Source-linked AI summary
LIGE-GR: A Smooth Leap from Ranking to Generative Recommendation in the LLM Era
Venkat Srinivas, Chenzhang He, Sam Woodmansee, Shawn Lian, Wenjie Hu, Renjie Jiang, Ziheng Huang, Xinyuan Zhang, Zhihao Zheng, Zhuoran Yu, Rui Li, Lei Yuan, Ziwei Li, Jimmy Jia, Mert Terzihan, Ekrem Kocaguneli, Yiming Liao, Zhichen Zhao, Yue Yin, Yue Weng, Wanlin Ma, Xufeng Cai, Weimiao Wu, Yezhou Huang, Du Zhang, Yukun Ding, Aaron Johnston, Yueming Wang, Zhaojie Gong, Yuting Zhang, Serena Li, Adithya Ganesh, Boying Liu, Haichuan Yang, Xialu Li, Matt Ma, Qunshu Zhang, John Joshua Miller, Praveen Rathinavelu, Cheng Huang, Aadhar Sachdeva, Josh Karns, Andres Aaron Gutierrez, Neil Agarwal, Gustas Pladis, Vladimir Batygin, Gopal Ray, Aditya Priyadarshi, Shantanu Patil, Zhe Wang, Penny Pan, Yiping Han, Arun Singh, Guangdeng Liao, Bi Xue, Xinyao Hu, Yang Song, Yisong Song, Meihong Wang, Haotian Wu, Deepak Agarwal, Ji Liu
TL;DR
LIGE-GR addresses how to bring sequence-level generation into mature recommender systems without the risks of wholesale replacement. It generalizes itemwise recommendation into listwise generation through contextual ranking, listwise value modeling, and RL-based decoding, improving time spent on Instagram Reels and Facebook Video.
Problem
It remains unclear how to incorporate listwise generation into mature recommender systems while preserving their accumulated technical, product, serving, and organizational investments.
Method
LIGE-GR upgrades existing itemwise recommenders with a context-aware ranking module, a listwise value model, and an RL-based sequence decoder.
Results
LIGE-GR increases time spent by 1.14% on Instagram Reels and 0.72% on Facebook Video against strong, optimized baselines.
Takeaways & Limitations
The framework provides a practical, low-resource transition from mature itemwise recommendation toward generative listwise recommendation while preserving compatibility with existing systems.
Takeaways & Limitations
The current framework handles candidate pools only on the order of hundreds and requires further integration for truly end-to-end recommendation.
Abstract
from arXiv · showhide
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 sequence-level generation and optimization from the LLM paradigm 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 propose LIGE-GR, a listwise generation and evaluation recommendation framework that upgrades from a traditional ranking system based on itemwise recommendation toward a generative recommendation paradigm. Instead of rebuilding the entire recommendation stack from scratch, LIGE-GR generalizes 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 validate LIGE-GR in short-video recommendation on Instagram Reels and Facebook Video. On these recommendation surfaces, LIGE-GR improves time spent by 1.14 percent on Instagram Reels and 0.72 percent on Facebook Video, while requiring only modest additional inference resources.
1 Introduction
Mature recommenders typically optimize items independently, while LIGE-GR upgrades them toward listwise generation without wholesale replacement. The framework improves time spent on two short-video surfaces with modest serving overhead.
- Existing paradigm: Traditional recommenders independently score candidate items, then sort their scalar values into an ordered list.Diversity and integrity adjustments may add sequence awareness, but are generally rule-based heuristics rather than learned listwise optimization.
- Motivation: Listwise recommendation evaluates the sequence jointly, requiring contextual value prediction, whole-sequence evaluation, and feasible sequence search.This contrasts with LLM-style generation, where each output is conditioned on previously generated context and judged by the sequence as a whole.
- Industrial barriers: Wholesale replacement remains difficult because mature systems contain accumulated model, product, serving, business, and organizational investments.The paper identifies system risks around baseline recovery, rollback, and reliability, alongside disruption to team boundaries, ownership, and planning.
- Proposed framework: LIGE-GR generalizes the existing itemwise system through additive, revertible upgrades to ranking, value modeling, and decoding.It adds a listwise ranking module, extends itemwise values to listwise values, and replaces greedy selection with an RL-based sequence decoder.
- Empirical validation: 1.14% higher time spent on Instagram Reels and 0.72% on Facebook Video validate LIGE-GR against strong, optimized baselines.On Instagram Reels, the framework adds roughly 10% inference resources for the context-free ranking component and approximately 7% end-to-end latency.
2 LIGE-GR: the Generative Paradigm
The generative reformulation starts from ordered recommendation lists and converts itemwise scoring into listwise sequence optimization. LIGE-GR retains the incumbent system while introducing contextual prediction, listwise valuation, and sequence decoding.
- Problem statement: Each request produces a feasible ordered list of distinct candidate items, typically with around 10 displayed positions.The objective is to maximize user experience and product quality as items are sequentially exposed.
- Itemwise recommendation: The incumbent system predicts engagement signals independently, combines them into item scores, and returns the top T items after product adjustments.Its control layer can enforce diversity and hard restrictions through mechanisms such as gap demotion, DPP scoring, and business rules.
- Listwise reformulation: LIGE-GR changes the objective from itemwise scores to a sequence-level value using predictions conditioned on preceding selected items.The context-aware predictor can represent repetition, saturation, complementarity, diversity, and user fatigue effects within a list.
- Three component upgrades: The upgrade has three parts: context-free to context-aware ranking, itemwise to listwise value modeling, and greedy to RL-based sequence decoding.These changes correspond to the three capabilities required for sequential optimization while preserving the surrounding recommendation structure.
- Compatibility: Reverting all three upgraded components recovers the incumbent recommender, enabling a smooth upgrade rather than a disruptive replacement.This strict generalization supports incremental introduction within mature systems.
3 LIGE-GR Design
LIGE-GR upgrades itemwise recommendation into listwise generation by adding context-aware prediction, listwise value modeling, and Palette decoding while preserving the mature system’s structure. Its objective accounts for sequence context and continuation probability, and its decoder searches feasible lists with future-value estimation.
- 3.1 Listwise Model: Context-Aware and Context-Free Predictors: Context-aware prediction captures listwise effects including repetition, saturation, complementarity, diversity, and user fatigue.The prediction for each candidate depends on the user, candidate, and preceding items in the requested list.
- 3.1 Listwise Model: Context-Aware and Context-Free Predictors: LIGE-GR preserves the existing itemwise predictor while adding a context-aware causal decoder that conditions each candidate’s predictions on previously selected items.The refinement module uses intermediate context-free representations and a lightweight four-layer, four-head causal Transformer.
- 3.2 Listwise VM: From Itemwise Value to Listwise Value: LIGE-GR weights each item’s value by the probability that the user reaches its position, producing a more faithful estimate of whole-list expected value.The listwise value model combines context-aware item value with control-layer adjustment and continuation probability.
- 3.3 Palette Decoder: Palette constructs lists sequentially with RL-based beam search, retaining the top-b feasible extensions instead of evaluating every possible full list.The decoder expands retained prefixes using ListVMgolden and applies control-layer constraints at each position.
- 3.3.1 Future-Value Estimation: The future-value component distinguishes Palette from standard beam search, while its lightweight estimate uses quantities already available during decoding.Richer learned value models or Monte Carlo tree search are left for future work.
- 3.3.1 Future-Value Estimation: Duration-aware estimation reduces the step estimator’s bias against prefixes ending in long items while retaining continuation information.It rescales continuation probability using the average duration of selected items rather than repeatedly applying the current item’s full duration.
4 Serving and Efficiency Optimization for LIGE-GR
LIGE-GR augments the existing ranking service with cached context-free computation and lightweight context-aware decoding. Serving optimizations restrict rescoring, batch beam evaluation, right-size generation, and preserve per-request fallback and exact global reversibility.
- 4.1 Serving LIGE-GR: The serving path has two phases: unchanged context-free embedding computation followed by context-aware list construction over retained beam prefixes.Algorithm 2 invokes Palette for the second phase using cached representations and the configured beam width and list length.
- 4.1 Serving LIGE-GR: LIGE-GR reuses cached context-free representations before autoregressive context-aware decoding, so the expensive forward pass runs only once per request.The context-aware module is a lightweight four-head, four-layer causal GPT-style decoder.
- 4.2 Reliability and Reversibility: If decoding exceeds the latency budget or a context-aware call fails, the request automatically falls back to itemwise behavior.Globally, disabling the context-aware path recovers the incumbent system by configuration switch, without retraining.
- 4.3 Efficiency Optimization: Rescoring only roughly the top third of context-free candidates bounds the context-aware overhead because lower-ranked candidates have low selection probability.The context-free ranking supplies a high-quality itemwise ordering used to trim the rescoring pool.
- 4.3 Efficiency Optimization: 10% additional inference resources accompany the base configuration across both evaluated settings, with approximately 7% Instagram Reels latency and 2.2% Facebook Video latency increases.The latency definitions and baselines differ across settings, so the magnitudes are not directly comparable.
5 Experimental Results
LIGE-GR improves prediction quality, online consumption outcomes, and list composition across Instagram Reels and Facebook Video, while adding modest serving costs.
- 5.1 Context-Aware vs. Context-Free Prediction: CA improves all six displayed Instagram Reels task families and all five displayed Facebook Video task families in normalized entropy.The complete Facebook Video evaluation improves prediction quality on 15 of 17 tasks, with regressions of at most 0.07% on two minor tasks.
- 5.3 Instagram Reels Online Validation: b = 6 Duration-Aware Configuration: The duration-aware b = 6 configuration improves all four reported Instagram Reels metrics relative to the b = 1 base configuration.The wider beam alone mainly improves likes and reshares, while requiring approximately 20% of the context-free component’s resources; comparable end-to-end latency is unavailable.
- 5.4 Impact on the Ecosystem: Counterfactual diagnostics compare 13,197 paired requests from 7,110 known logged users over a three-day window, with effects assessed at the 95% level.The request-level sample is skewed toward more-active viewers, and 1,498 requests without viewer IDs are conservatively counted as one user each.
- 5.4 Impact on the Ecosystem: LIGE-GR increases topic and creator diversity, reduces repeated-topic runs and adjacent-item similarity, and broadens exposure beyond familiar content.These gains involve tradeoffs: some immediate affinity matching is reduced, and the fraction of very fresh videos decreases.
6 Related Work
Related work spans itemwise ranking, listwise reranking, direct slate generation, and LLM-inspired recommendation. LIGE-GR differs by emphasizing an upgrade path from mature itemwise systems to listwise generation.
- LLM-Inspired Generative Recommendation: The paper’s distinguishing perspective is an upgrade path that preserves accumulated mature-system improvements while enabling richer listwise optimization.Its emphasis extends beyond a single model or algorithm to deployment compatibility in an iteratively refined industrial system.
- Itemwise Recommendation: Itemwise recommendation independently ranks items using user-item relevance, whereas LIGE-GR conditions each candidate on items already selected for the list.This distinction targets listwise effects such as repetition, saturation, complementarity, diversity, and user fatigue.
- Traditional Listwise Recommendation: Traditional listwise methods model cross-item context through reranking or generator–evaluator frameworks that score or select among candidate slates.Score-refinement methods sort contextualized scores, while generator–evaluator methods generate slates and select among them.
- Generative and Autoregressive Slate Construction: Generative slate-construction methods directly produce lists, including stochastic generators, autoregressive decoders, and approaches based on decomposed slate rewards.These methods differ in how they represent prior context, model inter-item interactions, and accelerate decoding.
- LLM-Inspired Generative Recommendation: LLM-inspired recommendation includes item tokenization, generative retrieval, unified text-to-text tasks, and wholesale funnel replacement.LIGE-GR instead reaches LLM-style listwise generation by upgrading the existing mature recommendation system rather than replacing the entire funnel.
7 Conclusion and Future Work
LIGE-GR provides a practical, low-resource framework for incrementally upgrading mature itemwise recommenders into generative, listwise systems. The framework preserves existing infrastructure while leaving room for richer architectures, objectives, decoding, and larger candidate spaces.
- LIGE-GR upgrades traditional itemwise recommendation into a generative, listwise framework without requiring disruptive replacement of existing infrastructure, models, or organizational ownership.It adds a context-aware module, converts the value model to listwise scoring, and uses an RL-based decoder.
- The framework is designed for industrial environments where migration effort, reliability, latency constraints, and cross-team ownership shape deployment decisions.
- LIGE-GR is a transition framework rather than the final form of generative recommendation, with potential improvements from more expressive architectures, richer list-level signals, and advanced decoding or reinforcement learning.
- The current framework handles candidate pools on the order of hundreds, so truly end-to-end recommendation requires technologies such as Semantic IDs for much larger candidate spaces.
- Streaming inference, early result emission, and dynamically scaled context-aware computation offer future opportunities to optimize the full serving stack.
A Beam-Width Replay Analysis
The replay analysis examines wider beam search under a fixed accumulated VM+CL score. Increasing beam width from 1 to 6 produces narrow reaction-metric gains but little change in consumption metrics.
- The b = 1 replay reproduces the incumbent itemwise greedy decoder’s mean and quartile score statistics to within 0.15% on 11,965 replayed requests.
- +0.74% likes and +1.21% reshares result from increasing b from 1 to 6, while time spent and video views remain near-neutral.These comparisons are relative to the matched b = 1 configuration under the accumulated VM+CL score.
- The wider beam improves reaction metrics but does not by itself convert offline score improvements into broad consumption improvements.