Self-Improving Agents
Agents that rewrite their own code, data, or skills
From self-play and test-time adaptation to executable subagents and autoresearch ratchets — how agents compound capability over time.
Reflexion
Verbal reinforcement learning — agent writes language 'lessons' after failures, reads them before retry. HumanEval 80% → 91% without weight updates.
SPIN
Self-play fine-tuning, model vs previous self, no extra annotations needed.
Cherry LLM
Self-guided data selection via IFD metric, 5% data outperforms full dataset.
RISE
Recursive introspection, multi-turn self-correction, +23.9% on GSM8K.
EvolveR
Experience-driven self-evolution, distill trajectories into abstract strategic principles.
Self-Improving at Test-Time
Detect weak spots → auto-generate data → LoRA at test time, +5.48% with 68× fewer samples.
Metacognitive Learning
Framework: agents need self-assessment, learning planning, and evaluation to truly self-improve.
AgentFactory
Preserves successful solutions as executable Python subagents, not text. Install→Self-Evolve→Deploy lifecycle, ~57% orchestration cost reduction.
autoresearch
Autonomous overnight ML research — agent edits train.py, runs 5-min experiments, keeps/reverts on val_bpb. program.md as lightweight skill.
Darwin Skill · 达尔文
Autoresearch ratchet applied to SKILL.md optimization — 8-dim rubric (structure + effectiveness), independent sub-agent scoring, git-revert on regression.