Peking U Paper: RSI Tax Shows Verification Eats 74.8-93.6% of Self-Improvement Cost
RobertTLange · x · 2026-11-18
Robert Lange highlights the Peking University paper The RSI Tax, which introduces token-cost accounting for recursive self-improvement (RSI).
- Across six methods and two models, a simple bounded-note baseline (SRN) hits 59.18% mean accuracy, beating Meta-Harness, EvoSkill, and RRSI with 65.6-71.0% total-token savings
- The RSI Tax decomposes into verification, exploration, and writing costs; verification accounts for 74.8-93.6%
- A large share of adaptation spend goes to artifacts never used later — performance gains alone overstate RSI's value
The author argues RSI/discovery harnesses should follow LLM benchmarks in plotting token efficiency alongside raw scores.
More from AGI Musings
- LLMs cram textbooks, genomic models discover new biology — julia_kiseleva · 2026-11-18
- SF AI circle's undercurrent of anxiety: only 1-5 years of intellectual alpha left before AGI — CatAstro_Piyush · 2026-11-18
- The nonhuman economy: AI compute buildout may crowd out land, labor and everything else — thomasahle · 2026-11-18
- Researchers push back on 'agent swarm' takeover premise: real-world 80% AI success is only ~15-minute tasks — rbhar90 · 2026-11-18
- AI weather model gives hurricane targets one extra day of warning, Nature paper shows — davidmanheim · 2026-11-18
- Software singularity must be a sigmoid: compute substrate caps AI takeoff curves — tszzl · 2026-11-18