Alphabell.
Paper · Training data

AdvSim2Real

AdvSim2Real: Training Web Agents Against Adaptive Prompt Injection in a Web World ModelA training framework co-evolves a task curriculum and an injection adversary inside a web world model to make a web agent more capable and robust against prompt injections.

AdvSim2Real
Figure from the paper, arXiv. Source
The loop

A web world model co-evolves a task curriculum and an injection adversary to train a web agent. The resulting training data improves the agent's robustness and capability, allowing it to handle more difficult tasks and stronger adversaries in the next iteration.

The loop
AdvSim2Real
Paper · arXiv
  1. Simulator co-evolves tasks and adversaries
  2. Generates challenging web agent scenarios
  3. Agent trains on resulting curriculum
  4. Agent handles harder tasks and attacks
  1. Simulator co-evolves tasks and adversaries
  2. Generates challenging web agent scenarios
  3. Agent trains on resulting curriculum
  4. Agent handles harder tasks and attacks
↻ The improved system does the next round, and the loop turns again.

Why it is a road to recursion

Co-evolving tasks and adversaries within a world model provides an automated curriculum that continuously pushes the agent's robustness and capabilities without requiring manual task design.

Evidence

Training in the simulator makes a 4B agent both more capable and more robust: its completion rises with and without attacks, holds against a frontier-model adversary it never trained against.

web-agentsadversarial-trainingprompt-injection

Related loops

More training data →
Nemotron-4 340B
PaperarXiv
Nemotron-4 340B
  1. Aligned model generates training data
  2. Reward model judges and filters it
  3. Data trains next intermediate instruct model
  4. Next model generates better alignment data
  5. repeat
Training dataNVIDIA · 2024
Recursive Self-Improvement in Unified Multimodal Models
PaperarXiv
Recursive Self-Improvement in Unified Multimodal Models
  1. UMM generates images and programs
  2. Execution verifies programs against specifications
  3. Verified renders and programs form training data
  4. Data trains UMM visual and text capabilities
  5. repeat
Training dataHuijuan Wang et al. · 2026
Tongyi DeepResearch
ProjectGitHub
Tongyi DeepResearch
  1. Agent synthesizes agentic pretraining data
  2. Pipeline adjusts training set in real time
  3. Data feeds back into model training
  4. repeat
Training dataTongyi Lab, Alibaba · 2025