RSI Is Deep Learning All Over Again: 11 Mechanisms from Adam to Dropout Mapped onto Agent Self-Improvement

青稞AI · wechat · 2026-10-21

A detailed Chinese-language walkthrough maps recursive self-improvement (RSI) methods onto classical deep learning mechanisms. When agents iteratively rewrite their Skills, the recurring failures — worse performance on unseen tasks, regressions on old ones, ever-growing documents — are exactly overfitting, catastrophic forgetting, and redundancy. Based on a recent RSI survey, the article covers 11 borrowed mechanisms across four angles:

Optimizers & backprop: SkillAdam adapts Adam's momentum and adaptive step-size ideas to decide what and how much to edit each round; TextGrad passes textual "gradients" backward through an agent pipeline to attribute failures to specific prompts or components.
Learning signals: MemRL adds a utility Q-value to memories for retrieval; SEAL does meta-learning by having the model write and test its own finetuning data; RISE extrapolates a stronger teacher from RL updates then distills back; ACE uses adversarial self-play between solver and bug-finding tester; DMoA learns a router for ensemble-style multi-agent selection.
Regularization & continual learning: RRSI limits edits per round and prunes unhelpful components; SkillEvoReg randomly masks Skill instructions (a Dropout analog) before revising.

The takeaway: RSI is not a brand-new paradigm — decades of deep learning tooling offer ready-made prototypes for agent self-improvement, and the mapping suggests many unexplored research directions.

Original post →

RSI Is Deep Learning All Over Again: 11 Mechanisms from Adam to Dropout Mapped onto Agent Self-Improvement

More from coding & agent