2,764 Parameters vs. 302M: Serving Sequential Recommenders on 20-Item Histories Without User Caches

Closing the Long-Short View Gap in Sequential Recommendation without Cached History

Lingfeng Shi, Chengkai Huang, Lina Yao, James Caverlee

cs.IR

2026-09-06

Cosine scoring plus softmax1 retraining, then a 2,764-parameter fix, serves 20-item histories at 95.9% of cached-baseline quality with zero per-user state.

What problem this solves

Sequential recommenders such as SASRec and HSTU are trained on long user histories, 202 to 402 interactions in this paper's setup, but latency budgets at serving time only allow the most recent 20. Naive truncation hurts: HSTU on Taobao drops from 0.5231 to 0.4446 NDCG@10, a 7.9-point loss. The standard remedy compresses each user's history into a persistent state retrieved at inference. PersRec, the strongest such baseline, requires full-parameter fine-tuning, a 16 KB per-user cache (1.44 GiB across Taobao users), and a 190-token prefill per request. Cold-start users get nothing, since there is no history to cache. The paper asks whether the backbone itself can be trained to tolerate short views, with no serving-time state at all.

Method

The diagnosis comes first. Two structural flaws:

Stage one retrains the long-view backbone with both fixes. Scoring switches from dot product to temperature-scaled cosine similarity (τ = 0.07), so ranking depends on direction alone. And the softmax gains a constant 1 in its denominator (softmax1), letting the model output near-zero total attention when nothing visible is informative, instead of being forced to spend a full unit of probability.

Stage two adapts to the short view by fine-tuning only the backbone's native bias and LayerNorm parameters, following BitFit. No new modules; on HSTU/Taobao that is 2,764 trainable parameters. The objective combines pointwise BCE on short views with KL distillation from the long-view teacher.

Why only bias and LayerNorm? Shortening the sequence disturbs LayerNorm statistics first, and these parameter types exist in every standard backbone, so the recipe stays architecture-agnostic.

Results

NDCG@10, sampled ranking with 1,000 negatives:

Serving inputSASRec·ML-10MHSTU·TaobaoSASRec·XLong
Original backbone, training-length0.23290.52310.2709
Original backbone, latest 200.21570.44460.2380
Proposed backbone, latest 200.25350.53960.3391
short-view adaptation0.26330.54670.3539
Proposed backbone, training-length0.26960.57640.3656

Three takeaways. Zero-shot, the retrained backbone on 20 items beats the original backbone's full-length inference in five of six settings (the exception is HSTU/XLong, 0.3324 vs 0.3441). After adaptation, SASRec/Taobao Hit@10 (0.6810) exceeds its own full-length ceiling (0.6677). Against PersRec: 2,764 versus 302,825,997 trainable parameters, 1.07 h versus 8.34 h of training, zero serving-time state, and 95.9% of its NDCG@10 (0.5467 vs 0.5703); on ML-10M and XLong the method wins outright.

A rolling evaluation over nine future targets with only 20 items visible is closer to real request streams: on XLong/HSTU it leads PersRec at all nine positions (average 0.3780 vs 0.3571); on Taobao/SASRec it recovers 97.4%.

Why it matters

This is a one-time, training-side change that removes serving-time state entirely. Cold-start users are covered by construction, and an adaptation with a few thousand parameters can be rerun frequently on fresh data. For recommendation engineers the menu is concrete: cosine scoring and softmax1 are backbone-level changes, and bias/LayerNorm-only tuning is a universal trick, validated on two backbones.

Limitations

The authors list longer sequences, semantic and multimodal signals, and LLM-based generative recommenders as future work. Beyond that: peak accuracy still belongs to the cached approach, and 95.9% is the evidence, so where caches are feasible the paper does not claim victory. The optimal loss balance shifts by dataset (BCE-only wins on HSTU/Taobao, distillation-only on SASRec/XLong) with no unified rule offered. Evaluation uses 1,000 sampled negatives rather than full-catalog ranking, which inflates absolute numbers, though all methods share the protocol. And on ML-10M, where norm correlates negatively with frequency, the popularity-shortcut explanation runs backwards and is argued most thinly.

Terms

Source

What people are saying

Related papers

All paper explainers