Today's edition highlights systems that automate their own optimization processes, moving beyond fixed loops to dynamic self-improvement.
The loop: The FREEEVOLVE agent automates the design of workflows and evaluation loops for language model agents. It improves its own evolution skill through meta-evolution by scoring each candidate skill on the fresh target agent it produces.
A framework that automates the optimization loop for agent workflows and uses meta-evolution to improve its own evolution skill.
The loop: A meta-agent diagnoses failures in robot manipulation tasks to revise its tools and skills. It also improves its own diagnostic tools, ensuring that retained revisions support both later action and subsequent self-improvement.
A framework for physical recursive self-improvement where a meta-agent diagnoses failures and revises its own diagnostic tools and skills.
The loop: An online task synthesis pipeline creates coding tasks calibrated to a 4B agent's current learnability frontier. The agent trains on these synthetic tasks via reinforcement learning, which shifts its frontier and prompts the pipeline to generate harder tasks for the next round.
A 4B coding agent post-trained exclusively via reinforcement learning on synthetic tasks generated at its current learnability frontier.
The loop: EBR-bench measures the ability of models to learn from their own experience. By evaluating how well models adapt, the benchmark provides a metric that tracks and guides the development of self-improving AI systems.
An update to Epoch AI's benchmark that tests the ability of models to learn from experience, including new rules to prevent exploit strategies.
The loop: The agentdescent framework treats agent parameters like skills and harnesses as optimizable components. This allows an optimizer to compute diffs as gradients and update the agents, enabling them to evolve their own components to improve performance.
A parallel, asynchronous framework that applies gradient descent concepts to self-evolving agents.
The loop: AutoRef automates the optimization of evaluation harnesses for image generation agents, improving the accuracy and efficiency of model development.
Listed in Awesome-AI4AI under: targets/harness
