Offline-compressed reusable history tokens lift ad conversion value 4.65% in production CTR ranking

X-split Argus: Learning Reusable History Tokens for Efficient CTR Prediction

Alexander Ploshkin, Nikita Stepanov, Alexey Pismenny, Viacheslav Ovchinnikov, Nikolay Savushkin

cs.IR

2026-11-18

X-split ARGUS compresses user histories into reusable offline tokens and cross-attends to them online, gaining 4.65% conversion value and 2.96% clicks in A/B at ~100k QPS.

What problem this solves

CTR models live or die on user representations that capture both long-term preference and fresh intent, and the sequential-modeling line of work (HSTU, ARGUS) keeps showing that longer histories help. The trouble is serving: a ranking request scores a batch of candidate ads, and re-encoding a long history per candidate burns GPU compute and memory traffic that a latency budget measured in hundreds of milliseconds cannot absorb. Once history encoding depends on the candidate, the computation cannot be shared across ads either.

Prior systems each took a slice of the problem. TransAct and LLaTTE pair precomputed user embeddings with an online pass over recent actions; VISTA learns reusable summary tokens; TM20K distills from a teacher that keeps the full sequence. This paper pins the question down: if the past is compressed into a small set of candidate-independent tokens, how much of the teacher's uncompressed-history benefit survives? The answer is quantified on a live RTB system.

Method

X-split ARGUS splits user history into two branches.

Training is end to end over both encoders, with two supervision sources: an autoregressive binary cross-entropy on observed clicks, where one shared pass over the sequence covers all prediction positions; and distillation from a larger teacher that scores each candidate with the uncompressed history, matched by a separate head under an L1 loss on raw logits. The key design choice is that distillation gradients flow through the decoder back into the compressor, so the teacher's judgments effectively decide what the K tokens keep. The teacher trains separately and never runs online. Two smaller touches: unclicked impressions get their training target lifted from zero to a small ε (asymmetric label smoothing, on the theory that the user may never have seen the ad), and K and M are direct knobs for the serving budget.

Results

Offline evaluation uses internal RTB logs: 365 days, about 10^9 training targets, histories capped at 8,192 events, global temporal split.

ConfigurationNE ↓Gain vs baselineTeacher gain retained
Context-aware ARGUS (production baseline)0.720400
X-split ARGUS0.7146+0.81%71.6%
Teacher (full history, unservable)0.7123+1.14%100%

On this test set a 0.1% NE change is statistically significant. The compressed student keeps 71.6% of the teacher's improvement, the single most important number in the paper.

Ablations: cutting K to 2 costs 0.19% while raising it to 32 buys only another 0.08%; shrinking the recent window M from 256 to 8 costs 0.34%; deepening both encoders from 6 to 10 layers lowers NE throughout; smoothing negative targets is a free 0.18% at K=2 with no change in training or serving cost.

The deployment covers the largest ad surface of the RTB system, about half its traffic at roughly 100k requests per second. A seven-day A/B test (6% of traffic per arm) gave +4.65% ACV and +2.96% clicks, p<0.001, and a reverse A/B a month later reproduced the gains. Cost: 15% more online GPUs of the same type within the same latency budget, plus daily precompute and a several-hundred-GB embedding store. A fixed checkpoint loses about 0.05 percentage points of relative NE advantage per day of staleness, so production fine-tunes daily on the last three days of data.

Why it matters

For ranking engineers this is a copyable serving pattern: long history leaves the request path entirely, recent-history encoding is shared within a request, and each candidate pays only for cross-attention. The 71.6% retention says the student sits not far below the teacher's ceiling, while the teacher itself could never serve under production constraints. A 4.65% ACV lift is a number industrial ranking teams have to take seriously, and the explicit cost accounting (15% more GPUs, same latency) makes the trade concrete. Tunable K and M let one architecture stretch across budgets.

Against the existing VISTA and TM20K line, the mechanism here is mostly combination and engineering; putting it into a 100k-QPS production surface with full cost accounting is exactly the part most papers lack.

Limitations

Stated by the authors: the depth ablation deepens both encoders at once, so capacity and receptive field are confounded; the teacher serves only as a quality reference; the linear staleness decay commits them to a permanent daily fine-tuning pipeline.

Beyond that list: the offline comparison has exactly one baseline, the in-house Context-aware ARGUS, with no head-to-head against VISTA or TM20K on the same data, so 71.6% retention is a within-family number; nothing explains the mechanism that turns a 0.81% offline NE gain into 4.65% online ACV; and with internal data plus hyperparameters relegated to an appendix, reproduction is off the table. The 15% figure covers online GPUs only, with precompute, storage, and fine-tuning described as small relative to serving but never quantified.

Terms

Source

What people are saying

Related papers

All paper explainers