ALIGN-HOLD: Experience Alignment for Real-Time Hold Control in Large-Scale Ride-Hailing Matching at DiDi
Zuhao Zhang, Xu Liu, Kai Wan, Zihao Lu, Li Ma, Shuai Li
cs.LG
Quick verdict
DiDi's RM-from-implicit-preference hold control beats EXHOLD in a 28-day A/B (+0.57% TC); strong production evidence, but label design and stats leave gaps.
Summary
ALIGN-HOLD replaces handcrafted reward engineering in DiDi's ride-hailing hold-control system with reward learning from implicit marketplace preferences. From production logs it builds pairwise preferences — completed vs. non-completed interactions sharing an order (10-minute window), a driver (1-hour window), or a local matching-graph snapshot — and trains a Transformer reward model with Bradley–Terry loss, using 1:1:1 balanced view sampling plus 20% model-adaptive hard-pair sampling (70th–90th loss percentiles). The frozen RM then supplies a dense, normalized reward and filters low-identifiability interactions via a label-free score-density band while a four-tier Transformer–LinUCB contextual bandit is trained in the production-calibrated matching simulator; tiers map to hold times through EXHOLD's unchanged guardrail optimizer. In a 28-day randomized A/B test across five Brazilian cities (~100,000 requests/day), ALIGN-HOLD raised trip completion by 0.57% and driver income by 0.64% (both significant), improved driver acceptance (+0.51%), cut passenger cancellations before/after acceptance by 1.85%/2.07%, and left the overall hold ratio nearly unchanged (+0.04%); the RM reaches 0.845 pairwise accuracy. Main limitations: preference labels are purely completion-based, arguably another handcrafted objective; RM evaluation reuses the same label rule; reported statistics lack confidence intervals, multiple-comparison correction, and ablation significance; data, simulator, and code are proprietary.
Review complete
Assessment at a glance
5
risk flags identified
● high 0
● medium 3
● low 2
5 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- Replacing EXHOLD's handcrafted reward with a reward model learned from multi-view implicit preference pairs yields statistically significant online gains over the production policy: TC +0.57% and DI +0.64% in a 28-day, five-city, ~100k-requests/day A/B (Table 1).
- Both RM-guided mechanisms are necessary: removing model-adaptive hard sampling (TC +0.10%, DI −0.26%) or low-identifiability filtering (TC +0.17%, DI −0.37%) eliminates the balanced gains of the full method (Table 3).
- The learned RM orders held-out interactions meaningfully: 0.845 pairwise accuracy and 0.852 macro outcome AUC (Tables 4–5), with monotonic outcome trends across score deciles (Fig. 6).
- The gains come from reallocating hold decisions across pairs, cohorts, hours, and supply regimes rather than increasing overall holding, since the aggregate hold ratio changes by only +0.04% (Figs. 4–5, §A.1).
Strengths
Large-scale randomized online evidence
28-day A/B (May 31–Jun 27, 2026) across five Brazilian cities, ~100k requests/day, with p-values per metric (Table 1: TC +0.57%, p=0.00011; DI +0.64%, p=0.00906), per-city and peak/off-peak consistency (Table 2), and a stable guardrail (hold ratio +0.04%).
Disciplined preference labeling
All three views share one rule (Eq. 4: completed ≻ non-completed) and the paper explicitly refuses to rank failure types (§3.1, §B.1), avoiding re-introduction of manual trade-offs; chronological 7:2:1 splits with trajectory-level isolation prevent preference leakage (§B.1).
Novel, well-diagnosed low-identifiability filtering
Unusually complete implementation detail
Behavioral analyses substantiate the mechanism
Weaknesses
The 'experience alignment' framing outruns the label design
Circular RM evaluation
Unaddressed off-policy confounding in RM training
Statistical reporting gaps
Proprietary everything; simulator validated only under the old policy
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
ALIGN-HOLD — Structured Preliminary Review
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- 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 does ALIGN-HOLD do in DiDi's ride-hailing system?
For each candidate driver–order pair, it decides whether to send the pair into matching immediately or temporarily 'hold' it to wait for a better match. ALIGN-HOLD learns this decision from implicit marketplace preferences: a reward model trained on comparisons of completed vs. non-completed interactions scores each candidate, a contextual-bandit policy maps that score to one of four experience tiers, and a guardrail-constrained optimizer converts the tier into a concrete hold duration.
How much did ALIGN-HOLD improve results in the A/B test?
In a 28-day randomized experiment across five Brazilian cities (about 100,000 passenger requests per day), trip completion rose 0.57% and driver income 0.64% versus the production EXHOLD policy (both p<0.01). Driver acceptance rose 0.51%; passenger cancellations before and after driver acceptance fell 1.85% and 2.07%. Driver cancellations after acceptance and call-acceptance time improved directionally but not significantly, and the overall hold ratio was essentially unchanged (+0.04%).
Is ALIGN-HOLD actually deployed?
Yes. After the online experiment, the paper reports the system was fully ramped up and is serving real traffic in DiDi's Brazil marketplace. The reward model and policy training run offline, so online latency is unaffected; production monitoring covers tier distributions, hold ratios, core marketplace indicators, RM-score drift, with versioned artifacts and rollback controls.
What is 'low-identifiability filtering' in this paper?
During policy training, some interactions — for example, very short, cheap pickups — have outcomes driven by incidental factors rather than match quality, so their completion/cancellation labels are weakly attributable. ALIGN-HOLD detects a dense band in the reward model's score distribution where different outcomes overlap, and excludes interactions in that band from policy updates, without using any outcome labels. Held-out diagnostics show these interactions concentrate among non-completed trips with unusually short pickup and waiting times.
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 1–6 are embedded graphics; only their captions and in-text statistics were extractable, so figure-internal details (e.g., exact curve shapes in Figs. 3, 5, 6) could not be independently verified.
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.