystyle\\bar\{s\}\(V\_\{t\}\)=\\frac\{1\}\{t\}\\sum\_\{\\tau=1\}^\{t\}\\left\[\\textbf\{itemVM\}\(\\textbf\{CA\}\(u,v\_\{i\}\\mid V\_\{\\tau\-1\}\)\)\+\\textbf\{CL\}\(v\_\{\\tau\}\\mid V\_\{\\tau\-1\}\)\\right\]\.\(11\)Assume each remaining step has the same continuation probability as the most recently selected item,CAcontinue​\(u,vt∣Vt−1\)\\textbf\{CA\}\_\{\\text\{continue\}\}\(u,v\_\{t\}\\mid V\_\{t\-1\}\)\. Then the probability of reaching thejj\-th future position is approximated byCAcontinue​\(u,vt∣Vt−1\)j\\textbf\{CA\}\_\{\\text\{continue\}\}\(u,v\_\{t\}\\mid V\_\{t\-1\}\)^\{j\}, giving

F^step​\(Vt\)=s¯​\(Vt\)⋅∑j=1T−tCAcontinue​\(u,vt∣Vt−1\)j,\\displaystyle\\widehat\{F\}\_\{\\text\{step\}\}\(V\_\{t\}\)=\\bar\{s\}\(V\_\{t\}\)\\cdot\\sum\_\{j=1\}^\{T\-t\}\\textbf\{CA\}\_\{\\text\{continue\}\}\(u,v\_\{t\}\\mid V\_\{t\-1\}\)^\{j\},\(12\)which is zero att=Tt=Tbecause the sum is empty\.

This step\-based estimate biases search toward shorter current items\.CAcontinue​\(u,vt∣Vt−1\)\\textbf\{CA\}\_\{\\text\{continue\}\}\(u,v\_\{t\}\\mid V\_\{t\-1\}\)is the probability of continuing after the entire itemvtv\_\{t\}; for the same per\-second exit propensity, a longervtv\_\{t\}has a lower continuation probability\. Reusing this whole\-item probability at every future step compounds the current item’s duration across all unfilled positions, reducingF^step​\(Vt\)\\widehat\{F\}\_\{\\text\{step\}\}\(V\_\{t\}\)and making prefixes ending in long items less likely to survive beam pruning\.

##### Duration\-aware estimate\.

To mitigate this repeated\-duration bias, rescale the most recent continuation probability to the average duration of the selected items:

CAcontinue​\(u,vt∣Vt−1\)d¯/dt\.\\displaystyle\\textbf\{CA\}\_\{\\text\{continue\}\}\(u,v\_\{t\}\\mid V\_\{t\-1\}\)^\{\\bar\{d\}/d\_\{t\}\}\.\(13\)Heredtd\_\{t\}is the duration of the most recently selected item,D⁡\(Vt\)D\(V\_\{t\}\)is the total duration of thettselected items, andd¯=D⁡\(Vt\)/t\\bar\{d\}=D\(V\_\{t\}\)/tis their average duration\. The exponentd¯/dt\\bar\{d\}/d\_\{t\}rescales the whole\-item continuation probability from durationdtd\_\{t\}to durationd¯\\bar\{d\}\. Using this rescaled probability for each future position gives

F^dur​\(Vt\)=s¯​\(Vt\)⋅∑j=1T−tCAcontinue​\(u,vt∣Vt−1\)j​d¯/dt,\\displaystyle\\widehat\{F\}\_\{\\text\{dur\}\}\(V\_\{t\}\)=\\bar\{s\}\(V\_\{t\}\)\\cdot\\sum\_\{j=1\}^\{T\-t\}\\textbf\{CA\}\_\{\\text\{continue\}\}\(u,v\_\{t\}\\mid V\_\{t\-1\}\)^\{j\\bar\{d\}/d\_\{t\}\},\(14\)which is again zero att=Tt=T, where the sum is empty\. The duration\-aware treatment keeps the score in value units while replacing the most recent item’s full duration with the prefix\-average duration\. It therefore reduces the step estimator’s preference for prefixes ending in shorter items while retaining the continuation signal\. We refer to the resulting decode\-time evaluator,ListVMgolden\+F^dur\\textbf\{ListVM\}\_\{\\text\{golden\}\}\+\\widehat\{F\}\_\{\\text\{dur\}\}, as the duration\-aware listwise VM\.

Palettecompletes theLIGE\-GRpipeline: the context\-aware model predicts candidate outcomes,ListVMgolden\\textbf\{ListVM\}\_\{\\text\{golden\}\}defines the value of an ordered list, andPalettesearches for the sequence returned to the user\. Section[5](https://arxiv.org/html/2609.18148#S5)evaluates the resulting end\-to\-end system and the contribution of its decoding configurations\.

## 4Serving and Efficiency Optimization forLIGE\-GR

This section describes how to serveLIGE\-GRand the optimizations to its additional requirements for inference resources\.

### 4\.1Serving for Decoding

LIGE\-GRaugments the existing ranking service rather than introducing a parallel serving stack\. Its request path consists of two phases \(Algorithm[2](https://arxiv.org/html/2609.18148#alg2)\):

- •Context\-free embedding computation\.This phase is identical to the conventional itemwise ranking pipeline\. TheCFmodule computes the intermediate representationv′v^\{\\prime\}\(encoding both user and item information, as illustrated in Figure[3](https://arxiv.org/html/2609.18148#S2.F3)\) for every candidate independently of the output list\. Since this computation is already performed by the existing ranking system, this phase introduces no additional inference\-resource requirements\.
- •Context\-aware decoding\.The decoder then constructs the output list autoregressively\. At each positiontt, the context\-aware module re\-scores the candidate set conditioned on the previously selected prefix\.

The additional inference\-resource requirements are modest because the expensive computation is performed only once per request\. Specifically, the context\-aware decoder reuses the cached representationsv′v^\{\\prime\}produced by the context\-free module, which dominates the computation of the original ranking model\. Moreover, the context\-aware module is intentionally lightweight: it is implemented as the four\-head, four\-layer causal GPT\-style decoder shown in Figure[3](https://arxiv.org/html/2609.18148#S2.F3), representing only a tiny fraction of the computation of the base ranking model\. Consequently, each decoding step requires only a lightweight forward pass over cached representations rather than re\-running the full ranking model\. Overall, listwise generation performsTTbatched lightweight decoding passes per request\. Increasing the beam widthbbenlarges the decoding batch rather than increasing the number of sequential forward passes, while the gains from wider beams quickly saturate \(Appendix[A](https://arxiv.org/html/2609.18148#A1)\)\.

Algorithm 2LIGE\-GRserving request path1:user context

uu, candidates

𝒞\\mathcal\{C\}, beam width

bb, list length

TT, per\-request latency budget

τ\\tau, and the inputs of Algorithm[1](https://arxiv.org/html/2609.18148#alg1)\(CA,

CAcontinue\\textbf\{CA\}\_\{\\mathrm\{continue\}\},itemVM,CL,

F^\\widehat\{F\}\)

2:compute and cache

CF​\(u,c\)\\textbf\{CF\}\(u,c\)and its intermediate representation

vc′v^\{\\prime\}\_\{c\},

∀c∈𝒞\\forall c\\in\\mathcal\{C\}⊳\\trianglerightPhase 1: the unchanged forward pass

3:

𝒱′←\{vc′\}c∈𝒞\\mathcal\{V\}^\{\\prime\}\\leftarrow\\\{v^\{\\prime\}\_\{c\}\\\}\_\{c\\in\\mathcal\{C\}\}⊳\\trianglerightcached once per request; no candidate re\-encoding during decoding

4:

V←Palette​\(u,𝒞,b,T,CA,CAcontinue,itemVM,CL,F^\)V\\leftarrow\\textsc\{Palette\}\(u,\\mathcal\{C\},b,T;\\,\\textbf\{CA\},\\textbf\{CA\}\_\{\\mathrm\{continue\}\},\\textbf\{itemVM\},\\textbf\{CL\},\\widehat\{F\}\)⊳\\trianglerightPhase 2==Algorithm[1](https://arxiv.org/html/2609.18148#alg1):TTbatched module invocations, each evaluating up tobbbeam prefixes over𝒱′\\mathcal\{V\}^\{\\prime\}

5:ifPhase 2 exceeds

τ\\taums or anyCAcall failsthen

6:return

Palette​\(u,𝒞,1,T,CF,1,itemVM,CL,0\)\\textsc\{Palette\}\(u,\\mathcal\{C\},1,T;\\,\\textbf\{CF\},1,\\textbf\{itemVM\},\\textbf\{CL\},0\)⊳\\trianglerightper\-request fallback: the itemwise decoder; reuses the cachedCFscores, noCAcalls

7:return

VV

### 4\.2Reliability and Reversibility

Reliability and continued iteration matter when upgrading an existing system\. TheLIGE\-GRframework supports configuration\-level reversion without retraining:

- •SwitchCApredictions toCF; no retraining is required\.
- •Recover the incumbent itemwise decoder by settingb=1b=1,pcontinue≡1p\_\{\\text\{continue\}\}\\equiv 1, andF^=0\\widehat\{F\}=0in Algorithm[1](https://arxiv.org/html/2609.18148#alg1)\.

The preserved context\-free path makesLIGE\-GRreversible at every granularity\.

Per request, if the decoding phase cannot complete within the request’s latency budgetτ\\tau\(Algorithm[2](https://arxiv.org/html/2609.18148#alg2)\), the system automatically falls back to the itemwise behavior for that request\. Globally, reverting to the baseline configuration is a switch rather than a migration: disabling the context\-aware path—and with it the listwise VM and beam search, which are built on its predictions—recovers the itemwise system exactly, by the strict\-generalization property of Section[2](https://arxiv.org/html/2609.18148#S2)\. The same additivity also decouples iteration: the context\-aware module can be updated independently of the base model, without touching the base model’s training or publishing flow\.

### 4\.3Efficiency Optimization

Section[4\.1](https://arxiv.org/html/2609.18148#S4.SS1)explained why the architecture is resource\-efficient by design: the forward pass is reused, and each decode step is a small forward pass over cached representations\. This section describes the serving\-side optimizations that control the remaining resource requirements—listwise construction over a full candidate set at serving traffic would still add unnecessary computation—and reports the measured requirements of the configuration\.

Restricting the re\-scoring pool\.The context\-aware path does not need to re\-score the full candidate set\. The context\-free scores from the first serving phase are already a high\-quality itemwise ranking, and candidates ranked far beyond the list length have low selection probability, so only the top\-ranked candidates—roughly a third of the set—are passed to the context\-aware module for listwise construction\. This bounds the overhead of the second phase regardless of the decoding configuration: in serving benchmarks across two GPU generations, the trimmed pool raises the context\-aware path’s throughput by roughly 60–80% relative to re\-scoring the full set, and the online improvements of Section[5](https://arxiv.org/html/2609.18148#S5)are obtained under the trimmed pool\. The context\-free ranking thus acts as a learned pre\-filter for the listwise stage—another way the preserved itemwise system contributes to the listwise stage\.

Batching the beam\.At each decode step, all beam continuations are evaluated in one batched forward pass of the lightweight context\-aware module\. At the same candidate\-pool size and traffic,b=6b=6requires roughly2\.1×2\.1\\timesthe inference resources ofb=1b=1, giving a derived estimate of approximately 20% of theCFcomponent’s resources\.

Right\-sizing generation\.The remaining knobs match compute to where quality still improves\. Beam width is set at the saturation knee of the offline gain curve \(Figure[5](https://arxiv.org/html/2609.18148#A1.F5)\): beyond small widths, additional beams provide little score improvement while still requiring more inference resources\. In one evaluation setting, the decoder also generates only as many positions as the response actually requests—response sizes vary at serving time—rather than always decoding the maximum list length and truncating\.

Across both evaluated settings, the base configuration requires additional inference resources equivalent to roughly 10% of those used by theCFcomponent\. On Instagram Reels, the evaluated upgrade increases end\-to\-end per\-request latency by roughly 7% relative to its baseline\. On Facebook Video, it increases average serving latency by about 2\.2% relative to its baseline\. The latency definitions and baselines differ across the two settings, so the magnitudes are not directly comparable\. These measured resource and latency changes accompany the online improvements reported in Section[5](https://arxiv.org/html/2609.18148#S5)\.

## 5Experimental Results

This section evaluatesLIGE\-GRin short\-video recommendation settings on Instagram Reels and Facebook Video\. Section[5\.1](https://arxiv.org/html/2609.18148#S5.SS1)compares context\-aware prediction \(CA\) with the context\-free baseline \(CF\) using normalized entropy \(NE\)\([Liu et al\., 2023a](https://arxiv.org/html/2609.18148#bib.bib21)\)\. Sections[5\.2](https://arxiv.org/html/2609.18148#S5.SS2)and[5\.3](https://arxiv.org/html/2609.18148#S5.SS3)report online results for theb=1b=1base configuration against the corresponding incumbent systems and for theb=6b=6duration\-aware configuration against theb=1b=1base configuration, respectively\. Section[5\.4](https://arxiv.org/html/2609.18148#S5.SS4)provides additional content\-ecosystem analysis\. In the evaluated settings, the candidate\-pool size\|𝒞\|\|\\mathcal\{C\}\|is on the order of10210^\{2\}, while the output\-list lengthTTis on the order of1010\(Eq\. \([1](https://arxiv.org/html/2609.18148#S2.E1)\)\)\.

### 5\.1Context\-Aware vs\. Context\-Free Prediction

Before evaluating end\-to\-end ranking outcomes, we isolate the effect of contextualization on prediction quality\. Table[1](https://arxiv.org/html/2609.18148#S5.T1)comparesCAwithCF; positive values indicate thatCAachieves lower \(better\) NE\. On Instagram Reels,CAimproves all six displayed task families, led by Continue \(1\.57%\)\. On Facebook Video, it improves all five task families shown\. Beyond this displayed subset, the complete 17\-task Facebook Video evaluation shows prediction\-quality improvements on 15 tasks and regressions of at most 0\.07% on two minor tasks\. These results establish the prediction\-quality benefit ofCA; Sections[5\.2](https://arxiv.org/html/2609.18148#S5.SS2)and[5\.3](https://arxiv.org/html/2609.18148#S5.SS3)next evaluate whether that benefit translates into online ranking outcomes\.

Table 1:Relative NE improvement of the context\-aware predictor \(CA\) against the context\-free baseline \(CF\)\. Positive values indicate lower \(better\) NE\. Instagram Reels values are accumulated over one online\-training run; Facebook Video values come from its context\-aware\-versus\-context\-free evaluation\. ‘—’ marks a task not tracked\.
### 5\.2Online Validation:b=1b=1Base Configuration

Table 2:Online A/B results for theb=1b=1base configuration in Eq\. \([8](https://arxiv.org/html/2609.18148#S3.E8)\), each against the corresponding incumbent recommendation baseline\. Sessions and DAU reflect overall product activity; the remaining metrics are restricted to the evaluated short\-video surface\. The Instagram Reels entry in the Likes / reactions row is likes; the Facebook Video counterpart is reactions\.†\\daggerdenotesp<0\.001p<0\.001\.Table 3:Online Instagram Reels A/B contrasts against theb=1b=1base configuration used in Table[2](https://arxiv.org/html/2609.18148#S5.T2)\. All configurations useCA, and each treatment row is compared independently with that base configuration; ‘—’ marks the reference row\.†\\daggerdenotesp<0\.001p<0\.001\.Table 4:Instagram Reels paired\-request list\-composition changes relative to the baseline\. Topic ranges span the taxonomies analyzed;greenmarks diagnostic improvements andredmarks the freshness regression\.We first test whether the prediction\-quality gains above translate into online gains whenLIGE\-GRis integrated into the ranking system\. The base configuration is

whereListVMvanilla\\textbf\{ListVM\}\_\{\\text\{vanilla\}\}is defined in Eq\. \([8](https://arxiv.org/html/2609.18148#S3.E8)\)\. Each product is evaluated against its own actively optimized incumbent baseline\.

For Instagram Reels, the treatment and baseline groups each included 1\.5% of users\. For Facebook Video, the treatment group included approximately 2% of users and was compared with a similarly sized control group\. Table[2](https://arxiv.org/html/2609.18148#S5.T2)reports seven\-day readouts for both products\. Across both evaluations,LIGE\-GRbroadly improves key consumption and engagement metrics, with the individual movements and tradeoffs reported in the table\. Across both settings, the base configuration requires additional inference resources equivalent to roughly 10% of those used by theCFcomponent\. In the Instagram Reels evaluation, end\-to\-end per\-request latency increases by roughly 7% relative to its baseline\. In the Facebook Video evaluation, average serving latency increases by about 2\.2% relative to its baseline\. Because the latency definitions and baselines differ, these measurements are not directly comparable\. Appendix[4](https://arxiv.org/html/2609.18148#S4)provides serving and efficiency details\.

These results establish online gains for theb=1b=1base configuration\. We next ask whether richer decoding adds value beyond this configuration, first by widening the beam and then by enabling the combined listwise, duration\-aware configuration\.

### 5\.3Instagram Reels Online Validation:b=6b=6Duration\-Aware Configuration

We evaluate this next step on Instagram Reels, using theb=1b=1base configuration from Section[5\.2](https://arxiv.org/html/2609.18148#S5.SS2)as the common baseline\. The duration\-aware configuration uses

HereListVMgolden\\textbf\{ListVM\}\_\{\\text\{golden\}\}andF^dur\\widehat\{F\}\_\{\\mathrm\{dur\}\}are defined in Eqs\. \([9](https://arxiv.org/html/2609.18148#S3.E9)\) and \([14](https://arxiv.org/html/2609.18148#S3.E14)\), respectively\. We compare three configurations with identical weights insideitemVM\. The baseline usesb=1b=1,ListVMvanilla\\textbf\{ListVM\}\_\{\\text\{vanilla\}\}, andF^=0\\widehat\{F\}=0\. The beam\-width arm increasesbbfrom 1 to 6 while keeping the objective and future\-value estimate unchanged\. The duration\-aware arm usesb=6b=6withListVMgolden\\textbf\{ListVM\}\_\{\\text\{golden\}\}andF^dur\\widehat\{F\}\_\{\\mathrm\{dur\}\}\. Appendix[A](https://arxiv.org/html/2609.18148#A1)motivates the choice ofb=6b=6\.

Eachb=6b=6treatment group included approximately 3% of users and was compared with a similarly sizedb=1b=1control group\. Table[3](https://arxiv.org/html/2609.18148#S5.T3)reports independent seven\-day A/B contrasts of eachb=6b=6treatment against the Instagram Reelsb=1b=1base configuration used in Table[2](https://arxiv.org/html/2609.18148#S5.T2)\. The beam\-width arm primarily improves likes and reshares, while the duration\-aware arm improves all four reported metrics\. At the same candidate\-pool size and traffic, theb=6b=6configuration is estimated to require additional inference resources equivalent to approximately 20% of those used by theCFcomponent\. A comparable end\-to\-end latency estimate is not available forb=6b=6; Appendix[4](https://arxiv.org/html/2609.18148#S4)provides details\.

### 5\.4Impact on the Ecosystem

Beyond online outcomes, we examine howLIGE\-GRin the setting of Section[5\.2](https://arxiv.org/html/2609.18148#S5.SS2)changes the ecosystem on Instagram Reels\. Table[4](https://arxiv.org/html/2609.18148#S5.T4)summarizes paired\-request diagnostics across topics, creators, exploration, length, and freshness\. The overall observation isLIGE\-GRimproves the product ecosystem from multiple dimensions with minor regression on content freshness\.

![Refer to caption](https://arxiv.org/html/2609.18148v1/figures/case_study.png)Figure 4:A representative paired request from the Instagram Reels evaluation’s counterfactual logs: the baseline list \(top\) and theLIGE\-GRlist \(bottom\) generated from identical inputs, with each item’s destination or origin position annotated\. Video thumbnails are AI\-generated stand\-in images matching each item’s topic \(one per unique video\); topic, duration, age, and movement annotations are from the logged request\.LIGE\-GRpreserves every baseline topic while introducing a new one \(Sports\), moves the 27\.5\-second video from position 2 to position 4, demotes the 70\-day\-old video to the final position, and removes one of the two Internet Culture items\.##### Methodology\.

To understand howLIGE\-GRchanges generated lists, we use counterfactual logging that records both the baseline pointwise list and theLIGE\-GRcontext\-aware list for the same request over identical inputs\. We analyze 13,197 paired requests from 7,110 known logged users\. 1,498 of these requests lacked logged viewer IDs and are conservatively counted as one user each, giving 8,608 effective users\. These data cover a three\-day window and are sampled at the request level within an experiment cohort, which skews the sample toward more\-active viewers\. Table[4](https://arxiv.org/html/2609.18148#S5.T4)in the main text summarizes the resulting composition shifts\. All reported effects are assessed to be statistically significant at the 95% level via separate Bayesian hierarchical linear models for each estimate that account for multiple requests by the same user\. We report effect sizes as relative changes compared to the baseline and interpret the main tradeoffs below\.

##### Diversity and repetition\.

Table[4](https://arxiv.org/html/2609.18148#S5.T4)shows thatLIGE\-GRincreases topic variety and topic coverage while reducing repeated\-topic runs\. Adjacent videos also become less similar to one another on average, improving list\-level variety while introducing a smoothness–diversity tradeoff that is invisible in purely pointwise evaluation\.

##### Creator mix and exploration\.

The generated lists include more distinct creators, lower same\-creator concentration, and higher creator entropy\. They also shift away from the largest creators and from content already familiar to the user or tightly matched to known interests\. We interpret these movements as an exploration tradeoff rather than an unqualified win:LIGE\-GRexposes a broader set of creators and topics, while giving up some immediate affinity matching\.

##### Length and freshness\.

Within\-request length entropy increases, so users see a wider variety of video lengths within the same request\. The fraction of very fresh videos decreases, matching the online recency guardrail movement; we therefore treat freshness as the table’s composition regression rather than as a quality improvement\.

Together, Figure[4](https://arxiv.org/html/2609.18148#S5.F4)and Table[4](https://arxiv.org/html/2609.18148#S5.T4)show thatLIGE\-GRchanges ecosystem exposure across topics, creators, familiarity, length, and freshness\. These diagnostics characterize how generated lists redistribute exposure across the content and creators represented in those lists, while the online experiments above measure the corresponding product impact\.

## 6Related Work

We review the related work in this section from four aspects: itemwise recommendation, traditional listwise recommendation, generative and autoregressive slate optimization, and LLM\-inspired generative recommendation\. The difference from prior work is that this paper addresses recommendation from an upgrade\-path perspective\. Instead of focusing on improving a single model or algorithm, it studies how to upgrade a mature itemwise recommendation system—refined through years of iteration—into an LLM\-inspired generative recommendation paradigm\. The emphasis is not only on enabling richer listwise optimization, but also on providing a practical migration path that minimizes disruption to existing infrastructure, product logic, and engineering investment\.

Itemwise Recommendation\.Early work on recommendation primarily follows a simple yet effective itemwise paradigm\([Wang et al\., 2021b](https://arxiv.org/html/2609.18148#bib.bib45);[Wang et al\., 2017](https://arxiv.org/html/2609.18148#bib.bib44);[Cheng et al\., 2016](https://arxiv.org/html/2609.18148#bib.bib7);[Zhou et al\., 2018](https://arxiv.org/html/2609.18148#bib.bib50);[Ma et al\., 2018](https://arxiv.org/html/2609.18148#bib.bib24);[Rendle, 2010](https://arxiv.org/html/2609.18148#bib.bib34);[Ning and Slim,](https://arxiv.org/html/2609.18148#bib.bib26);[Sarwar et al\., 2001](https://arxiv.org/html/2609.18148#bib.bib35);[Guo et al\., 2017](https://arxiv.org/html/2609.18148#bib.bib13)\)\. Item and user features are first compressed by feature encoders\([Wang et al\., 2021b](https://arxiv.org/html/2609.18148#bib.bib45)\), and items are then ranked independently by their relevance to the user\. Research along this line explores better feature modeling, user–item relevance scoring, and multi\-task learning\. A parallel line applies sequence modeling to user interaction histories while still scoring each candidate independently, through self\-attention\([Kang and McAuley, 2018](https://arxiv.org/html/2609.18148#bib.bib17);[Sun et al\., 2019](https://arxiv.org/html/2609.18148#bib.bib37)\), recurrent\([Hidasi et al\., 2016](https://arxiv.org/html/2609.18148#bib.bib14)\)and convolutional\([Tang and Wang, 2018](https://arxiv.org/html/2609.18148#bib.bib38)\)encoders, target\-aware attention\([Zhou et al\., 2018](https://arxiv.org/html/2609.18148#bib.bib50);[Xia et al\., 2023](https://arxiv.org/html/2609.18148#bib.bib48)\), memory\-based modeling of lifelong behavior\([Pi et al\., 2019](https://arxiv.org/html/2609.18148#bib.bib31)\), and industrial sequential transduction\([Zhai et al\., 2024](https://arxiv.org/html/2609.18148#bib.bib49)\)\.LIGE\-GRinstead conditions each candidate prediction on the items already selected for the current list\.

Traditional Listwise Recommendation\.Listwise Recommendation was introduced to address the above problem\([Cao et al\., 2007](https://arxiv.org/html/2609.18148#bib.bib6);[Ai et al\., 2019](https://arxiv.org/html/2609.18148#bib.bib2)\), while a large body of work introduces cross\-item modeling at a dedicated*reranking*stage\. DLCM\([Ai et al\., 2018](https://arxiv.org/html/2609.18148#bib.bib1)\)encodes ranking context with a GRU to refine scores; PRM\([Pei et al\., 2019](https://arxiv.org/html/2609.18148#bib.bib30)\)and SetRank\([Pang et al\., 2020](https://arxiv.org/html/2609.18148#bib.bib28)\)apply bidirectional self\-attention over the candidate set to capture mutual influence in a single pass; PEAR\([Li et al\., 2022](https://arxiv.org/html/2609.18148#bib.bib19)\)adds personalized contextualized transformers; MIR\([Xi et al\., 2022](https://arxiv.org/html/2609.18148#bib.bib47)\)jointly models set\-level candidates and user history; and PIER\([Shi et al\., 2023](https://arxiv.org/html/2609.18148#bib.bib36)\)selects among candidate permutations end\-to\-end\. A prominent instantiation is the “generator–evaluator” \(G–E\) framework, which pairs a list generator with a list evaluator that scores the quality of a generated sequence\([Wang et al\., 2019](https://arxiv.org/html/2609.18148#bib.bib43);[Feng et al\., 2021b](https://arxiv.org/html/2609.18148#bib.bib10);[Feng et al\., 2021a](https://arxiv.org/html/2609.18148#bib.bib9);[Ren et al\., 2024](https://arxiv.org/html/2609.18148#bib.bib33)\)\. The evaluator guides the generator so that together they produce a high\-quality list in terms of relevance or engagement\. NAR4Rec\([Ren et al\., 2024](https://arxiv.org/html/2609.18148#bib.bib33)\)in particular constructs multiple slates and selects the best via a learned slatewise evaluator\. Score\-refinement methods form the final slate by sorting contextualized scores, whereas generator–evaluator methods generate candidate slates and select among them\. Both families are typically introduced through dedicated reranking machinery on top of the existing system\.LIGE\-GRdelivers the G–E benefits*inside*the existing ranking stage, without adding a dedicated reranking stage\. Within this framing,LIGE\-GRinterleaves generation and evaluation: it scores partial lists and prunes low\-valued prefixes before completion\. This contrasts with generate\-then\-evaluate variants that evaluate completed candidate slates\.

Generative and Autoregressive Slate Construction\.Rather than re\-sorting scores, another family directly generates the slate\. List\-CVAE\([Jiang et al\., 2018](https://arxiv.org/html/2609.18148#bib.bib16)\)uses a conditional variational auto\-encoder to generate stochastic lists, and pivot\-CVAE\([Liu et al\., 2021](https://arxiv.org/html/2609.18148#bib.bib22)\)improves it to guarantee list variation and mitigate over\-concentration\. GFN4Rec\([Liu et al\., 2023b](https://arxiv.org/html/2609.18148#bib.bib23)\)adapts GFlowNet\([Bengio et al\., 2021](https://arxiv.org/html/2609.18148#bib.bib4)\)to sample sequences with probability proportional to the listwise reward\. A closely related thread constructs the slate*autoregressively*\. Seq2Slate\([Bello et al\., 2018](https://arxiv.org/html/2609.18148#bib.bib3)\)selects items one at a time with an RNN encoder–decoder, compressing prior context into a fixed\-size hidden state\. SlateQ\([Ie et al\., 2019](https://arxiv.org/html/2609.18148#bib.bib15)\)decomposes the combinatorial slate reward into per\-item Q\-values under a user choice model, but does not model inter\-item interactions during construction\. GFN4Rec\([Liu et al\., 2023b](https://arxiv.org/html/2609.18148#bib.bib23)\)also decodes autoregressively but without direct attention over previously selected items\. More recent work accelerates or reshapes this decoding\. GReF\([Lin et al\., 2025](https://arxiv.org/html/2609.18148#bib.bib20)\)uses ordered multi\-token prediction and replaces the evaluator with preference\-based training, and HiGR\([Pang et al\., 2025](https://arxiv.org/html/2609.18148#bib.bib29)\)adds hierarchical planning with multi\-objective preference alignment\. Alternative generative paradigms avoid sequential decoding entirely\.[Tomasi et al\. \(2025\)](https://arxiv.org/html/2609.18148#bib.bib40)cast slate construction as parallel denoising diffusion, which removes sequential latency but cannot enforce per\-step constraints\.LIGE\-GRalso adopts the autoregressive formulation, but unlike Seq2Slate’s fixed hidden state or GFN4Rec’s attention\-free decoding, it applies full causal attention over all previously selected items within a single architecture\. A practical advantage is thatLIGE\-GRhas an explicit objective, and generation can be controlled by changing objective weights or VM forms\.

LLM\-Inspired Generative RecommendationFollowing the success of LLMs, the recommendation community has increasingly adopted their techniques in ranking systems\([Wu et al\., 2024](https://arxiv.org/html/2609.18148#bib.bib46)\)\. One idea is to*tokenize*items so that recommendation can be cast as sequence generation\. At the retrieval stage, TIGER\([Rajput et al\., 2023](https://arxiv.org/html/2609.18148#bib.bib32)\), building on the differentiable search index\([Tay et al\., 2022](https://arxiv.org/html/2609.18148#bib.bib39)\), encodes each item into multi\-modal semantic IDs and autoregressively predicts the next ID to be consumed, which is effectively a generative retrieval paradigm\. In a similar spirit, streaming VQ\([Bin et al\., 2025](https://arxiv.org/html/2609.18148#bib.bib5)\)tokenizes items with VQ\-VAE\([van den Oord et al\., 2017](https://arxiv.org/html/2609.18148#bib.bib41)\), upgrading the industrial index into a learnable, instantly updatable, and balanced structure, while P5\([Geng et al\., 2022](https://arxiv.org/html/2609.18148#bib.bib11)\)unifies diverse recommendation tasks under a single text\-to\-text formulation\. Pushing this direction further, OneRec\([Deng et al\., 2025](https://arxiv.org/html/2609.18148#bib.bib8)\)couples generative retrieval with an Iterative Preference Alignment module and is the first approach to replace the entire recommendation funnel with one unified model\. Unlike OneRec’s wholesale replacement of the funnel,LIGE\-GRreaches listwise, LLM\-style generation by upgrading the existing ranking stage*in place*: it captures contextual cues, makes listwise predictions, and delivers the generator–evaluator benefits without adding a stage or replacing the stack\.

## 7Conclusion and Future Work

In this paper, we present a practical and low\-resource\-requirement framework for upgrading traditional itemwise recommendation systems toward a generative, listwise recommendation paradigm\. The proposed approach introduces a context\-aware module on top of existing ranking models, upgrades the itemwise value model into a listwise value model, and replaces conventional itemwise greedy decoding with an RL\-based decoder\. Through validation on Instagram Reels and Facebook Video, we demonstrate the framework in real\-world industrial recommendation systems\. A design goal of this work is to enable a smooth and incremental transition from mature itemwise systems to listwise generative systems\. Rather than requiring a disruptive replacement of existing infrastructure, models, or organizational ownership, the proposed framework generalizes and extends the existing recommendation stack\. This makes the approach more practical for industrial environments, where technical migration effort, system reliability, latency constraints, and cross\-team ownership all matter\.

This work should not be viewed as the final form of generative recommendation\. Instead, it provides a transition framework with headroom\. For example, more expressive architectures may further improve the model’s ability to capture list\-level dependencies; the listwise objective can be extended to incorporate richer list\-level signals that are difficult to define in an itemwise system; more advanced decoding and reinforcement learning methods may further improve the diversity and long\-term quality of generated recommendation lists\. The currentLIGE\-GRframework can only handle an input candidate pool on the order of hundreds\. To achieve truly end\-to\-end recommendation, it will need to be integrated with technologies such as Semantic IDs that can efficiently support much larger candidate spaces\.

LIGE\-GRoffers additional future potential for end\-to\-end optimization across the full serving stack\. In latency\-sensitive serving applications, with streaming inference, we can stream out individual results and serve them to the user as they become available\. We can further optimize the case of greedy search by immediately yielding the first result before applying the CA pass, as the first result will not be affected\. The complexity of the CA pass also gives additional potential for dynamic compute complexity scaling based on varying compute supply and demand\. This work provides a foundation for future research and system development toward more expressive, controllable, and value\-aligned recommendation systems\.

## Acknowledgments

We thank Rex Cheung, Erica Li, Lars Backstrom, Max Eulenstein, Jayant Subramanian, Bruce Deng, Yimin Tan, Qichao Que, Jerry Fu, Congle Zhang, Lihong Li, Fei Sha, and Mahesh Srinivasan for their insightful technical discussions, constructive feedback, and many valuable contributions throughout this work\. We are particularly grateful to Shilin Ding for championing this project from its inception and providing steadfast support throughout its 0\-to\-1 journey\.

## References

- Ai et al\. \(2018\)Qingyao Ai, Keping Bi, Jiafeng Guo, and W Bruce Croft\.Learning a deep listwise context model for ranking refinement\.In*Proc\. SIGIR’18*, 2018\.
- Ai et al\. \(2019\)Qingyao Ai, Xuanhui Wang, Nadav Golbandi, Mike Bendersky, and Marc Najork\.Learning groupwise scoring functions using deep neural networks\.In*Proc\. International Workshop On Deep Matching In Practical Applications’19*, 2019\.
- Bello et al\. \(2018\)Irwan Bello, Sayali Kulkarni, Sagar Jain, Craig Boutilier, Ed Chi, Elad Eban, Xiyang Luo, Alan Mackey, and Ofer Meshi\.Seq2slate: Re\-ranking and slate optimization with rnns\.*arXiv preprint arXiv:1810\.02019*, 2018\.
- Bengio et al\. \(2021\)Emmanuel Bengio, Moksh Jain, Maksym Korablyov, Doina Precup, and Yoshua Bengio\.Flow network based generative models for non\-iterative diverse candidate generation\.In*Proc\. NeurIPS’21*, 2021\.
- Bin et al\. \(2025\)Xingyan Bin, Jianfei Cui, Wujie Yan, Zhichen Zhao, Xintian Han, Chongyang Yan, Feng Zhang, Xun Zhou, Xiao Yang, and Zuotao Liu\.Real\-time indexing for large\-scale recommendation by streaming vector quantization retriever\.In*Proc\. KDD’25*, 2025\.
- Cao et al\. \(2007\)Zhe Cao, Tao Qin, Tie\-Yan Liu, Ming\-Feng Tsai, and Hang Li\.Learning to rank: from pairwise approach to listwise approach\.In*Proc\. ICML’07*, 2007\.
- Cheng et al\. \(2016\)Heng\-Tze Cheng, Levent Koc, Jeremiah Harmsen, Tal Shaked, Tushar Chandra, Hrishi Aradhye, Glen Anderson, Greg Corrado, Wei Chai, Mustafa Ispir, et al\.Wide & deep learning for recommender systems\.In*Proc\. Workshop on Deep Learning for Recommender Systems’16*, 2016\.
- Deng et al\. \(2025\)Jiaxin Deng, Shiyao Wang, Kuo Cai, Lejian Ren, Qigen Hu, Weifeng Ding, Qiang Luo, and Guorui Zhou\.Onerec: unifying retrieve and rank with generative recommender and iterative preference alignment\.*arXiv preprint arXiv:2502\.18965*, 2025\.
- Feng et al\. \(2021a\)Yufei Feng, Yu Gong, Fei Sun, Junfeng Ge, and Wenwu Ou\.Revisit recommender system in the permutation prospective\.*arXiv preprint arXiv:2102\.12057*, 2021a\.
- Feng et al\. \(2021b\)Yufei Feng, Binbin Hu, Yu Gong, Fei Sun, Qingwen Liu, and Wenwu Ou\.Grn: Generative rerank network for context\-wise recommendation\.*arXiv preprint arXiv:2104\.00860*, 2021b\.
- Geng et al\. \(2022\)Shijie Geng, Shuchang Liu, Zuohui Fu, Yingqiang Ge, and Yongfeng Zhang\.Recommendation as language processing \(rlp\): A unified pretrain, personalize, prompt and predict paradigm \(p5\)\.In*Proc\. RecSys’22*, 2022\.
- Gong et al\. \(2021\)Yu Gong, Xi Jiang, Yuwei Wang, Lin Lin, Kai Feng, and Hongyan Liang\.Edge\-cloud polarized reranking system for web\-scale video recommendation\.In*Proc\. SIGIR’21*, 2021\.
- Guo et al\. \(2017\)Huifeng Guo, Ruiming Tang, Yunming Ye, Zhenguo Li, and Xiuqiang He\.Deepfm: a factorization\-machine based neural network for ctr prediction\.In*Proc\. IJCAI’17*, 2017\.
- Hidasi et al\. \(2016\)Balázs Hidasi, Alexandros Karatzoglou, Linas Baltrunas, and Domonkos Tikk\.Session\-based recommendations with recurrent neural networks\.In*Proc\. ICLR’16*, 2016\.
- Ie et al\. \(2019\)Eugene Ie, Vihan Jain, Jing Wang, Sanmit Narvekar, Ritesh Agarwal, Rui Wu, Heng\-Tze Cheng, Tushar Chandra, and Craig Boutilier\.SLATEQ: A tractable decomposition for reinforcement learning with recommendation sets\.In*Proc\. IJCAI’19*, 2019\.
- Jiang et al\. \(2018\)Ray Jiang, Sven Gowal, Timothy A Mann, and Danilo J Rezende\.Beyond greedy ranking: Slate optimization via list\-cvae\.*arXiv preprint arXiv:1803\.01682*, 2018\.
- Kang and McAuley \(2018\)Wang\-Cheng Kang and Julian McAuley\.Self\-attentive sequential recommendation\.In*Proc\. ICDM’18*, 2018\.
- Li et al\. \(2018\)Chengtao Li, Sriram Shamaiah, Zhen Zhe, Hongyan Liang, and Yiyang Yang\.Fast greedy map inference for determinantal point process to improve recommendation diversity\.In*Proc\. NeurIPS’18*, 2018\.
- Li et al\. \(2022\)Yi Li, Jieming Zhu, Weiwen Liu, Liangcai Su, Guohao Cai, Qi Zhang, Ruiming Tang, Xi Xiao, and Xiuqiang He\.PEAR: Personalized re\-ranking with contextualized transformer for recommendation\.In*Companion Proc\. TheWebConf’22*, 2022\.
- Lin et al\. \(2025\)Zhijie Lin, Zhuofeng Li, Chenglei Dai, Wentian Bao, Shuai Lin, Enyun Yu, Haoxiang Zhang, and Liang Zhao\.GReF: A unified generative framework for efficient reranking via ordered multi\-token prediction\.In*Proc\. CIKM’25*, 2025\.
- Liu et al\. \(2023a\)Hanyang Liu, Shuai Yang, Feng Qi, and Shuaiwen Wang\.Learning to rank normalized entropy curves with differentiable window transformation\.*arXiv preprint arXiv:2301\.10443*, 2023a\.
- Liu et al\. \(2021\)Shuchang Liu, Fei Sun, Yingqiang Ge, Changhua Pei, and Yongfeng Zhang\.Variation control and evaluation for generative slate recommendations\.In*Proc\. TheWebConf’21*, 2021\.
- Liu et al\. \(2023b\)Shuchang Liu, Qingpeng Cai, Zhankui He, Bowen Sun, Julian McAuley, Dong Zheng, Peng Jiang, and Kun Gai\.Generative flow network for listwise recommendation\.In*Proc\. KDD’23*, 2023b\.
- Ma et al\. \(2018\)Jiaqi Ma, Zhe Zhao, Xinyang Yi, Jilin Chen, Lichan Hong, and Ed H Chi\.Modeling task relationships in multi\-task learning with multi\-gate mixture\-of\-experts\.In*Proc\. KDD’18*, 2018\.
- Meng et al\. \(2019\)Shuchang Meng, Xiaolin Zhang, Minh\-Thang Xuan, Eric Zhan, and Yiyang Yang\.Tensorized determinantal point processes for recommendation\.In*Proceedings of the 25th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining \(KDD\)*, pages 1885–1894, 2019\.
- \(26\)X Ning and G Karypis Slim\.Sparse linear methods for top\-n recommender systems\.In*Proceedings of the 2011 IEEE 11th International Conference on Data Mining*, pages 497–506\.
- Pan et al\. \(2020\)Yushun Pan, Fu\-Lai Qian, Li Chen, Hongyan Liang, and Yiyang Yang\.Purs: Personalized unexpected recommender system for improving user satisfaction\.In*Proceedings of the 14th ACM Conference on Recommender Systems \(RecSys\)*, pages 283–292, 2020\.
- Pang et al\. \(2020\)Liang Pang, Jun Xu, Qingyao Ai, Yanyan Lan, Xueqi Cheng, and Jirong Wen\.SetRank: Learning a permutation\-invariant ranking model for information retrieval\.In*Proceedings of the 43rd International ACM SIGIR Conference on Research and Development in Information Retrieval*, pages 499–508, 2020\.
- Pang et al\. \(2025\)Yunsheng Pang, Zijian Liu, Yudong Li, Shaojie Zhu, Zijian Luo, Chenyun Yu, Sikai Wu, Shichen Shen, Congying Xia, Yanchi Liu, Haifeng Chen, and Liang Wang\.HiGR: Efficient generative slate recommendation via hierarchical planning and multi\-objective preference alignment\.*arXiv preprint arXiv:2512\.24787*, 2025\.
- Pei et al\. \(2019\)Changhua Pei, Yi Zhang, Yongfeng Zhang, Fei Sun, Xiao Lin, Hanxiao Sun, Jian Wu, Peng Jiang, Junfeng Ge, Wenwu Ou, et al\.Personalized re\-ranking for recommendation\.In*Proceedings of the 13th ACM conference on recommender systems*, pages 3–11, 2019\.
- Pi et al\. \(2019\)Qi Pi, Weijie Bian, Guorui Zhou, Xiaoqiang Zhu, and Kun Gai\.Practice on long sequential user behavior modeling for click\-through rate prediction\.In*Proc\. KDD’19*, 2019\.
- Rajput et al\. \(2023\)Shashank Rajput, Nikhil Mehta, Anima Singh, Raghunandan Keshavan, Trung Vu, Lukasz Heidt, Lichan Hong, Yi Tay, Vinh Q\. Tran, Jonah Samost, Maciej Kula, Ed H\. Chi, and Maheswaran Sathiamoorthy\.Recommender systems with generative retrieval\.In*Proc\. NeurIPS’23*, 2023\.
- Ren et al\. \(2024\)Yuxin Ren, Qiya Yang, Yichun Wu, Wei Xu, Yalong Wang, and Zhiqiang Zhang\.Non\-autoregressive generative models for reranking recommendation\.In*Proc\. KDD’24*, 2024\.
- Rendle \(2010\)Steffen Rendle\.Factorization machines\.In*Proc\. ICDM’10*, 2010\.
- Sarwar et al\. \(2001\)Badrul Sarwar, George Karypis, Joseph Konstan, and John Riedl\.Item\-based collaborative filtering recommendation algorithms\.In*Proc\. WWW’01*, 2001\.
- Shi et al\. \(2023\)Xiaowen Shi, Fan Yang, Ze Wang, Xiaoxu Wu, Muzhi Guan, Guogang Liao, Yongkang Wang, Xingxing Wang, and Dong Wang\.PIER: Permutation\-level interest\-based end\-to\-end re\-ranking framework in e\-commerce\.In*Proc\. KDD’23*, 2023\.
- Sun et al\. \(2019\)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*Proc\. CIKM’19*, 2019\.
- Tang and Wang \(2018\)Jiaxi Tang and Ke Wang\.Personalized top\-n sequential recommendation via convolutional sequence embedding\.In*Proc\. WSDM’18*, 2018\.
- Tay et al\. \(2022\)Yi Tay, Vinh Q\. Tran, Mostafa Dehghani, Jianmo Ni, Dara Bahri, Sanket Mehta, Zhen Qin, Kai Hui, Zhe Zhao, Jai Gupta, Tal Schuster, William W\. Cohen, and Donald Metzler\.Transformer memory as a differentiable search index\.In*Proc\. NeurIPS’22*, 2022\.
- Tomasi et al\. \(2025\)Federico Tomasi, Francesco Fabbri, Mounia Lalmas, and Zhenwen Dai\.Prompt\-to\-slate: Diffusion models for prompt\-conditioned slate generation\.In*Proc\. RecSys’25*, 2025\.
- van den Oord et al\. \(2017\)Aaron van den Oord, Oriol Vinyals, and Koray Kavukcuoglu\.Neural discrete representation learning\.In*Proc\. NeurIPS’17*, 2017\.
- Wang et al\. \(2021a\)Bo Wang, Fan Sun, Erli Zhang, Zhe Tao, Linju Ju, Yihang Gao, and Xiao\-Yong Zhang\.Sliding spectrum decomposition for diversified recommendation\.In*Proc\. KDD’21*, pages 3696–3704, 2021a\.
- Wang et al\. \(2019\)Fan Wang, Xiaomin Fang, Lihang Liu, Yaxue Chen, Jiucheng Tao, Zhiming Peng, Cihang Jin, and Hao Tian\.Sequential evaluation and generation framework for combinatorial recommender system\.*arXiv preprint arXiv:1902\.00245*, 2019\.
- Wang et al\. \(2017\)Ruoxi Wang, Bin Fu, Gang Fu, and Mingliang Wang\.Deep & cross network for ad click predictions\.In*Proc\. ADKDD’17*\. 2017\.
- Wang et al\. \(2021b\)Ruoxi Wang, Rakesh Shivanna, Derek Cheng, Sagar Jain, Dong Lin, Lichan Hong, and Ed Chi\.Dcn v2: Improved deep & cross network and practical lessons for web\-scale learning to rank systems\.In*Proc\. TheWebConf’21*, 2021b\.
- Wu et al\. \(2024\)Likang Wu, Zhi Zheng, Zhaopeng Qiu, Hao Wang, Hongchao Gu, Tingjia Shen, Chuan Qin, Chen Zhu, Hengshu Zhu, Qi Liu, Hui Xiong, and Enhong Chen\.A survey on large language models for recommendation\.volume 27, 2024\.
- Xi et al\. \(2022\)Yunjia Xi, Weiwen Liu, Jieming Zhu, Xilong Zhao, Xinyi Dai, Ruiming Tang, Weinan Zhang, Rui Zhang, and Yong Yu\.Multi\-level interaction reranking with user behavior history\.In*Proc\. SIGIR’22*, 2022\.
- Xia et al\. \(2023\)Xue Xia, Pong Eksombatchai, Nikil Pancha, Dhruvil Deven Badani, Po\-Wei Wang, Neng Gu, Saurabh Vishwas Joshi, Nazanin Farahpour, Zhiyuan Zhang, and Andrew Zhai\.TransAct: Transformer\-based realtime user action model for recommendation at Pinterest\.In*Proc\. KDD’23*, 2023\.
- Zhai et al\. \(2024\)Jiaqi Zhai, Lucy Liao, Xing Liu, Yueming Wang, Rui Li, Xuan Cao, Leon Gao, Zhaojie Gong, Fangda Gu, Jiayuan He, Yinghai Lu, and Yu Shi\.Actions speak louder than words: trillion\-parameter sequential transducers for generative recommendations\.In*Proc\. ICML’24*, 2024\.
- Zhou et al\. \(2018\)Guorui Zhou, Xiaoqiang Zhu, Chenru Song, Ying Fan, Han Zhu, Xiao Ma, Yanghui Yan, Junqi Jin, Han Li, and Kun Gai\.Deep interest network for click\-through rate prediction\.In*Proc\. KDD’18*, 2018\.

## Appendix ABeam\-Width Replay Analysis

##### Setb=6b=6under accumulated VM\+CL scoring\.

To choosebb, we vary it while holding the accumulated VM\+CL score fixed, withpcontinue≡1p\_\{\\mathrm\{continue\}\}\\equiv 1andF^=0\\widehat\{F\}=0\. Theb=1b=1equivalence check uses 11,965 replayed requests and reproduces the incumbent itemwise greedy decoder’s mean and quartile score statistics to within 0\.15%\. In a separate outlier\-filtered replay of 5,815 requests, Figure[5](https://arxiv.org/html/2609.18148#A1.F5)shows diminishing returns beyond small beam widths; the twob=6b=6configurations in Table[3](https://arxiv.org/html/2609.18148#S5.T3)useb=6b=6\.

![Refer to caption](https://arxiv.org/html/2609.18148v1/figures/beam_width.png)Figure 5:Beam\-width selection on 5,815 outlier\-filtered replayed requests\. Relative accumulated VM\+CL score improvement over theb=1b=1reference, withpcontinue≡1p\_\{\\mathrm\{continue\}\}\\equiv 1,F^=0\\widehat\{F\}=0, and onlybbvarying\. Beam widthb=6b=6lies in the saturated regime and is used in the final two configurations of Table[3](https://arxiv.org/html/2609.18148#S5.T3)\.Increasingbbfrom 1 to 6 under the accumulated VM\+CL score yields narrow online improvements: time spent and video views remain near\-neutral, while likes rise by\+0\.74%\+0\.74\\%and reshares by\+1\.21%\+1\.21\\%relative to the matchedb=1b=1configuration \(Table[3](https://arxiv.org/html/2609.18148#S5.T3)\)\. The wider beam improves reaction metrics but does not by itself convert the offline score improvement into broad consumption improvements\.