Selected work

Representative Papers

Rounding in Preconditioner Space: Redesigning 4-bit AdamW Optimizer-State Quantization

Oct 08, 2026

This study addresses the issue of rounding errors accumulating recursively through moment estimation during optimizer state quantization in AdamW, which leads to preconditioner distortion and perturbed adaptive updates. To mitigate this, the work reformulates 4-bit AdamW quantization from a rounding-space perspective, proposing ZIP-SR (preconditioner-space stochastic rounding) and ZE-EDEN (zero-exclusion calibration) strategies, integrated with the NF4 format to optimize quantization dynamics. Evaluated on models ranging from 130M to 2.7B parameters, the proposed approach reduces the validation loss gap by 70%, achieving training performance closely approximating full 32-bit precision.

0 citationsRead paper

From Valid to Useful: Post-Verification Acquisition for Recursive Self-Improving Recommendation

Oct 03, 2026

This study addresses the training bias arising from uneven source contributions of synthetic sequences in recursive self-improving recommendation (RSIR). We propose the DA-RSIR framework, which establishes "post-validation acquisition" as an independent control point to decouple validation from selection without requiring additional labels or teacher models. Furthermore, it introduces Bayesian Active Learning by Disagreement (BALD) scoring with Monte Carlo Dropout estimation to enable unsupervised sample importance evaluation based on predictive disagreement, thereby optimizing training data selection for subsequent iterations. Experiments across four datasets demonstrate that our method comprehensively outperforms the retain-all strategy, achieving single-round performance that surpasses five-round cumulative gains with high statistical significance.

0 citationsRead paper

Screen Before You Serve: Simulation for Production Customer Experience AI Agents at 140M Scale

Sep 24, 2026

This study addresses the insufficient test coverage for customer service AI agents in regulated industries and the risk that online experimentation poses to user trust. To overcome these challenges, this work proposes a hypothesis-driven simulation evaluation workflow. By leveraging the Snowglobe simulator and synthetic data generation techniques, the method enables multi-turn interaction simulation without invoking production backends. Furthermore, end-to-end binary evaluators are employed to pre-screen candidate agents for safety, facilitating large-scale configuration exploration with zero user risk. Experimental results demonstrate that simulated scores align closely with production environments. In A/B testing, the proposed approach yields a 36.69-point improvement in tNPS and an 8.82% increase in self-service rates, thereby achieving safe, iterative development of AI agents within regulated settings.

0 citationsRead paper

Building Customer Support AI Agents at 100M-User Scale: An Evaluation-Driven Framework

Jun 07, 2026

This work addresses the deployment blind spots commonly encountered in developing production-grade AI customer service agents for hundreds of millions of users, which often stem from fragmented workflows across evaluation, context engineering, training, and online metrics. To bridge this gap, the authors propose the first unified development framework centered on evaluation-driven design. The framework integrates structured context engineering, human-in-the-loop prompt iteration, an LLM-based judging mechanism with consistency guarantees (including GEPA optimization), and an end-to-end validation pipeline, achieving strong alignment between offline metrics and online performance. Evaluated across five real-world customer service scenarios, the approach substantially enhances user experience—evidenced by a 37-percentage-point increase in AI Net Promoter Score and a 29-percentage-point rise in self-service rate in the card delivery scenario—with most scenarios approaching expert human-level performance.

0 citationsRead paper

Your Spending Needs Attention: Modeling Financial Habits with Transformers

Jul 31, 2025

To address the challenges of high-dimensional heterogeneity and costly manual feature engineering in financial user behavior data (e.g., transaction logs, app events, customer service records), this paper proposes nuFormer—the first end-to-end user representation learning framework that jointly encodes merchant textual features (e.g., business names) and structured fields (e.g., amount, timestamp) into a Transformer architecture. Leveraging self-supervised pretraining on raw transaction sequences, nuFormer jointly optimizes user embeddings and conventional handcrafted features without requiring auxiliary data sources. Empirical evaluation on Nubank’s large-scale recommendation system demonstrates substantial improvements in core metrics—including CTR and Recall—while also exhibiting strong generalization performance in risk prediction and fraud detection tasks. These results validate nuFormer’s industrial-grade effectiveness, scalability, and methodological novelty.

0 citationsRead paper

Recent publications

Latest Papers

Rounding in Preconditioner Space: Redesigning 4-bit AdamW Optimizer-State Quantization

Oct 08, 2026

This study addresses the issue of rounding errors accumulating recursively through moment estimation during optimizer state quantization in AdamW, which leads to preconditioner distortion and perturbed adaptive updates. To mitigate this, the work reformulates 4-bit AdamW quantization from a rounding-space perspective, proposing ZIP-SR (preconditioner-space stochastic rounding) and ZE-EDEN (zero-exclusion calibration) strategies, integrated with the NF4 format to optimize quantization dynamics. Evaluated on models ranging from 130M to 2.7B parameters, the proposed approach reduces the validation loss gap by 70%, achieving training performance closely approximating full 32-bit precision.

0 citationsRead paper

From Valid to Useful: Post-Verification Acquisition for Recursive Self-Improving Recommendation

Oct 03, 2026

This study addresses the training bias arising from uneven source contributions of synthetic sequences in recursive self-improving recommendation (RSIR). We propose the DA-RSIR framework, which establishes "post-validation acquisition" as an independent control point to decouple validation from selection without requiring additional labels or teacher models. Furthermore, it introduces Bayesian Active Learning by Disagreement (BALD) scoring with Monte Carlo Dropout estimation to enable unsupervised sample importance evaluation based on predictive disagreement, thereby optimizing training data selection for subsequent iterations. Experiments across four datasets demonstrate that our method comprehensively outperforms the retain-all strategy, achieving single-round performance that surpasses five-round cumulative gains with high statistical significance.

0 citationsRead paper

Screen Before You Serve: Simulation for Production Customer Experience AI Agents at 140M Scale

Sep 24, 2026

This study addresses the insufficient test coverage for customer service AI agents in regulated industries and the risk that online experimentation poses to user trust. To overcome these challenges, this work proposes a hypothesis-driven simulation evaluation workflow. By leveraging the Snowglobe simulator and synthetic data generation techniques, the method enables multi-turn interaction simulation without invoking production backends. Furthermore, end-to-end binary evaluators are employed to pre-screen candidate agents for safety, facilitating large-scale configuration exploration with zero user risk. Experimental results demonstrate that simulated scores align closely with production environments. In A/B testing, the proposed approach yields a 36.69-point improvement in tNPS and an 8.82% increase in self-service rates, thereby achieving safe, iterative development of AI agents within regulated settings.

0 citationsRead paper

Building Customer Support AI Agents at 100M-User Scale: An Evaluation-Driven Framework

Jun 07, 2026

This work addresses the deployment blind spots commonly encountered in developing production-grade AI customer service agents for hundreds of millions of users, which often stem from fragmented workflows across evaluation, context engineering, training, and online metrics. To bridge this gap, the authors propose the first unified development framework centered on evaluation-driven design. The framework integrates structured context engineering, human-in-the-loop prompt iteration, an LLM-based judging mechanism with consistency guarantees (including GEPA optimization), and an end-to-end validation pipeline, achieving strong alignment between offline metrics and online performance. Evaluated across five real-world customer service scenarios, the approach substantially enhances user experience—evidenced by a 37-percentage-point increase in AI Net Promoter Score and a 29-percentage-point rise in self-service rate in the card delivery scenario—with most scenarios approaching expert human-level performance.

0 citationsRead paper

Your Spending Needs Attention: Modeling Financial Habits with Transformers

Jul 31, 2025

To address the challenges of high-dimensional heterogeneity and costly manual feature engineering in financial user behavior data (e.g., transaction logs, app events, customer service records), this paper proposes nuFormer—the first end-to-end user representation learning framework that jointly encodes merchant textual features (e.g., business names) and structured fields (e.g., amount, timestamp) into a Transformer architecture. Leveraging self-supervised pretraining on raw transaction sequences, nuFormer jointly optimizes user embeddings and conventional handcrafted features without requiring auxiliary data sources. Empirical evaluation on Nubank’s large-scale recommendation system demonstrates substantial improvements in core metrics—including CTR and Recall—while also exhibiting strong generalization performance in risk prediction and fraud detection tasks. These results validate nuFormer’s industrial-grade effectiveness, scalability, and methodological novelty.

0 citationsRead paper