Today's edition highlights the growing trend of evaluator-first evolution and unsupervised tool refinement, moving beyond simple prompt optimization.
The loop: VeriEvolve constructs and revises its own executable evaluator using agent run outcomes, which provides a better fitness signal for the next round of agent evolution.
An evaluator-first evolution framework that synthesizes and refines executable evaluators to provide reliable fitness signals for self-improving agents.
The loop: MLEvolve uses a multi-agent framework to discover machine learning algorithms, which updates its Retrospective Memory and search strategy, making the agent better at subsequent algorithm discovery.
An LLM-based multi-agent framework that discovers machine learning algorithms and evolves by accumulating experience in a dynamic memory.
The loop: AtomSD extracts fine-grained procedural knowledge from a model's own successful trajectories and distills it back into the evolving policy to improve its performance on complex tasks.
An atom-guided self-distillation framework where a learner continually extracts fine-grained procedural knowledge from its successful trajectories to improve its own policy.
The loop: TEAL evolves a library of tools and usage skills for a coding agent by analyzing its trajectories, which the agent then uses to solve analytical questions more effectively.
An unsupervised framework that automatically builds and refines reusable tools and usage skills for coding agents answering analytical questions over data lakes.
The loop: HarnessDev enables LLMs to design and evolve their own agent harness, which then improves the execution of the LLMs themselves.
A benchmark and framework exploring whether LLMs can create and evolve their own agent harnesses.
The loop: Auto-research loops recursively scale to optimize an efficient agent harness, which improves the infrastructure for further agent research.
A system that recursively scales auto-research loops to optimize efficient agent harnesses.
The loop: AI agents are evaluated on their ability to adapt to an evolving harness, which improves the evaluation of agentic R&D systems.
A benchmark that evaluates the ability of AI agents to adapt to and keep pace with an evolving harness.
The loop: RoboDoctor performs recursive self-improvement on robot policies, which makes the system better at controlling robots.
An official code implementation for training-free recursive self-improvement for robot policies.
