Reusable parts let a single fix repair many broken AI training environments

By decomposing each simulated workplace into shared intents, rules, records and tool servers, the authors raised the share of environments an agent can actually solve from 71.6% to 88.3% in one round of repair.

Chinese Tech

Zhuosong Jiang · Chengxu Wu · Junqi Liu · Yujie Yi · Siqin Wang · Zhixuan Huang · +8 more

Shanghai Jiao Tong University · Tencent · Emory University · National University of Singapore · Peking University

Research Digest·3 November 2026·3 min read

Illustration: Reusable parts let a single fix repair many broken AI training environments

In one line

Building agent training environments from reusable, evolvable pieces, then repairing the pieces when an environment breaks, raises how many work and improves agent training.

Before you read this

Three ideas this paper builds on

harness evolutionIteratively adapting an agent's environment, prompts, or evaluation setup to improve task performance and generalization.60 papers →verificationChecking outputs for correctness before accepting them, often to catch errors or reward hacking.284 papers →reinforcement learning with verifiable rewardsReinforcement learning that uses rewards from objectively checkable signals like correct answers or verification feedback.143 papers →

Read it as

AI agents, software that carries out tasks by taking actions, learn by working inside simulated workplaces. Each workplace is usually hand-built for one task, and building a fresh one for every task is slow and expensive.

This work assembles each simulated workplace from four reusable parts: a goal, a business rule, a set of data records, and a server supplying the tools to act with. When a workplace failed because a part was wrong, they fixed that part, so every later workplace using it improved too.

One repair round raised the share of workplaces that worked from 71.6% to 88.3%. Agents trained inside them scored substantially better afterwards, and faults traced back to a few shared causes.

RSS

The Daily Edition

The front page in your inbox each weekday morning: the lead, the picks, and why each one earned its place.

Concepts

harness evolution verification reinforcement learning with verifiable rewards failure attribution supervised fine-tuning