Awesome Recursive Self-Improvement 🧬

Autonomous agents that evolve by rewriting their own source code based on interaction data, moving beyond simple skill-file updates to core architecture modification.

Scaling stateful AI agents with rapid checkpoint and rollback of the complete sandbox state, enabling high-frequency exploration and tree-search.

Training for diversity improves test-time search. Integration with AlphaEvolve for autonomous discovery of task-specific reward functions.

Proposes a lifecycle-aware runtime harness that improves frozen LLM agents by converting recurring interaction failures into reusable interventions. average relative improvement of 88.5%.

Long-term memory framework that decouples retrieval from distillation. Zero commercial-API token cost for memory operations.

Exploits agent idle time (waiting for tool outputs) to generate plan candidates, lifting accuracy on complex benchmarks like GAIA and MLE-Bench.

A self-evolution method using vanilla GRPO interleaved with offline self-distillation (OFSD) to improve search-augmented agents without external supervision.

Proposes DecentMem, a decentralized memory framework where agents maintain dual-pool memory to guarantee global reachability and improve accuracy by up to 23.8%.

Extracts low-rank capability subspaces from gradients to fine-tune models on their own outputs without external signals.

A multi-reward RLIF framework using answer-level and completion-level rewards with KL-Cov regularization to stabilize long-horizon reasoning.

Lifelong learning framework that enables agents to refine their own execution policies continuously. A significant step toward autonomous evolution.

Introduction of a framework that allows agents to learn on the fly by accumulating memory and reflection, enabling RSI without backpropagation.

Proposes a Searcher-Navigator framework where agents assemble research evidence like a jigsaw puzzle, outperforming brute-force parallel rollouts.