Alibaba's ReSID recovers long-tail items in generative recommendation with beam 20 beating beam 200

ReSID: Rethinking Long-Tail Retrieval in Generative Recommendation through Item Reachability

Muyu Zou, Shunye Tang, Yi Xu, Haibo Xing, Yandong Wang, Lingyu Mu, Jinxin Hu

cs.IR

2026-11-12

ReSID tolerates 3 mismatched tokens in SID matching and debiases beam search, doubling tail R@200 at a tenth of the beam budget; deployed on Alibaba ad traffic with +1.3% revenue.

What problem this solves

Generative recommendation (GR) recasts retrieval as sequence generation: each item is compressed offline into a string of discrete tokens called a Semantic ID (SID), an autoregressive model generates SIDs online, and complete sequences are matched exactly against the catalog. The pipeline works for popular items and collapses on the tail. Interaction distributions on real platforms are heavily skewed; in the paper's in-house data, long-tail items are more than half the catalog but receive only 20-22% of training interactions.

Prior work on long-tail GR stays on the training side: CRAB splits overheated codes, Ghost applies unlikelihood training so tail items inherit head prefixes. The representation changes; inference does not, remaining standard beam search (keep only the top-B prefixes per step) plus exact full-SID matching. The paper asks the earlier question: do learned tail signals survive to final retrieval at all?

Mostly no. With beam 20, two decoding steps in, 62.5% of tail targets have already fallen out of the beam; after all six steps, 90.2% are gone for good (head: 58.7%). Even the relaxed criterion of keeping one prefix completable within three mismatches fails for half the tail by step 4. The paper names this the long-tail reachability bottleneck.

Method

ReSID modifies three stages of the pipeline.

The paper's own analysis supplies the quantitative case: tolerating 1-3 mismatches lifts tail reachability from 9.75% to 41.56% (head: 59.77%), and one third of missed tail targets share at least half their tokens with some decoded sequence. But recoverable prefixes get pruned upstream, so guided decoding and tolerant matching have to arrive as a pair.

Results

SettingMetricResult
In-houseTail R@20036.62 (Ghost 18.94)
In-houseOverall R@20048.28 (TIGER 37.65)
Two public datasetsTail Recall / NDCG+41.2% / +36.6% over Ghost
Online A/B (15% traffic per arm, 12 days)Ad revenue / CTR / GMV+1.3% / +1.1% / +0.8%
Online A/BLong-tail exposure share+4.0%

Beam widths tell the cost story: ReSID runs 20 against the baselines' 200 on public datasets, and 100 against 300 in-house. End-to-end latency is 1/2 to 1/8 of the baselines, throughput is 3.7-16.7× higher, and online mean response time is 10.3 ms.

Ablations pin the dominant gain on partial matching: removing it drops Office tail R@200 by 66.1% and crashes tail exposure coverage from 99.31% to 63.73%. Removing popularity calibration or user-semantic alignment costs 8.7% and 6.2% of tail R@200; swapping back to sequential SIDs costs 17.3%. Mechanistically, 51.26% of tail hits on Industrial (66.09% on Office) depend strictly on partial matching. Hyperparameters are insensitive, and tolerance saturates at r=3.

Why it matters

For industrial recommendation engineers the paper pays out twice. First, diagnosis: it decomposes why GR loses the tail into three separately measurable stages (SID structure, decoding pruning, matching rule) with instrumented localization; even without the full framework, the parallel-over-sequential and tolerant-matching findings travel on their own. Second, economics: tail recall is normally bought with wider beams, and here better tail results come from a beam ten times narrower, with latency and throughput improving at the same time. The system serves main traffic in production, with code and checkpoints released.

The components themselves are familiar; popularity debiasing is logit adjustment moved to inference time. The contribution is placing known tools where a systematic analysis says they belong, then validating the assembly with online revenue.

Limitations

No dedicated limitations section; these read out of the numbers:

Terms

Source

What people are saying

Related papers

All paper explainers