Live #44: Multi‑Agent Automation of Rust Code Verification & 30× Storage Reduction for User Representations

Paper 1 – StarVerus: LLM‑Powered Multi‑Agent Collaboration for Industrial Rust Code Verification Automation

Industrial systems built with Rust require strict code contracts to ensure safety, but manually writing these contracts is costly and demands expert knowledge. Existing LLM‑based methods only generate proofs for predefined contracts and rely on fixed relation‑extraction modules, which cannot create full specifications from scratch or handle complex dynamic dependencies.

StarVerus addresses these gaps with a multi‑agent framework that automatically generates complete contracts during the generation phase and corrects them in a repair phase using contract alignment and proof‑repair cascades. The framework extracts bidirectional context from the function‑call graph and introduces a planner‑repairer‑executor‑rewriter collaboration mechanism to enhance proof repair capability.

Experimental results show that StarVerus achieves a highest verification pass rate of 88.6% across five benchmark test sets and successfully verifies 68.7% of conditions in the Asterinas operating system, demonstrating strong industrial applicability.

Paper 2 – FOUNDv2: Learning Unified User Quantized Tokenizers for User Representation

User representation learning underpins personalization on large‑scale web platforms, yet traditional continuous embeddings struggle with multi‑source data fusion, storage efficiency, and multi‑scale modeling. FOUNDv2 proposes a Unified User Quantized Tokenizer (U2QT) that compresses heterogeneous user data into a unified discrete token space.

The method follows a two‑stage pipeline: first, compact feature extraction maps raw data to a dense representation; second, a multi‑view RQ‑VAE together with shared and source‑specific codebooks discretizes the features into high‑efficiency tokens. A multi‑scale alignment objective simultaneously captures fine‑grained behavior dependencies and macro‑level periodicity, boosting predictive power.

Extensive experiments reveal that FOUNDv2 consistently outperforms task‑specific baselines on multiple benchmarks while reducing storage and computation costs (30× storage reduction, 3.5× training speedup). The approach has been deployed at Alipay, confirming its scalability and practicality in industrial settings.

Paper 3 – Query as Anchor: Scenario‑Adaptive User Representation via Large Language Model

Traditional user modeling treats representations as static and task‑agnostic, which limits adaptability to diverse industrial scenarios. This work upgrades the paradigm to a dynamic, query‑aware representation that balances generality with task sensitivity.

The authors construct an industrial‑scale pretraining dataset (UserU) and design a Q‑Anchor Embedding architecture that aligns multi‑modal behavior sequences, applies hierarchical encoding, and jointly optimizes contrastive and autoregressive objectives. Cluster‑based soft‑prompt tuning further enhances scenario‑specific modality modeling.

Empirical results demonstrate leading performance on ten Alipay industrial benchmarks, and an online A/B test confirms the method’s tangible business impact.

Paper 4 – HIVE: Hierarchical Generation of Integrated and Varied Ensembles for Efficient Out‑of‑Distribution Generalization

Model‑as‑a‑Service platforms must serve clients from heterogeneous domains, each potentially presenting unseen out‑of‑distribution (OOD) scenarios. Privacy constraints, low‑latency requirements, and limited compute further complicate deployment.

HIVE reframes the supplier role from delivering a single static model to providing a set of accurate, complementary, and efficiently deployable models. The pipeline first clusters probe‑sample predictions via a counterfactual mechanism to identify environments with similar prediction logic. It then employs multi‑scenario learning and a dual‑refinement transfer path to boost global stability and local specificity. Finally, quality filtering and diversity‑regularized pruning select a compact, functionally complementary model ensemble for lightweight client adaptation.

Results on five public TableShift datasets show consistent gains, especially under severe distribution shift. Since its production launch in January 2024, HIVE has served over 50 enterprise customers, with more than 95% opting to replace a standard model with the HIVE solution. Offline evaluation on a million‑scale OOD test set further validates its effectiveness in real industrial systems.