Today's edition highlights systems that learn from their own structural experiments, whether by co-evolving weights and harnesses, distilling multi-agent workflows, or reflecting on evolutionary history.
The loop: A shared model searches for better multi-agent workflows and selects the best executions. These executions are then used to train the model, improving its base capabilities.
A shared model acts as a team of subagents to solve open-ended research tasks, then trains on its own best executions to improve its performance.
The loop: A policy proposes and reflects on harness edits, then uses the outcomes of those edits to train itself. This improves its ability to evolve the agent harness.
A framework co-evolves an agent's weights and its harness by using the agent to propose edits and then training it on the outcomes.
The loop: An agentic mutator optimizes AI systems by retrieving evidence from its own evolutionary history and probing the live system. This allows it to verify and propose new mutations more effectively.
An evolutionary optimizer uses an agentic mutator that learns from its own history to improve production AI systems like data pipelines and workflows.
The loop: A policy judges and ranks its own debugging trajectories to generate auxiliary rewards. These rewards are then used to improve the policy through reinforcement learning.
A reward-refinement method allows a software engineering agent to judge its own debugging behavior and use those judgments to improve its performance.
The loop: A predictive world model selects learning data and guides actions. The resulting feedback is then used to adapt the world model and improve the policy.
A domain-general framework uses predictive world models to guide action and learning-data selection for recursive self-improvement.
The loop: RSI-Anything uses a predictive world model to select learning data and guide actions, then uses the resulting feedback to adapt the world model and improve the policy, making the system better at future selection and execution.
RSI-Anything: Unlocking Recursive Self-Improvement with World Models
