NOVA: A Verification-Aware Agent Harness for Architecture Evolution in Industrial Recommender Systems

Shaohua Liu ∗ , Liang Fang ∗ , Yilong Sun ∗ , Shudong Huang ∗† , Qingsong Luo, Shaoxin Liu, Xiaoyang Chen, Dongqiang Liu, Chuangang Ma, Zhenzhen Chai, Henghuan Wang, Shijie Quan, Changyuan Cui, Zhangbin Zhu, Peng Chen, Wei Xu, Lei Xiao, Haijie Gu, Jie Jiang Tencent Inc. {jimmyshliu,franklfang,elonsun,ericdhuang,qsongluo,orionsxliu,noreenchen, dongcialiu,chuangangma,njzzchai,asherwang,justinquan,shoncui,defyzhu, pengchen,davidxu,shawnxiao,jerrickgu,zeus}@tencent.com · 2026

Read NOVA: A Verification-Aware Agent Harness for Architecture Evolution in Industrial Recommender Systems by Shaohua Liu ∗ , Liang Fang ∗ , Yilong Sun ∗ , Shudong Huang ∗† , Qingsong Luo, Shaoxin Liu, Xiaoyang Chen, Dongqiang Liu, Chuangang Ma, Zhenzhen Chai, Henghuan Wang, Shijie Quan, Changyuan Cui, Zhangbin Zhu, Peng Chen, Wei Xu, Lei Xiao, Haijie Gu, Jie Jiang Tencent Inc. {jimmyshliu,franklfang,elonsun,ericdhuang,qsongluo,orionsxliu,noreenchen, dongcialiu,chuangangma,njzzchai,asherwang,justinquan,shoncui,defyzhu, pengchen,davidxu,shawnxiao,jerrickgu,zeus}@tencent.com (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2606-27243.txt.

Description

Liu and colleagues present NOVA, an agent harness that evolves recommender system architectures inside a live industrial advertising platform. AutoML mostly tunes settings, while generic coding agents produce runnable code that may still fail silently as an architecture. In 2026, NOVA guides edits with an architecture gradient, a non-numeric update signal blending past modifications, verification diagnostics, metric feedback, and trajectory memory.

A verification cascade checks structure, local execution, offline quality, and online impact, blocking bad candidates early. Failure patterns are logged as forbidden directions for later rounds. Four control levels match automation to task risk, routing dangerous work to human-supervised copilot mode. NOVA posts the best effective pass rates on mid-level scaling and literature-to-production tasks. One such cycle shrinks human-attended time more than thirteenfold. Online tests show revenue gains with sharply reduced prediction bias.

Catalog record

CREATORShaohua Liu ∗ , Liang Fang ∗ , Yilong Sun ∗ , Shudong Huang ∗† , Qingsong Luo, Shaoxin Liu, Xiaoyang Chen, Dongqiang Liu, Chuangang Ma, Zhenzhen Chai, Henghuan Wang, Shijie Quan, Changyuan Cui, Zhangbin Zhu, Peng Chen, Wei Xu, Lei Xiao, Haijie Gu, Jie Jiang Tencent Inc. {jimmyshliu,franklfang,elonsun,ericdhuang,qsongluo,orionsxliu,noreenchen, dongcialiu,chuangangma,njzzchai,asherwang,justinquan,shoncui,defyzhu, pengchen,davidxu,shawnxiao,jerrickgu,zeus}@tencent.com
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE59 KB
LICENSEhttp://creativecommons.org/licenses/by-sa/4.0/
RIGHTSCC BY SA
SOURCEarXiv

Source: https://arxiv.org/abs/2606.27243

Browse the full library · About the Tower of Babel