OneBid: A Unified Auto-Bidding Foundation Model for Diverse oCPX Advertising Scenarios
Yewen Li, Yitian Li, Pengfei Lv, Xialong Liu, Peng Jiang, Qingpeng Cai
cs.LG · cs.AI
Quick verdict
Claims first auto-bidding foundation model; strong production A/B gains (+2.2% ADVV), but evidence is proprietary and offline checks are critic-based proxies.
Summary
OneBid is a Kuaishou production system that unifies auto-bidding across seven oCPX conversion scenarios (purchase, ROAS, activation, form submit, leads submit, etc.) in a single Decision Transformer-based 'foundation model', replacing one-model-per-scenario pipelines. Pre-training on ~70M proprietary bidding transitions uses two conditioning tokens — Return-to-Go (built from cost pacing as a dense proxy) and a novel Cost-to-Go for cost-ratio control — plus auxiliary IQL-style Q/V heads and a sequence-level Mixture-of-Experts (S-MoE) that routes whole sequences by scenario identity to keep per-request computation fixed. Post-training introduces CROP, a GRPO-style policy improvement that replaces online rewards with a learned critic scoring sampled candidate actions, using group-relative advantages, clipping, and KL regularization toward the pretrained policy. The authors report power-law scaling (R²=0.82), offline gains of CROP over AWR/GAS on Q-value improvement (1.90% vs 0.12%/0.06%), and online A/B tests (33% traffic, at least 5 days) showing +2.2% overall ADVV and up to +13.1% in ROAS, fully deployed since May 2026. Main limitations: all data proprietary and non-reproducible; offline metrics (next-action MSE, Q-delta) are proxies; online baselines restricted to the in-production DT; and the ROAS confidence interval is very wide ([5.0%, 28%]).
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
- OneBid is, to the authors' knowledge, the first auto-bidding foundation model: one pretrained backbone across seven oCPX conversion scenarios that already beats a heavily tuned per-scenario production Decision Transformer online.
- Decomposing conditioning into Return-to-Go and Cost-to-Go tokens plus sequence-level MoE yields multi-objective control and latency-bounded capacity scaling with power-law scaling behavior (L(N,D) ~ 8.5e-4 * N^-0.24 * D^-0.84, R2 = 0.82).
- CROP enables safe GRPO-style offline policy improvement by substituting critic-based group-relative evaluation for online rewards, improving Q-values more than AWR (1.90% vs 0.12%) and GAS (0.06%).
- Fully deployed on Kuaishou oCPX ads since May 2026, OneBid delivers +2.2% overall ADVV across scenarios (up to +13.1% in ROAS) while largely maintaining or improving cost-ratio validity.
Strengths
Production-scale online validation with sound A/B hygiene
33% production traffic for at least 5 days with pre/post AA validation; per-scenario confidence intervals reported (Table 4, e.g. Form Submit +1.9%, CI [1.0%, 2.8%]); successive-generation comparison in Table 2 isolates pretrained vs finetuned OneBid against the production DT. Fully deployed at Kuaishou since May 2026 with +2.2% overall ADVV.
Coherent challenge-to-mechanism mapping with ablations
Each of the three stated challenges gets a concrete component: RTG/CTG dual conditioning (§4.1), S-MoE (§4.3), CROP (§5.2). Figure 3 ablates CTG removal (worse NAP MSE) and S-MoE vs a dense model with matched activated parameters (lower NAP MSE).
Quantified scaling and latency engineering
Candid limitations section
Sensible offline adaptation of GRPO
Weaknesses
Evaluation circularity in the offline post-training evidence
Headline ROAS number rests on a very wide interval
The conversion-control signal is actually a pacing proxy
Weak and narrow baseline scope
Zero external reproducibility
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
OneBid (Kuaishou, arXiv 2609.21550) proposes what the authors call the first *foundation model for auto-bidding*: a single Decision Transformer (DT) backbone pre-trained on ~70M transitions from seven heterogeneous oCPX conversion scenarios (Purchase, ROAS, Activation, Form Submit, Leads Submit, App Invocation, Private-message Sent; Appendix D), then adapted per scenario. Three claimed challenges structure the design: multi-objective control (§4.1: RTG plus a novel Cost-to-Go conditioning token), capacity scaling under tens-of-milliseconds latency (§4.3: Sequence-level MoE), and safe offline policy improvement (§5.2: CROP). Headline results: power-law scaling of next-action-prediction loss (Eq 18, R²=0.82), CROP beating AWR/GAS on Q-value improvement (Table 1: 1.90% vs 0.12%/0.06%), and fully deployed online gains of +2.2% ADVV overall, up to +13.1% in ROAS (Table 4). Claims are ambitious but consistently scoped to oCPX, and Appendix H is candid about proprietary-data constraints.
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 is OneBid?
OneBid is Kuaishou's unified auto-bidding foundation model: a single Decision Transformer-based system that serves seven different oCPX advertising conversion scenarios (purchase, ROAS, app activation, form submit, leads submit, app invocation, private message), replacing the earlier practice of training one bidding model per scenario. It was fully deployed in Kuaishou's ad system in May 2026.
How much did OneBid improve advertising results?
In online A/B tests on Kuaishou production traffic (33% of traffic for at least five days), OneBid improved the platform's revenue metric ADVV by +2.2% overall across oCPX ads, with per-scenario gains from +1.8% (Activation) to +13.1% (ROAS, though that estimate has a wide confidence interval of 5.0%-28%), while mostly maintaining or improving the share of campaigns with valid cost ratios.
Is OneBid's code or dataset publicly available?
No. The model is trained and evaluated exclusively on Kuaishou's proprietary advertising logs (about 70 million bidding transitions across seven conversion types). The paper's limitations section calls for public multi-objective auto-bidding benchmarks but releases no code, data, or model weights.
What is CROP and why is it needed?
CROP (Critic-guided Relative Offline Policy optimization) is OneBid's fine-tuning method. GRPO-style improvement normally needs online trial-and-error, which is unsafe for live ad bidding. CROP instead samples candidate bid actions from the pretrained policy, scores them with a learned value critic, normalizes scores within each group of candidates, and updates the policy with clipping and a KL constraint that keeps it close to the pretrained model.
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 2-6 (scaling curves, ablations, t-SNE, CROP training curves) are available only as text descriptions in the extraction; exact curve values and shapes could not be verified.
- The extraction omits some state-feature details (what the state token aggregates beyond ad/control/traffic sides) and Appendix D scenario statistics, limiting precision of the methodology assessment.
- All quantitative results are single-platform (Kuaishou); no external replication exists to cross-check any number.
- Paper metadata shows a duplicated author name (Peng Jiang, two email handles) and one name/email mismatch, likely benign internal-directory artifacts but worth confirming for citation accuracy.
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.