AI makes AI better.

5 picked from 8 candidates · 31 sources read · 61 loops indexed

1. 01
   Loop of the dayPaperarXiv7 Oct 2026Training data

   AdvSim2Real: Training Web Agents Against Adaptive Prompt Injection in a Web World Model

   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.

   Co-evolving a task curriculum and an adversary inside a simulated web environment creates a robust training loop that hardens agents against adaptive prompt injections.

   loop fit 10/10Sarim Hashmi, Mukul Ranjan, Kshitij Mishra, Mikhail Kuznetsov, Praneeth...via arXiv
   
2. 02
   PaperarXiv6 Oct 2026Architecture and optimizer search

   RLDiscover: LLM-Driven Co-Evolution of Reinforcement Learning Algorithms

   The loop: The RLDiscover framework progressively co-evolves the components of model-free deep reinforcement learning algorithms. The discovered algorithms improve the learning efficiency of the agents that use them, providing better fitness signals for the framework to discover even stronger algorithms.

   loop fit 9/10Haoran Li, Zengle Ge et al.via arXiv
   
3. 03
   PaperarXiv8 Oct 2026Tools and harness

   Harness Evolution Hits a Ceiling: When Weight Training Should Begin

   The loop: A self-evolving harness loop repairs process failures to generate successful execution trajectories for an LLM agent. These trajectories are then used to train the model's weights, internalizing the gains and making the model better under the original harness.

   loop fit 9/10Yuan Tian, Bing Hu et al.via arXiv
   
4. 04
   PaperarXiv6 Oct 2026Self-modifying agents

   Learning from Revision Consequences: Hindsight Meta-Experience Distillation for Self-Improving Agents

   The loop: A self-improving agent constructs meta-experience by re-executing incumbent and revised meta-skills from the same restored discovery state. This hindsight distillation improves the agent's meta-skills, making it better at discovering and refining future task-skills.

   loop fit 9/10Qianhan Feng, Zhongzhen Huang et al.via arXiv
   
5. 05
   PaperarXiv8 Oct 2026Training data

   SynCo: Data Synthesis Co-Training for Self-Evolving LLMs via Multi-Agent Reinforcement Learning

   The loop: A Synthesizer agent constructs training tasks based on a Reasoner agent's current capabilities, and the Reasoner learns from the resulting experience. The outcomes of the Reasoner's rollouts provide complementary rewards that jointly optimize both agents, making the Synthesizer better at generating informative tasks.

   loop fit 9/10Wei Yang, Shawn Li et al.via arXiv
   
6. 06
   PaperarXiv9 Oct 2026Self-reward and self-play

   The Red Queen Gödel Machine: Co-Evolving Agents and Their Evaluators

   The loop: A coder agent and a reviewer agent co-evolve under non-stationary utilities. The reviewer grades patches to guide the coder's search, while the coder's output helps the reviewer refine its grading rubric, making both better at their respective tasks.

   loop fit 10/10Alex Iacob, Andrej Jovanovi\'c, William F. Shen, Daniel Burkhardt, Meghdad...via arXiv
   
7. 07
   PaperarXiv7 Oct 2026Kernels and compilers

   KernelOPT: Dispatch-Aware Agentic Search for GPU Kernel Optimization

   The loop: A multi-agent system profiles and optimizes generated Triton sub-kernels within compiled models. This improves the execution efficiency of the models and allows the agents to verify the re-stitched model end-to-end.

   loop fit 9/10Aheli Poddar, Sanskar Prasad, Arindam Samanta, Subha Chakraborty, Vishal...via arXiv
   
8. 08
   PaperarXiv8 Oct 2026Research automation

   Language Models as AI Research World Models

   The loop: Language models act as Research World Models to predict the outcomes of candidate AI research experiments. This improves the selection of future interventions and generates new experimental data to further train the world model.

   loop fit 9/10Zijun Wang, Zewen Liu et al.via arXiv