Decoding ML Decision: An Agentic Reasoning Framework for Large-Scale Ranking System

Longfei Yun; Yihan Wu; Haoran Liu; Xiaoxuan Liu; Ziyun Xu; Yi Wang; Yang Xia; Pengfei Wang

cs.AIFeb 20, 2026

Decoding ML Decision: An Agentic Reasoning Framework for Large-Scale Ranking System

Authors: Longfei Yun, Yihan Wu, Haoran Liu, Xiaoxuan Liu, Ziyun Xu, Yi Wang, Yang Xia, Pengfei Wang, +6 more

Organizations: Meta

Abstract

Modern large-scale ranking systems operate within a sophisticated landscape of competing objectives, operational constraints, and evolving product requirements. Progress in this domain is increasingly bottlenecked by the engineering context constraint: the arduous process of translating ambiguous product intent into reasonable, executable, verifiable hypotheses, rather than by modeling techniques alone. We present GEARS (Generative Engine for Agentic Ranking Systems), a framework that reframes ranking optimization as an autonomous discovery process within a programmable experimentation environment. Rather than treating optimization as static model selection, GEARS leverages Specialized Agent Skills to encapsulate ranking expert knowledge into reusable reasoning capabilities, enabling operators to steer systems via high-level intent vibe personalization. Furthermore, to ensure production reliability, the framework incorporates validation hooks to enforce statistical robustness and filter out brittle policies that overfit short-term signals. Experimental validation across diverse product surfaces demonstrates that GEARS consistently identifies superior, near-Pareto-efficient policies by synergizing algorithmic signals with deep ranking context while maintaining rigorous deployment stability.

Figures & tables

Figure 1. End-to-end GEARS pipeline. The system accepts a natural-language intent, generates and filters candidate policies through programmatic execution (Stage 1), interprets them using domain-specific skills loaded on demand (Stage 2), and validates production readiness through deterministic governance hooks (Stage 3).

Figure 2. Tolerance-based frontier expansion. (a) Standard Pareto search selects only convex-hull solutions, discarding near-optimal candidates in non-convex regions. (b) GEARS admits policies within an uncertainty-aware tolerance band $\delta=\gamma\cdot\sigma$ , retaining near-Pareto candidates that may offer better stability or deployability.

\epsilon_{m}(p)\triangleq\tau\cdot\sigma_{m}(p).

Algorithm 1 Tolerance-based Personalization: Random-weight Top-

K

+ Tolerance Pareto Filtering

Figure 3. Anonymized trace of a production GEARS session. The system filters 29 candidates to 16 via programmatic guardrail checks (Stage 1), verifies feature stability through domain-specific skills (Stage 2), and validates statistical significance before promoting the policy (Stage 3). The targeted policy neutralizes the guardrail regression present in the naive baseline.

Method	Ranking Quality			Precision@k			Global	Recall@k			Top-1 Performance
	nDCG@1	nDCG@3	nDCG@5	Prec@1	Prec@3	Prec@5	Rank Corr.	Rec@1	Rec@3	Rec@5	Top-1 Acc	Top-1 in GT
Naive	0.57	0.70	0.74	0.57	0.39	0.28	0.20	0.36	0.67	0.77	0.44	0.57
CoT ( Wei et al., 2022 )	0.68	0.80	0.83	0.68	0.44	0.30	0.31	0.43	0.73	0.82	0.57	0.68
Self-Consistency ( Wang et al., 2022 )	0.57	0.74	0.76	0.57	0.39	0.28	0.13	0.33	0.67	0.78	0.37	0.57
Self-Refine ( Madaan et al., 2023 )	0.61	0.78	0.80	0.61	0.44	0.31	0.34	0.37	0.74	0.83	0.50	0.61
Code-as-Action ( Wang et al., 2024 )	0.77	0.87	0.87	0.77	0.52	0.33	0.59	0.45	0.84	0.88	0.68	0.77

Table 1. Quantitative comparison of policy selection performance. We evaluate the proposed GEARS framework against five state-of-the-art baselines. Performance is measured across five key dimensions: Ranking Quality (nDCG@

k

), Precision (Prec@

k

), Global Rank Correlation, Recall (Rec@

k

), and Top-1 Performance.

Feature Class	Feature Name	Shift (Quantile)	Shift (Binary)	Status
Baseline	Feature Set $\mathcal{S}$	6%	2%	Benchmark
Product	Feature 2	16%	4%	Stable
Engagement	Feature 3	30%	10-12%	Stable (Binary)
Product	Feature 4	$\sim$ 50%	$\sim$ 20%	Unstable
Product	Feature 5	N/A	$\sim$ 30%	Unstable

Table 2. Feature Stability Benchmark (6-Month Window). Binary cuts align engagement feature stability with feature set

\mathcal{S}

baselines.

Figure 4. Pareto efficiency of generated policies. We plot the performance of all candidate policies, with the Pareto frontier (dark blue line) indicating the optimal trade-off curve. Annotated stars mark key policies of interest.

Figure 5. The backtest results indicate that the metrics improvement achieved by the selected policy remains consistent over a period of one month.

Domain	Metric 1 (%)	Metric 2 (%)	Metric 3 (%)
Surface 1	0.14	—	—
Surface 2	—	—	0.08
Surface 3	0.10	0.089	—
Surface 4	0.10	—	—
Surface 5	0.042	—	—
Surface 6	0.011	0.017	0.08

Table 3. Online performance improvement of GEARS in different products at Meta. A dash (—) indicates that the specific metric was not applicable or not the primary optimization target for that surface.

Treatment Arm	Metrics 1	Metrics 2
Treatment 1	$-0.049\%\pm 0.043$ ✗	+0.282% $\pm$ 0.074 ✓
Treatment 2	+0.036% $\pm$ 0.034 ✓	$-0.289\%\pm 0.073$ ✗

Table 4. Quantitative results for Baseline Treatment Arms. The table reports the percentage lift relative to the control group (

\text{mean}\pm\text{standard error}

	Metric 1		Metric 2
Policy	Lift (%)	95% CI	Lift (%)	95% CI
1	$-0.032$	$\pm 0.057$	$\mathbf{+0.350}$	$\pm 0.092$
2	$-0.029$	$\pm 0.057$	$\mathbf{+0.349}$	$\pm 0.092$
3	$-0.025$	$\pm 0.057$	$\mathbf{+0.338}$	$\pm 0.092$

Table 5. Top-3 policies recommended by GEARS for the case study. All three achieve statistically significant gains on Metric 2 while maintaining neutrality on Metric 1 (95% CI includes zero).

Explore similar work

RankEvolve: A Reliable Multi-Agent Auto-Research Harness for Evolving Ranking Models
Sep 30, 2026Zheng Chen, Linfeng Liu, Hong Li +1Mle-Bench Lite Reliable
Meta
Agentic ML Exploration (A-MLE) for Ads Ranking
Sep 8, 2026Erwin Gao, Vinodh Kumar Sunkara, Jingyi Guan +36Ads Agentic
Meta Platforms, Inc. Menlo Park and New York, USA
AgenticRecTune: Multi-Agent with Self-Evolving Skillhub for Recommendation System Optimization
Apr 21, 2026Xidong Wu, Yue Zhuan, Ruoqiao Wei +7Agentic Optimization Recommendation
Google Mountain View, California, USA

Sep 30, 2026, cs.AIJ/K move · Enter open · S save

RankEvolve: A Reliable Multi-Agent Auto-Research Harness for Evolving Ranking Models

Zheng Chen, Linfeng Liu, Hong Li, Hong Yan

Meta

Mle-Bench Lite Reliable

Open paper AskPDF arXiv

Auto-research agents, LLM systems that propose, implement, train, and evaluate model changes across iterations, promise to automate applied ML's experimental loop. Over long horizons, execution accuracy is a binding constraint: a change can silently leak held-out data, omit normalization, disconnect a gradient, or leave a train/eval flag unwired, invalidating expensive runs and compounding error across iterations. We present RankEvolve, an auto-research framework for evolving generative ranking models. An Executable Operating Protocol (EOP) declares phases, gates, branches, and loops, and the runtime enforces the compiled state machine. A meta-meta-harness composes complete black-box coding-agent products, including Claude Code and Codex, as execution-graph nodes that review and repair one another's work. In a budget-matched evaluation, heterogeneous composition raises all-oracle execution accuracy from the best single-product baseline of 45.8 percent to 62.5 percent (paired +16.7 points, 95 percent CI [6.6, 26.7]) while achieving a 10.4 percent silent critical-defect rate. An implemented knowledge layer carries findings, including negative results, across iterations. In a twelve-iteration deployment on the open-source HSTU recommender, RankEvolve reported NDCG@10 of 0.2192 on MovieLens-20M LARGE (+4.48 percent over the published anchor) and 0.1948 on BASE (+2.80 percent). ExecML-HSTU, seeded by incidents from that deployment, provides the oracle benchmark for the execution-accuracy evaluation. A pre-specified LitGPT transfer split replicates the heterogeneous-composition effect beyond recommendation (+12.5 points, 95 percent CI [3.0, 22.0]), and a paired ablation isolates per-step from full-protocol instruction injection. These results characterize when runtime-controlled composition of coding-agent products improves execution accuracy.

Abstract

Explore similar work

RankEvolve: A Reliable Multi-Agent Auto-Research Harness for Evolving Ranking Models

Agentic ML Exploration (A-MLE) for Ads Ranking

AgenticRecTune: Multi-Agent with Self-Evolving Skillhub for Recommendation System Optimization