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, et al.
cs.LG · cs.IR
Quick verdict
LIGE-GR upgrades rankers to listwise generative recommendation in place, lifting watch time ~1% for ~10% compute; internal A/Bs only, unverifiable outside Meta.
Summary
The paper proposes LIGE-GR, a Meta framework that upgrades production itemwise ranking recommenders toward LLM-inspired listwise (generative) recommendation without rebuilding the stack. It adds three components to an existing recommender: a lightweight four-head, four-layer causal Transformer context-aware predictor conditioned on the already-selected prefix; a listwise value model that weights items by continuation probability; and a beam-search decoder ('Palette') with a closed-form, duration-aware future-value estimate. The design strictly generalizes the incumbent itemwise system, enabling configuration-level reversion and per-request latency fallback. On Instagram Reels and Facebook Video, the b=1 configuration improves time spent by 1.14% and 0.72% respectively (p<0.001), with ~10% additional inference resources relative to the context-free ranker and ~7%/2.2% latency overhead; on Reels, a b=6 duration-aware configuration adds a further +0.69% time spent. Counterfactual list-composition analysis shows broader topic/creator diversity and more exploration, but a small freshness regression (-0.66% under-72-hour videos). Main limitations: no public code or data, all evidence is internal A/B testing without confidence intervals or multiple-comparison discussion, the 'RL-based' framing overstates what is essentially heuristic beam search, and key choices (trimmed candidate pool, beam width, continuation calibration) lack isolated online ablations.
Review complete
Assessment at a glance
5
risk flags identified
● high 1
● medium 3
● low 1
5 strengths · 5 weaknesses · 5 author questions
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- LIGE-GR upgrades itemwise ranking systems into listwise generative recommenders by adding a context-aware predictor, a listwise value model, and a beam-search decoder, while strictly generalizing the incumbent system so it can be exactly recovered by configuration.
- On Instagram Reels and Facebook Video, the b=1 base configuration improves time spent by 1.14% and 0.72% (p<0.001) with only ~10% additional inference resources relative to the context-free ranking component.
- The b=6 duration-aware configuration (ListVMgolden plus the duration-aware future-value estimator) adds +0.69% time spent and +1.82% video views over the b=1 base on Instagram Reels, showing listwise value modeling and beam decoding contribute beyond contextualized prediction alone.
- Generated lists improve ecosystem diversity and exploration (topic entropy +1.14% to +2.37%, familiar-creator content -6.23%) at the cost of a small freshness regression (-0.66% of videos under 72 hours old).
Strengths
Strict-generalization design makes the upgrade operationally safe
Sections 2.3/4.2 and Algorithm 2: setting CA to CF, b=1, pcontinue=1, and the future-value term to 0 exactly recovers the incumbent decoder; config-level reversion without retraining plus a per-request latency-budget fallback are rare, deployment-grade contributions that directly address the replacement-risk argument in Section 1.
Large-scale online validation with guardrails and honest cost accounting
Tables 2-3: seven-day A/Bs on two production surfaces with p<0.001 on core consumption metrics, sessions/DAU reported as guardrails, and measured overheads stated up front (~10% of CF compute, +7% Reels per-request latency, ~2.2% FB Video latency), including the admission that the two latency figures are not directly comparable.
Ecosystem evaluation goes beyond engagement metrics
The listwise value model addresses a real modeling gap, backed by prediction-quality evidence
Serving design is concretely engineered, not hand-waved
Weaknesses
The 'RL-based decoder' framing overstates the method
Statistical reporting is thin for the centrality of the claims
No external reproducibility path
Key design choices lack isolated online ablation, and the offline proxy demonstrably mispredicts online gains
Facebook Video shows an unexplained significant regression, and its configuration is untested beyond b=1
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 5 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What is LIGE-GR?
LIGE-GR is a framework from Meta that upgrades an existing ranking-based recommender system (which scores each item independently and sorts them) into a listwise, generative recommender. It adds a lightweight Transformer that re-scores each candidate conditioned on the items already chosen for the list, a list-level value model weighted by the probability the user keeps watching, and a beam-search decoder that builds the whole list jointly, all without rebuilding the production system.
How much did LIGE-GR improve recommendation quality?
In Meta's online A/B tests, the base configuration increased user time spent by 1.14% on Instagram Reels and 0.72% on Facebook Video (p<0.001), with additional inference cost of roughly 10% of the existing ranking model's compute and a 7% (Reels) or 2.2% (Facebook Video) increase in serving latency. A stronger beam-width-6 configuration added a further 0.69% time spent on Reels.
How is LIGE-GR different from OneRec or other generative recommenders?
OneRec-style systems replace the entire recommendation funnel with a single generative model, which is technically risky and organizationally disruptive for mature products. LIGE-GR instead upgrades the existing ranking stage in place: it reuses the current ranker's computation, keeps a fallback path that exactly reproduces the old behavior, and can be reverted by a configuration switch rather than a migration.
Is LIGE-GR's code or data publicly available?
No. The paper is an industrial report from Meta with no released code, datasets, or public benchmark results; all evidence comes from internal A/B experiments on Instagram Reels and Facebook Video, so external replication is not currently possible.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Figures 1-5 are available only as captions in the extracted text; their plotted content could not be verified (e.g., the beam-width saturation curve in Figure 5).
- Mathematical notation is partially garbled by text extraction (the future-value estimator appears as 'bF' where F-hat is meant, exponents and subscripts are flattened); interpretations of Eqs. 10-14 assume the standard intended forms.
- The metadata lists 9 authors while the full text lists roughly 70 authors from Meta; the review treats the full-text author list as authoritative.
- Only the abstract and embedded full text were used; no external sources were consulted.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.