Listwise Generation Raises Short-Video Engagement in Existing Rankers

Most large-scale recommenders still score candidates independently, then greedily assemble a feed. That procedure is efficient and compatible with mature serving stacks, but it cannot directly ask whether the combination of items is good: adjacent videos may be repetitive, creators may be over-concentrated, and a locally high-scoring placement can constrain the rest of the list. LIGE-GR reframes ranking as sequential list construction while retaining the incumbent ranker as part of the system.

Core Contribution

The paper’s contribution is an upgrade path rather than a wholesale generative-recommender replacement. LIGE-GR moves three components from itemwise to listwise operation: the ranker is made context-aware, the objective scores a complete sequence rather than summing independent item scores, and greedy decoding is replaced by an RL-based search procedure. The authors’ central claim is that these changes let a production ranker optimize interactions among placements without replacing its models, value functions, or serving infrastructure.

That distinction matters operationally. Prior generator-evaluator systems commonly add a dedicated reranking stage or generate then score complete candidate slates. LIGE-GR instead interleaves construction and evaluation inside the ranking stage, scoring partial lists and pruning weak prefixes before completion. It is therefore aimed at systems where latency, reliability, and ownership boundaries make a clean-slate LLM-style recommender impractical.

Technical Approach

The architecture retains a context-free module that scores every candidate from user features through an interaction network and per-task heads. Its intermediate candidate representations are passed to a lightweight context-aware module: a four-layer causal Transformer. At position $t$, this module conditions its prediction for the current candidate on items placed at earlier positions, allowing the score to change with the evolving list rather than the user alone.

Figure 3 makes the migration path clear: the incumbent context-free scorer remains the first component, while the Transformer refines its per-task predictions using only preceding placements. The task heads share weights across candidates and positions, which keeps the context-aware addition comparatively small.

A listwise value model then evaluates the sequence as a whole. The decoder uses RL-based beam search rather than committing to the single greedy path through the position-by-candidate lattice. In replay, increasing beam width improves the accumulated VM+CL score relative to a width-1 reference, but the returns flatten: 3.16% at width 2, 5.63% at width 4, 6.72% at width 6, and 7.42% at width 8. The authors select $b=6$ because it is already in that flatter regime.

Results and Analysis

The production evidence is more useful than the replay curve alone. On Instagram Reels, the paper reports a 1.14% increase in time spent; on Facebook Video, it reports 0.72%. In the beam-width analysis, moving from width 1 to 6 left time spent and video views near neutral while increasing likes by 0.74% and reshares by 1.21%. That is a useful warning against treating an offline value gain as a direct proxy for broad consumption: a wider search improved reaction metrics, but not the main consumption measures in that comparison.

The paper also examines counterfactual paired requests on Reels rather than limiting evaluation to engagement. Across 13,197 paired requests from 7,110 logged users, it reports more topical variety and exploration: topic entropy rises 1.14% to 2.37%, distinct topic clusters 1.34% to 2.40%, distinct creators 0.66%, and interest-matched topic content 1.65%. Repetition declines, with the longest same-topic run down 4.03% to 7.4% and adjacent-video cosine similarity down 3.68%.

Those gains come with a trade-off. Same-creator concentration falls 2.98%, familiar-creator content falls 6.23%, and videos under 72 hours old fall 0.66%. The authors characterize the latter as a composition regression, not a quality improvement. The evidence supports the narrower conclusion that listwise construction can produce modest online engagement gains while redistributing exposure toward broader topics and creators. It does not establish that the method is the final form of generative recommendation: the conclusion explicitly notes that the current system handles input pools on the order of hundreds and would need technologies such as semantic IDs for much larger candidate spaces.

Limits in Practice

The online tests cover two short-video surfaces, so the reported gains are meaningful but do not establish transfer to other recommendation domains. The framework also depends on a learned listwise value and an RL decoder, making it sensitive to how those objectives encode product trade-offs. The observed freshness decline and near-neutral consumption response in the beam analysis show why list composition should be monitored alongside aggregate engagement.