A mathematical derivation of recursive self-improvement: why fixed improvers decay and RSI opens ceilings
青稞AI · wechat · 2026-11-23
A long-form technical article from 青稞AI unifies self-refinement, self-improvement, and recursive self-improvement (RSI) as operator iterations with full derivations.
Key takeaways:
- A fixed improvement operator necessarily yields geometrically decaying returns: the trajectory is fully determined by the initial state and the operator, capped by its fixed point;
- RSI introduces a meta-improvement operator forming a dual loop that raises the ceiling, but changing the method guarantees neither improvement nor increasing returns — evaluation is what guards against regressions;
- Any improvement round decomposes into proposer–evaluator–learner, instantiated via Self-Refine, Darwin Gödel Machine, STaR, Self-Rewarding, and AbsoluteZero, with derivations of STaR's equivalence to maximum-likelihood training and error-probability formulas for evaluator mistakes;
- The article contrasts inference-time improvements (output revision, memory, scaffolding) with training-time ones (self-generated reasoning data, self-reward, self-play curricula), closing the loop with Darwin Gödel Machine (SWE-bench 20.0%→50.0%) and AlphaEvolve.
More from AGI Musings
- Biologists push back on AI figures claiming biology is "solvable" — GaryMarcus · 2026-11-23
- Stripe cofounders declare the world 'entered the singularity' as Silicon Valley echoes — coherence · 2026-11-23
- Kevin Roose takes his AGI-pill quest to New Yorker editor David Remnick — kevinroose · 2026-11-23
- 'We were here first': an argument that humanity deserves a branch even in an ASI future — ZeroStateReflex · 2026-11-23
- Writer Predicts LLMs Will Train on Bot Output as Human Writing Dwindles — TuhinChakr · 2026-11-23
- Apple's appenz: agent swarms are filling in NVIDIA's CUDA moat — appenz · 2026-11-23