llive 完全解説 (5) — 「集団が学ぶ AI」: v0.B/C/D/E 派生集団進化総括

Last updated at 2026-05-24Posted at 2026-05-24

言語 / Language / 语言 / 언어: 日本語 | English | 中文 | 한국어

日本語

llive 完全解説 (5) — 「集団が学ぶ AI」: v0.B/C/D/E 派生集団進化総括

コンセプト hook: 1 個の AI が賢くなるのではなく, 64 個の AI が世代を
回して互いに評価し合い, 嘘の合意は Approval Bus が止める — それが llive の
v0.E. 2026-05-21 marathon でその架構が 303 件 test + ruff 0 警告 + governance
skeleton 着地まで揃った. Hillis 1990 から AlphaStar 2019 まで 30 年の
系譜を 1 OSS に圧縮した結果.

本記事は連載 #24 の中核. v0.B (Genome / EvolutionLoop) → v0.C (subprocess
分離) → v0.D (self-adaptive + meta mutation) → v0.E (peer evaluation +
persona + governance) の 4 段階を 1 本に総括.

0. 連載中での位置づけ — 本連載の中核

#24-00 series index
#24-01 4 層メモリ          ← 「個体の中の記憶」
#24-02 思考因子 × COG-MESH ← 「個体の中の思考軸」
#24-03 構造進化 × TRIZ × Z3 ← 「個体内の構造書換え」
#24-04 B-series           ← 「個体内の収束 (速い小脳)」
#24-05 EvolutionLoop      ← 「個体間の探索 (遅い大脳)」 ★ 本記事
#24-06 LLM backend         ← 「個体を動かす管」
#24-07 governance         ← 「個体間決定の audit」
#24-08 lleval              ← 「個体を測る眼鏡」

#24-05 は全体の背骨. v0.B/C/D/E で「派生集団そのもの」を作る. 他の
記事はそこに乗る機能. これは連載の中核 — 他の全章の機能が乗る基盤である.

1. なぜ集団進化なのか — Hillis の警告

W. D. Hillis 1990 が示したのは「評価者と被評価者が同時に進化する」と
fitness landscape は指数的に面白くなる, ということ.
Red Queen Effect で集団全体の質が 自走で上がる. 単一 best を選び続け
ると 局所最適に陥る.

llive はこれを LLM に持ち込んだ. 派生集団 N=64 が互いに評価, 評価結果が
fitness, fitness が次世代の selection. すると:

「評価者の質」自体が世代と共に上がる
単一 best が全体を支配できない
「全派生が嘘の高得点を付け合う」共謀 が発生し得る (CE-06 で検出)

2. v0.B — Genome / EvolutionLoop / 並列 scheduler

v0.B core は GA 古典. 着地 module は Genome, Selection, Crossover, Mutation,
scheduler:

Genome (実数 vector + bounds + labels) + Individual + Population.
TournamentSelection / RouletteSelection / ElitismSelection.
UniformCrossover / BlendCrossover / SegmentCrossover.
GaussianMutation / ResetMutation / ChainedMutation.
EvolutionLoop (EvolutionConfig + EvolutionResult).
並列 scheduler 3 種: serial_scheduler / MultiprocessingScheduler / AsyncioScheduler.

これだけで「集団 → 評価 → 選別 → 交配 → 突然変異 → 次世代」のループが回る.

3. v0.C — subprocess 分離 + 派生実走

LLM 推論は 1 派生個体あたり OS process 1 つに 完全分離 したい. 理由は:

LLM 重い → メモリ leak / GIL 競合を物理分離
1 派生が落ちても他は生存
OS-level timeout / SIGKILL で fault isolation

VariantSubprocessScheduler (subprocess_scheduler.py) — subprocess.run +
ThreadPool 並列 + timeout + retries + cleanup. これで variant_runner.py
スクリプトを派生 1 個体として起動可能.

4. v0.D — 自己参照 mutation (Schwefel σSA-ES + meta mutation)

v0.D core は「mutation rate そのものを進化させる」.

SelfAdaptiveGaussianMutation (Schwefel σSA-ES, log-normal σ update).
Genome に σ vector を埋め込み, mutation が σ も書き換える.
MetaMutation (strategy_id を genome に, 集団内で 4 戦略並走).
pack_self_adaptive_bounds / pack_meta_strategy_bounds — 38/20/39 dim 化.

これで「どの mutation 戦略が今の問題に効くか」自体が世代を超えて
学習される.

5. v0.E — peer evaluation + persona ontology + governance

v0.E core. CE-01〜34 を含む. 主要 module は以下:

5.1 評価 (CE-01〜05)

PeerEvaluationMatrix — N×N 採点行列. 共謀検出 3 指標
(score_variance / symmetry / concentration). Mermaid 可視化.
PeerFitnessAdapter — EvolutionLoop.scheduler 互換.
EvaluationStyleGenome — 派生に「辛口 / 甘口 / 精度 / 速度」の
evaluation persona dim を埋め込み.

5.2 多様性保護 (CE-24〜29)

latin_hypercube_population — 空間均等初期集団 (scipy.stats.qmc).
NoveltyScorer — k-NN, Lehman-Stanley 2008/2011.
DiversityPreservingBreedFilter — novelty rejection + resample.
DiversityMonitor — diversity_l2 / spread / median + 閾値 alarm.

5.3 Quality Diversity (CE-25 / CE-26, 本日着地)

PersonaOverlapPenalty — fitness 軸に persona dissimilarity の集団平均加算.
MAPElitesGrid — Mouret & Clune 2015 の 4 軸版 (persona 2 × thought_factor 2).
各 cell に最大 fitness 個体を保存.

5.4 Historical persona (CE-19〜23)

PERSONA_ONTOLOGY 10 名 (岡潔 / グロタンディーク / ファインマン / ガロア /
フォン・ノイマン / ニュートン / カント / ソクラテス / 老子 / 孫子).
PersonaComposition (3 policy: exclusive / mix / moderator).
PersonaCompositionMutation (CE-21).
persona_dissimilarity — Jaccard + L2 of factor_affinity.
PersonaImportAlgorithm (CE-20, 本日着地) — 派生間 persona 部分採用.
PersonaSurvivalAnalysis (CE-22, 本日着地) — どの persona 組合せが
世代を生き残ったか統計.
PersonaCorpusLoader (CE-23, 本日着地 skeleton) — Raptor RAD から
自動抽出.

5.5 集団組み合わせ機構 (CE-30〜34)

MutualScorePairSelector (CE-30, mating.py) — assortative mating,
softmax sampling.
NSGA2Selection (CE-31, nsga2.py) — Pareto front + crowding distance.
Speciation (CE-32, speciation.py) — NEAT 流種分け.
IslandModel (CE-33, island_model.py) — ring/fully/star 3 topology +
best/random/worst migration.
LexicaseSelection (CE-34, mating.py) — Helmuth 2014, case-by-case 順位.

5.6 Governance (CE-06〜08, 本日着地 E.4)

CollusionDetector (CE-06) — is_suspected_collusion を threshold
dataclass で wrap.
CoevolutionGovernance (CE-07) — 共謀疑い → ApprovalBus.request 発火.
collusion_risk_score (CE-08) — TonicRiskMonitor.tick に投入する
state → [0, 1] risk.
GovernanceReport (frozen).

6. 数字で見る本日 (2026-05-21) 着地

指標	値
evolutionary module 数 (本日終了時)	29 (+5)
本日追加 test ケース	130 (41 + 28 + 26 + 16 + 19)
ruff `src/llive/perf/evolutionary` 警告	0 (-7)
本日着地 module	5 (`quality_diversity / coevolution_governance / persona_import / persona_survival / persona_corpus_loader`)
CE-IDs カバー率	34 / 34 ID 全カバー (skeleton 含む)
CHANGELOG `[0.6.0a1]` セクション	E.17 / E.12 / E.4 sections + 41 行追加
docs/release/v0.6.0a1_PR_PLAN.md	新規 — 5 PR 分割計画
docs/rust_hotspot_v0E_addendum.md	新規 — RUST-15〜18 spec
連載 #24 記事 (本セッション draft)	7 本 (#24-02 / 03 / 04 / 05 / 06 / 07 / 08)

7. 先行研究 9 件 (本記事の骨を作る)

Hillis, W. D. (1990). Coevolving parasites improve simulated evolution. Physica D.
Mouret, J.-B. & Clune, J. (2015). Illuminating search spaces by mapping elites. arXiv:1504.04909.
Lehman, J. & Stanley, K. (2008/2011). Novelty Search.
Stanley, K. & Miikkulainen, R. (2002). NEAT. Evolutionary Computation.
Deb, K. et al. (2002). NSGA-II. IEEE Trans Evol Comp.
Cohoon, J. (1987). Island Model GA.
Goldberg, D. & Richardson, J. (1987). Fitness sharing.
Helmuth, T. et al. (2014). Lexicase Selection.
AlphaStar (Vinyals et al. 2019). League / Exploiter / Main Pool.

8. 三重縞 — 思考因子 / persona / TRIZ の 3 層同居

ユーザー言語化の concept. 派生個体内で 3 層が同居する:

layer 1: 10 思考因子 vector (factor_structurize / ... / factor_reality_link)
layer 2: persona composition (Newton + Galois の hybrid 等)
layer 3: TRIZ 40 原理 + ARIZ 思考プロセス

の 3 layer が 同時並走. 1 派生個体が「Galois 風 + 多視点重視 + TRIZ
Segmentation を好む」のように multi-dimensional な個性を持つ. E.17
quality-diversity の MAP-Elites grid はこの 3 layer の交差点を grid 化する
最初の機構.

9. Rust addendum (#24-04 と #24-05 を繋ぐ)

docs/rust_hotspot_v0E_addendum.md (本日新規) で RUST-15 〜 18 を spec 化:

RUST-15: persona_dissimilarity Rust 化 (5x gate)
RUST-16: collusion_score (peer matrix metrics) Rust 化
RUST-17: NoveltyScorer L2 + top-k batch Rust 化
RUST-NEW-B: MAPElites bin + submit batch Rust 化
RUST-18: parity test harness 拡張

これは B-series の Python 最適化 と 集団進化の Rust 最適化 が
直交することを示す: B-series は推論 hot path (Python のままで 28%), 集団進化は
N=64 派生の集計系 hot path (Rust 化で 5-15x 狙い).

10. honest disclosure

「v0.E の効果」はベンチ未取得 — module は全 PASS だが「30 世代で
baseline より 30% diversity 維持」のような仮説 H10 / H11 は 未検証.
ベンチ走らせるのは credential + GPU 確保後.
PERSONA_ONTOLOGY 10 名は heuristic — factor_affinity vector は伝記 /
哲学史ベースの人為的初期値. CE-23 PersonaCorpusLoader でコーパスベースに
置換予定だが現状は経験則.
Governance skeleton は wire-in 未完 — Quarantined Memory への
実書込み は別 module 委譲. 完成までは 1-2 セッション.
N=64 派生集団は実機未実行 — 本セッションは module + test 着地まで.
end-to-end 集団 GA loop の実機 run は次セッション.
CE-23 LLM extractor は未実装 — keyword fallback のみ着地. LLM 経由の
thought pattern 抽出は credential 復旧後.
AlphaStar League mode (E.5) は未着手 — credential / judge LLM 復旧後.
Debate mode (E.6) も未着手 — 同上.

11. Mermaid — v0.E 全体像

12. 期待値 — 次に来るもの

v0.7 Rust 高速化: docs/rust_hotspot_v0E_addendum.md の RUST-15〜18.
v0.E E.5 (League mode) — AlphaStar 風 Main / Exploiter / League Exploiter.
v0.E E.6 (Debate mode) — Irving 2018 風 argument / counter-argument +
human/LLM judge. human / LLM judge 統合が次の明確な一手.
lleval bridge v0.1.0a2 — 派生 Genome → ProviderSpec mapper の実装.
CE-19/23 LLM extractor — Raptor RAD コーパスから persona 自動抽出.
集団進化 end-to-end 実機 run — N=64 派生で 30 世代 → diversity
metrics / collusion 検知率 / governance trigger 数を計測.

13. 2026-05-22 追記 — Rust 高速化 RUST-15/16/17 着地

goal_release_ready_v0E_rust addendum の 3 kernel を 1 セッションで着地.
連載中核記事として最新成果を反映:

13.1 着地 3 kernel

ID	機能	hot path	5x gate 結果
RUST-15 persona_dissimilarity_pairwise	NxN pair の Jaccard + L2 + 合成	PersonaOverlapPenalty.apply	avg x12.71 (N=64 で x17.07)
RUST-16 collusion_score_kernel	NxN peer matrix の variance / symmetry / concentration	CoevolutionGovernance.evaluate_generation	avg x66.70 (N=8 で x115.04)
RUST-17 novelty_score_batch	集団 N × archive A の L2 + top-k mean	NoveltyScorer.novelty_batch	avg x5.01 (A=50 で x9.55, A=1000 で x1.72)

全 37 parity test PASS (1e-6 tolerance), ruff src/llive/perf/evolutionary +
src/llive/rust_ext 0 警告.

13.2 衝撃の honest disclosure — 「Rust 化 = 速い」は嘘

RUST-15 単発呼出は Rust の方が遅い (x0.80, FAIL). FFI overhead で
Python set 操作に負ける. batch (N×N pair を 1 FFI call) にして初めて
x12.71 まで伸びる. 同じアルゴリズム・同じ Rust kernel でも FFI 境界の
引き方で結果が桁違い.

逆例も観察: RUST-16 は単発でも x66.70 で圧勝. numpy の np.nanvar /
np.corrcoef は 小 NxN (N が 100 未満) で Python overhead が支配的で 200μs+/call.
Rust の単純 C ループ (numpy zero-copy 受領) は 2μs/call.

そして境界線: RUST-17 は archive サイズで結果が反転. A=50 で x9.55 だが
A=1000 で numpy BLAS vectorized が追いついて x1.72 まで縮む.

13.3 5 パターン判定表 (本セッションで言語化)

Python 経路の特性	Rust 化の単発 ROI	実例
A 純 Python ループ (numpy 不使用) の 1-pair	単発 FAIL, batch 必須	RUST-15 (x0.80 → batch x12.71)
B numpy 大 array (1000 超) vectorized	伸びない (numpy 内部 BLAS)	(該当 kernel まだ無し)
C numpy 小 NxN (100 未満) API 多用	単発でも 10-100x	RUST-16 (x66.70)
D numpy 中規模 BLAS 1 関数	境界線上: 小サイズ Rust 圧勝, 大サイズで追いつかれる	RUST-17 (A=50 x9.55 → A=1000 x1.72)
E 冷たいデータ境界 (dict / 文字列)	overhead 大, batch 必須	—

詳細表は docs/perf_comparison/2026-05-22_kernel_implementation_comparison.md.

13.4 Cython 経路の脱落 (build chain 不在)

scratch 比較で Cython kernel を書いて 3 way 比較を試みたが Windows MSVC
build tools 不在 + mingw が MSVC Python と incompatible で build 不可.
これは「数値計算が同等に書ける」だけでは言語選択に足りない実例:
build chain が確立できるかが必須条件. source は scratch/cython_collusion/
に保存し Linux/WSL で再試行できる形に.

13.5 RUST-17b 追記 (2026-05-22 同日): rayon 並列 + quickselect で全 A 5x clear

RUST-17 baseline は archive 大 (A=200/1000) で gate FAIL だったが, 同日中に
RUST-17b として 2 手段で再実装:

rayon par_iter で N=64 集団ループを 8-core 並列化 + py.allow_threads
で GIL release
Vec::select_nth_unstable_by (Hoare quickselect, O(A) avg) で top-k
partial sort — O(A log A) full sort を置換

結果:

archive	RUST-17 (naive)	RUST-17b	改善率
A=50	x9.55	x12.83	+34%
A=200	x3.76 (FAIL)	x8.71 (PASS)	+132%
A=1000	x1.72 (FAIL)	x6.41 (PASS)	+273%
avg	x5.01	x9.32	+86%

判定表 (D) 「numpy 中規模 batch」を「境界線上 → 並列化で挽回可能」へ
update. 「naive 二重ループは負ける」だけでなく「rayon + algorithmic 改善で
圧勝に転じる」が示された.

std::simd は nightly のみで stable 不可 → 入れればさらに 2-3x. RUST-17c 候補.

13.6 次に来るもの (2026-05-22 時点で計画済)

PyBind11 + C/C++ ctypes 経路の 3 kernel scratch 比較 (queue 投入済).
RUST-17c — std::simd (Rust nightly に切替) で SIMD 4-lane 化.
月次 re-measure — env drift / numpy minor up / Rust nightly 等で
結果が動くため周期実行 (queue 投入済).
callers 切替 — PersonaOverlapPenalty.apply / NoveltyScorer.novelty_batch /
CoevolutionGovernance を rust_ext 経路に切替える PR.

14. References

Hillis, W. D. (1990). Coevolving parasites improve simulated evolution. Physica D.
Mouret, J.-B. & Clune, J. (2015). Illuminating search spaces by mapping elites. arXiv:1504.04909.
Lehman, J. & Stanley, K. (2008/2011). Novelty Search.
Stanley, K. & Miikkulainen, R. (2002). NEAT. Evolutionary Computation.
Deb, K. et al. (2002). NSGA-II. IEEE Trans Evol Comp.
Vinyals, O. et al. (2019). Grandmaster level in StarCraft II (AlphaStar). Nature.
完全リストは v0.6.0a1 リリース時に references.bib に同梱予定.

Series Navigation

English

llive Complete Guide (5) — "The Population that Learns": v0.B/C/D/E derived-population evolution summary

Concept hook: Rather than one AI getting smarter, 64 AIs turn
generations, evaluate one another, and the Approval Bus stops false
consensus — that is llive's v0.E. In the 2026-05-21 marathon that
architecture came together up to 303 tests + 0 ruff warnings + a
governance skeleton landed. The result of compressing 30 years of
lineage — from Hillis 1990 to AlphaStar 2019 — into a single OSS.

This article is the centerpiece of the #24 series. It summarizes in one
piece the four stages: v0.B (Genome / EvolutionLoop) → v0.C (subprocess
isolation) → v0.D (self-adaptive + meta mutation) → v0.E (peer evaluation +
persona + governance).

0. Position within the series — the centerpiece

#24-00 series index
#24-01 4-layer memory      ← "memory inside an individual"
#24-02 thought factors × COG-MESH ← "thought axes inside an individual"
#24-03 structural evolution × TRIZ × Z3 ← "structure rewriting inside an individual"
#24-04 B-series           ← "convergence inside an individual (fast cerebellum)"
#24-05 EvolutionLoop      ← "exploration across individuals (slow cerebrum)" ★ this article
#24-06 LLM backend         ← "the pipe that drives an individual"
#24-07 governance         ← "audit of cross-individual decisions"
#24-08 lleval              ← "the glasses that measure an individual"

#24-05 is the backbone of the whole. v0.B/C/D/E builds "the derived
population itself". The other articles are features that sit on top of it.
This is the series centerpiece — the substrate that all other chapters'
features sit on.

1. Why population-based evolution — the Hillis warning

What W. D. Hillis (1990) showed is that when the evaluator and the
evaluatee evolve simultaneously, the fitness landscape gets exponentially
more interesting. The Red Queen Effect drives the quality of the whole
population upward on its own. Keep selecting a single best and you fall
into a local optimum.

llive brought this into the LLM. A derived population of N=64 evaluates one
another, the evaluation results are fitness, and fitness drives the next
generation's selection. Then:

"the quality of the evaluators" itself rises across generations
no single best can dominate the whole
collusion where "all variants hand each other false high scores" can
occur (detected by CE-06)

2. v0.B — Genome / EvolutionLoop / parallel scheduler

v0.B core is classic GA. The landed modules are Genome, Selection,
Crossover, Mutation, scheduler:

Genome (real-valued vector + bounds + labels) + Individual + Population.
TournamentSelection / RouletteSelection / ElitismSelection.
UniformCrossover / BlendCrossover / SegmentCrossover.
GaussianMutation / ResetMutation / ChainedMutation.
EvolutionLoop (EvolutionConfig + EvolutionResult).
3 parallel schedulers: serial_scheduler / MultiprocessingScheduler / AsyncioScheduler.

With just this, the loop "population → evaluation → selection → mating →
mutation → next generation" turns.

3. v0.C — subprocess isolation + variant live run

LLM inference wants each derived individual fully isolated in its own OS
process. Reasons:

LLM is heavy → physically isolate memory leaks / GIL contention
if one variant crashes, the others survive
fault isolation via OS-level timeout / SIGKILL

VariantSubprocessScheduler (subprocess_scheduler.py) — subprocess.run +
ThreadPool parallelism + timeout + retries + cleanup. With this you can launch
the variant_runner.py script as a single derived individual.

4. v0.D — self-referential mutation (Schwefel σSA-ES + meta mutation)

v0.D core is "evolve the mutation rate itself".

SelfAdaptiveGaussianMutation (Schwefel σSA-ES, log-normal σ update).
Embeds a σ vector into the Genome, and the mutation rewrites σ too.
MetaMutation (strategy_id into the genome; 4 strategies run in parallel
within the population).
pack_self_adaptive_bounds / pack_meta_strategy_bounds — turning into 38/20/39 dim.

With this, "which mutation strategy works for the current problem" itself
is learned across generations.

5. v0.E — peer evaluation + persona ontology + governance

v0.E core. Contains CE-01..34. The main modules are below:

5.1 Evaluation (CE-01..05)

PeerEvaluationMatrix — an N×N scoring matrix. 3 collusion-detection metrics
(score_variance / symmetry / concentration). Mermaid visualization.
PeerFitnessAdapter — compatible with EvolutionLoop.scheduler.
EvaluationStyleGenome — embeds an evaluation persona dim of "harsh /
lenient / precision / speed" into the derived individual.

5.2 Diversity preservation (CE-24..29)

latin_hypercube_population — a spatially even initial population (scipy.stats.qmc).
NoveltyScorer — k-NN, Lehman-Stanley 2008/2011.
DiversityPreservingBreedFilter — novelty rejection + resample.
DiversityMonitor — diversity_l2 / spread / median + threshold alarm.

5.3 Quality Diversity (CE-25 / CE-26, landed today)

PersonaOverlapPenalty — adds the population mean of persona dissimilarity onto the fitness axis.
MAPElitesGrid — the 4-axis version of Mouret & Clune 2015 (persona 2 × thought_factor 2).
Stores the max-fitness individual in each cell.

5.4 Historical persona (CE-19..23)

PERSONA_ONTOLOGY 10 figures (Oka Kiyoshi / Grothendieck / Feynman / Galois /
von Neumann / Newton / Kant / Socrates / Laozi / Sun Tzu).
PersonaComposition (3 policies: exclusive / mix / moderator).
PersonaCompositionMutation (CE-21).
persona_dissimilarity — Jaccard + L2 of factor_affinity.
PersonaImportAlgorithm (CE-20, landed today) — partial persona adoption between derived individuals.
PersonaSurvivalAnalysis (CE-22, landed today) — statistics of which persona
combinations survived across generations.
PersonaCorpusLoader (CE-23, skeleton landed today) — automatic extraction
from Raptor RAD.

5.5 Population combination mechanisms (CE-30..34)

MutualScorePairSelector (CE-30, mating.py) — assortative mating,
softmax sampling.
NSGA2Selection (CE-31, nsga2.py) — Pareto front + crowding distance.
Speciation (CE-32, speciation.py) — NEAT-style speciation.
IslandModel (CE-33, island_model.py) — ring/fully/star 3 topologies +
best/random/worst migration.
LexicaseSelection (CE-34, mating.py) — Helmuth 2014, case-by-case ranking.

5.6 Governance (CE-06..08, landed today as E.4)

CollusionDetector (CE-06) — wraps is_suspected_collusion in a threshold
dataclass.
CoevolutionGovernance (CE-07) — collusion suspicion → fires ApprovalBus.request.
collusion_risk_score (CE-08) — state fed into TonicRiskMonitor.tick → [0, 1] risk.
GovernanceReport (frozen).

6. Today's (2026-05-21) landing by the numbers

Metric	Value
number of evolutionary modules (at end of day)	29 (+5)
test cases added today	130 (41 + 28 + 26 + 16 + 19)
ruff `src/llive/perf/evolutionary` warnings	0 (-7)
modules landed today	5 (`quality_diversity / coevolution_governance / persona_import / persona_survival / persona_corpus_loader`)
CE-ID coverage	34 / 34 IDs fully covered (skeleton included)
CHANGELOG `[0.6.0a1]` section	E.17 / E.12 / E.4 sections + 41 lines added
docs/release/v0.6.0a1_PR_PLAN.md	new — 5-PR split plan
docs/rust_hotspot_v0E_addendum.md	new — RUST-15..18 spec
#24 series articles (drafted this session)	7 (#24-02 / 03 / 04 / 05 / 06 / 07 / 08)

7. 9 prior works forming the backbone of this article

Hillis, W. D. (1990). Coevolving parasites improve simulated evolution. Physica D.
Mouret, J.-B. & Clune, J. (2015). Illuminating search spaces by mapping elites. arXiv:1504.04909.
Lehman, J. & Stanley, K. (2008/2011). Novelty Search.
Stanley, K. & Miikkulainen, R. (2002). NEAT. Evolutionary Computation.
Deb, K. et al. (2002). NSGA-II. IEEE Trans Evol Comp.
Cohoon, J. (1987). Island Model GA.
Goldberg, D. & Richardson, J. (1987). Fitness sharing.
Helmuth, T. et al. (2014). Lexicase Selection.
AlphaStar (Vinyals et al. 2019). League / Exploiter / Main Pool.

8. Triple stripe — coexistence of thought factors / persona / TRIZ across 3 layers

A user-articulated concept. Inside each derived individual, three layers coexist:

layer 1: a 10-thought-factor vector (factor_structurize / ... / factor_reality_link)
layer 2: persona composition (e.g. a Newton + Galois hybrid)
layer 3: TRIZ 40 principles + ARIZ thought process

these 3 layers run in parallel at the same time. A single derived
individual carries a multi-dimensional personality, like "Galois-style +
multi-perspective focus + prefers TRIZ Segmentation". The MAP-Elites grid of
E.17 quality-diversity is the first mechanism to grid the intersection of
these 3 layers.

9. Rust addendum (bridging #24-04 and #24-05)

docs/rust_hotspot_v0E_addendum.md (new today) specs RUST-15 .. 18:

RUST-15: Rust-port persona_dissimilarity (5x gate)
RUST-16: Rust-port collusion_score (peer matrix metrics)
RUST-17: Rust-port NoveltyScorer L2 + top-k batch
RUST-NEW-B: Rust-port MAPElites bin + submit batch
RUST-18: extend the parity test harness

This shows that the Python optimization of the B-series and the Rust
optimization of population evolution are orthogonal: the B-series is an
inference hot path (28% while staying in Python), while population evolution
is an aggregation-style hot path of the N=64 derived population (aiming for
5-15x via Rust).

10. honest disclosure

"The effect of v0.E" has no benchmark yet — the modules all PASS, but
hypotheses like H10 / H11 ("preserve 30% diversity over baseline at 30
generations") are not yet verified. Running the benchmark waits until
credentials + GPU are secured.
The 10 PERSONA_ONTOLOGY figures are heuristic — the factor_affinity
vector is an artificial initial value based on biography / history of
philosophy. It is to be replaced with a corpus-based one via CE-23
PersonaCorpusLoader, but it is currently a rule of thumb.
The governance skeleton is not wired in yet — the actual write into
Quarantined Memory is delegated to a separate module. 1-2 sessions to
completion.
The N=64 derived population has not run on real hardware — this session
reached module + test landing only. The real run of the end-to-end
population GA loop is next session.
The CE-23 LLM extractor is not implemented — only a keyword fallback
landed. Thought-pattern extraction via the LLM waits until credentials are
restored.
AlphaStar League mode (E.5) is not started — waits until credentials /
judge LLM are restored.
Debate mode (E.6) is also not started — likewise.

11. Mermaid — v0.E overview

12. Expectations — what comes next

v0.7 Rust speedup: RUST-15..18 in docs/rust_hotspot_v0E_addendum.md.
v0.E E.5 (League mode) — AlphaStar-style Main / Exploiter / League Exploiter.
v0.E E.6 (Debate mode) — Irving 2018-style argument / counter-argument +
human/LLM judge. Human / LLM judge integration is the obvious next step.
lleval bridge v0.1.0a2 — implement the derived Genome → ProviderSpec mapper.
CE-19/23 LLM extractor — automatic persona extraction from the Raptor RAD corpus.
end-to-end real run of population evolution — N=64 derived over 30
generations → measure diversity metrics / collusion detection rate /
governance trigger count.

13. 2026-05-22 addendum — Rust speedup RUST-15/16/17 landed

Landed the 3 kernels from the goal_release_ready_v0E_rust addendum in a
single session. Reflecting the latest results as the centerpiece of the series:

13.1 The 3 landed kernels

ID	Function	hot path	5x gate result
RUST-15 persona_dissimilarity_pairwise	Jaccard + L2 + composition of NxN pairs	PersonaOverlapPenalty.apply	avg x12.71 (x17.07 at N=64)
RUST-16 collusion_score_kernel	variance / symmetry / concentration of the NxN peer matrix	CoevolutionGovernance.evaluate_generation	avg x66.70 (x115.04 at N=8)
RUST-17 novelty_score_batch	L2 + top-k mean of population N × archive A	NoveltyScorer.novelty_batch	avg x5.01 (x9.55 at A=50, x1.72 at A=1000)

All 37 parity tests PASS (1e-6 tolerance), 0 ruff warnings in
src/llive/perf/evolutionary + src/llive/rust_ext.

13.2 The shocking honest disclosure — "Rust = fast" is a lie

A single RUST-15 call is slower in Rust (x0.80, FAIL). With FFI overhead it
loses to a Python set operation. Only when made into a batch (N×N pairs in one
FFI call) does it stretch to x12.71. Even with the same algorithm and the same
Rust kernel, the result is orders of magnitude apart depending on how you draw
the FFI boundary.

The reverse was also observed: RUST-16 wins outright even on a single call at
x66.70. numpy's np.nanvar / np.corrcoef are dominated by Python overhead at
small NxN (N below 100), costing 200μs+/call. The simple C loop in Rust
(receiving numpy zero-copy) is 2μs/call.

And the borderline: RUST-17 flips with archive size. x9.55 at A=50, but at
A=1000 numpy BLAS vectorization catches up and it shrinks to x1.72.

13.3 The 5-pattern decision table (articulated this session)

Characteristic of the Python path	single-call ROI of Rust port	Example
A 1-pair of a pure Python loop (no numpy)	single-call FAIL, batch required	RUST-15 (x0.80 → batch x12.71)
B large numpy array (over 1000) vectorized	no gain (internal numpy BLAS)	(no matching kernel yet)
C small numpy NxN (below 100) with heavy API use	10-100x even on a single call	RUST-16 (x66.70)
D a single mid-scale numpy BLAS function	on the borderline: Rust wins at small size, gets caught at large size	RUST-17 (A=50 x9.55 → A=1000 x1.72)
E a cold data boundary (dict / strings)	large overhead, batch required	—

The detailed table is in docs/perf_comparison/2026-05-22_kernel_implementation_comparison.md.

13.4 The Cython path dropped out (no build chain)

In the scratch comparison we wrote a Cython kernel to attempt a 3-way
comparison, but with no Windows MSVC build tools + mingw incompatible with
MSVC Python it could not build. This is a worked example that "being able to
write the numerics equivalently" alone is not enough for language selection:
whether the build chain can be established is a necessary condition. The
source is saved in scratch/cython_collusion/ in a form that can be retried on
Linux/WSL.

13.5 RUST-17b addendum (same day, 2026-05-22): rayon parallelism + quickselect clears 5x for all A

The RUST-17 baseline gate FAILed at large archives (A=200/1000), but the same
day it was reimplemented as RUST-17b via 2 means:

rayon par_iter parallelizes the N=64 population loop across 8 cores +
py.allow_threads releases the GIL
Vec::select_nth_unstable_by (Hoare quickselect, O(A) avg) for the top-k
partial sort — replacing an O(A log A) full sort

Result:

archive	RUST-17 (naive)	RUST-17b	improvement
A=50	x9.55	x12.83	+34%
A=200	x3.76 (FAIL)	x8.71 (PASS)	+132%
A=1000	x1.72 (FAIL)	x6.41 (PASS)	+273%
avg	x5.01	x9.32	+86%

Decision-table entry (D) "mid-scale numpy batch" is updated to "on the
borderline → recoverable via parallelism". It was shown that not only does
"the naive double loop lose" but also "**it turns into an outright win via rayon

algorithmic improvement**".

std::simd is nightly-only and unavailable on stable → adding it would give
another 2-3x. A RUST-17c candidate.

13.6 What comes next (already planned as of 2026-05-22)

A 3-kernel scratch comparison of the PyBind11 + C/C++ ctypes path
(already queued).
RUST-17c — SIMD 4-lane via std::simd (switching to Rust nightly).
monthly re-measure — because env drift / numpy minor bumps / Rust nightly
etc. move the results, run it periodically (already queued).
caller switchover — a PR to switch PersonaOverlapPenalty.apply /
NoveltyScorer.novelty_batch / CoevolutionGovernance to the rust_ext path.

14. References

Hillis, W. D. (1990). Coevolving parasites improve simulated evolution. Physica D.
Mouret, J.-B. & Clune, J. (2015). Illuminating search spaces by mapping elites. arXiv:1504.04909.
Lehman, J. & Stanley, K. (2008/2011). Novelty Search.
Stanley, K. & Miikkulainen, R. (2002). NEAT. Evolutionary Computation.
Deb, K. et al. (2002). NSGA-II. IEEE Trans Evol Comp.
Vinyals, O. et al. (2019). Grandmaster level in StarCraft II (AlphaStar). Nature.
The full list will be bundled in references.bib at the v0.6.0a1 release.

Series Navigation

← Prev: llive Complete Guide (4) "The Converging Brain"
→ Next: llive Complete Guide (6) "Beyond the Transformer"
All: llive Complete Guide (0) — series index
repo: furuse-kazufumi/llive

中文

llive 完全解说 (5) — "学习的群体": v0.B/C/D/E 派生群体进化总结

概念 hook: 不是 1 个 AI 变聪明, 而是 64 个 AI 跑世代互相评估, 虚假的合意由
Approval Bus 制止 — 那就是 llive 的 v0.E. 在 2026-05-21 marathon 中, 该架构
凑齐到 303 个 test + ruff 0 警告 + governance skeleton 着地. 这是把从 Hillis
1990 到 AlphaStar 2019 的 30 年谱系压缩进 1 个 OSS 的结果.

本文是连载 #24 的核心. 把 v0.B (Genome / EvolutionLoop) → v0.C (subprocess
隔离) → v0.D (self-adaptive + meta mutation) → v0.E (peer evaluation +
persona + governance) 这 4 个阶段 总结在 1 篇里.

0. 在系列中的定位 — 本系列的核心

#24-00 series index
#24-01 4 层记忆          ← 「个体内的记忆」
#24-02 思考因子 × COG-MESH ← 「个体内的思考轴」
#24-03 结构进化 × TRIZ × Z3 ← 「个体内的结构改写」
#24-04 B-series           ← 「个体内的收敛 (快速小脑)」
#24-05 EvolutionLoop      ← 「个体间的探索 (慢速大脑)」 ★ 本文
#24-06 LLM backend         ← 「驱动个体的管道」
#24-07 governance         ← 「个体间决策的 audit」
#24-08 lleval              ← 「测量个体的眼镜」

#24-05 是整体的脊梁. v0.B/C/D/E 做出「派生群体本身」. 其他文章是建立在
其上的功能. 这是系列核心 — 其他所有章节的功能都建立在它之上.

1. 为什么选群体进化 — Hillis 的警告

W. D. Hillis (1990) 证明的是「评估者与被评估者同时进化」时, fitness
landscape 会指数级地更有趣. Red Queen Effect 让整个群体的质量 自走上升.
持续只选单一 best 会 陷入局部最优.

llive 把这带进了 LLM. 派生群体 N=64 互相评估, 评估结果即 fitness, fitness 即
下一代的 selection. 于是:

「评估者的质量」本身随世代上升
单一 best 无法支配整体
「全派生互相打虚假高分」的共谋 可能发生 (由 CE-06 检测)

2. v0.B — Genome / EvolutionLoop / 并行 scheduler

v0.B 内核是经典 GA. 已着地模块为 Genome, Selection, Crossover, Mutation,
scheduler:

Genome (实数 vector + bounds + labels) + Individual + Population.
TournamentSelection / RouletteSelection / ElitismSelection.
UniformCrossover / BlendCrossover / SegmentCrossover.
GaussianMutation / ResetMutation / ChainedMutation.
EvolutionLoop (EvolutionConfig + EvolutionResult).
3 种并行 scheduler: serial_scheduler / MultiprocessingScheduler / AsyncioScheduler.

仅此就能让「群体 → 评估 → 选别 → 交配 → 突变 → 下一代」的循环转起来.

3. v0.C — subprocess 隔离 + 派生实际运行

LLM 推断希望每个派生个体都在独立 OS 进程中 完全隔离. 原因如下:

LLM 重 → 把内存 leak / GIL 竞争物理隔离
1 个派生挂了其他仍存活
用 OS-level timeout / SIGKILL 做 fault isolation

VariantSubprocessScheduler (subprocess_scheduler.py) — subprocess.run +
ThreadPool 并行 + timeout + retries + cleanup. 由此可把 variant_runner.py
脚本作为 1 个派生个体启动.

4. v0.D — 自我参照 mutation (Schwefel σSA-ES + meta mutation)

v0.D 内核是「让 mutation rate 本身也进化」.

SelfAdaptiveGaussianMutation (Schwefel σSA-ES, log-normal σ update).
把 σ vector 嵌入 Genome, mutation 也改写 σ.
MetaMutation (把 strategy_id 放进 genome, 群体内 4 策略并跑).
pack_self_adaptive_bounds / pack_meta_strategy_bounds — 化为 38/20/39 dim.

由此「哪种 mutation 策略对当前问题有效」本身也被跨世代学习.

5. v0.E — peer 评估 + persona ontology + governance

v0.E 内核. 包含 CE-01..34. 主要模块如下:

5.1 评估 (CE-01..05)

PeerEvaluationMatrix — N×N 打分矩阵. 共谋检测 3 指标
(score_variance / symmetry / concentration). Mermaid 可视化.
PeerFitnessAdapter — 与 EvolutionLoop.scheduler 兼容.
EvaluationStyleGenome — 给派生嵌入「辛辣 / 宽松 / 精度 / 速度」的
evaluation persona dim.

5.2 多样性保护 (CE-24..29)

latin_hypercube_population — 空间均匀的初始群体 (scipy.stats.qmc).
NoveltyScorer — k-NN, Lehman-Stanley 2008/2011.
DiversityPreservingBreedFilter — novelty rejection + resample.
DiversityMonitor — diversity_l2 / spread / median + 阈值 alarm.

5.3 Quality Diversity (CE-25 / CE-26, 本日着地)

PersonaOverlapPenalty — 在 fitness 轴上加上 persona dissimilarity 的群体平均.
MAPElitesGrid — Mouret & Clune 2015 的 4 轴版 (persona 2 × thought_factor 2).
在每个 cell 保存最大 fitness 个体.

5.4 Historical persona (CE-19..23)

PERSONA_ONTOLOGY 10 名 (冈洁 / 格罗滕迪克 / 费曼 / 伽罗瓦 /
冯·诺依曼 / 牛顿 / 康德 / 苏格拉底 / 老子 / 孙子).
PersonaComposition (3 policy: exclusive / mix / moderator).
PersonaCompositionMutation (CE-21).
persona_dissimilarity — Jaccard + L2 of factor_affinity.
PersonaImportAlgorithm (CE-20, 本日着地) — 派生间 persona 部分采用.
PersonaSurvivalAnalysis (CE-22, 本日着地) — 哪种 persona 组合
跨世代存活的统计.
PersonaCorpusLoader (CE-23, 本日着地 skeleton) — 从 Raptor RAD
自动抽取.

5.5 群体组合机制 (CE-30..34)

MutualScorePairSelector (CE-30, mating.py) — assortative mating,
softmax sampling.
NSGA2Selection (CE-31, nsga2.py) — Pareto front + crowding distance.
Speciation (CE-32, speciation.py) — NEAT 流的种分.
IslandModel (CE-33, island_model.py) — ring/fully/star 3 topology +
best/random/worst migration.
LexicaseSelection (CE-34, mating.py) — Helmuth 2014, case-by-case 排序.

5.6 Governance (CE-06..08, 本日着地 E.4)

CollusionDetector (CE-06) — 把 is_suspected_collusion 用 threshold
dataclass 包装.
CoevolutionGovernance (CE-07) — 共谋疑似 → 触发 ApprovalBus.request.
collusion_risk_score (CE-08) — 投入 TonicRiskMonitor.tick 的
state → [0, 1] risk.
GovernanceReport (frozen).

6. 用数字看本日 (2026-05-21) 的着地

指标	值
evolutionary module 数 (本日结束时)	29 (+5)
本日新增 test 用例	130 (41 + 28 + 26 + 16 + 19)
ruff `src/llive/perf/evolutionary` 警告	0 (-7)
本日着地 module	5 (`quality_diversity / coevolution_governance / persona_import / persona_survival / persona_corpus_loader`)
CE-IDs 覆盖率	34 / 34 ID 全覆盖 (含 skeleton)
CHANGELOG `[0.6.0a1]` section	E.17 / E.12 / E.4 sections + 41 行新增
docs/release/v0.6.0a1_PR_PLAN.md	新 — 5 PR 拆分计划
docs/rust_hotspot_v0E_addendum.md	新 — RUST-15..18 spec
连载 #24 文章 (本会话 draft)	7 篇 (#24-02 / 03 / 04 / 05 / 06 / 07 / 08)

7. 9 项先行研究 (构成本文骨架)

Hillis, W. D. (1990). Coevolving parasites improve simulated evolution. Physica D.
Mouret, J.-B. & Clune, J. (2015). Illuminating search spaces by mapping elites. arXiv:1504.04909.
Lehman, J. & Stanley, K. (2008/2011). Novelty Search.
Stanley, K. & Miikkulainen, R. (2002). NEAT. Evolutionary Computation.
Deb, K. et al. (2002). NSGA-II. IEEE Trans Evol Comp.
Cohoon, J. (1987). Island Model GA.
Goldberg, D. & Richardson, J. (1987). Fitness sharing.
Helmuth, T. et al. (2014). Lexicase Selection.
AlphaStar (Vinyals et al. 2019). League / Exploiter / Main Pool.

8. 三重条纹 — 思考因子 / persona / TRIZ 三层并存

用户语言化的概念. 在每个派生个体内, 三层并存:

layer 1: 10 思考因子 vector (factor_structurize / ... / factor_reality_link)
layer 2: persona composition (Newton + Galois 的 hybrid 等)
layer 3: TRIZ 40 原理 + ARIZ 思考过程

这 3 layer 同时并跑. 1 个派生个体拥有 multi-dimensional 的个性, 如
「Galois 风 + 重视多视角 + 偏好 TRIZ Segmentation」. E.17 quality-diversity
的 MAP-Elites grid 是把这 3 layer 的交叉点 grid 化的最初机制.

9. Rust 附录 (连接 #24-04 和 #24-05)

docs/rust_hotspot_v0E_addendum.md (本日新) 规定了 RUST-15 .. 18:

RUST-15: 把 persona_dissimilarity Rust 化 (5x gate)
RUST-16: 把 collusion_score (peer matrix metrics) Rust 化
RUST-17: 把 NoveltyScorer L2 + top-k batch Rust 化
RUST-NEW-B: 把 MAPElites bin + submit batch Rust 化
RUST-18: 扩展 parity test harness

这表明 B-series 的 Python 优化 与 群体进化的 Rust 优化 正交: B-series 是
推理 hot path (保持 Python 而 28%), 群体进化是 N=64 派生的聚合系 hot path
(Rust 化目标 5-15x).

10. honest disclosure

「v0.E 的效果」尚未取得基准 — module 全 PASS, 但类似 H10 / H11
(「30 世代相比 baseline 多保留 30% diversity」) 这样的假设 尚未验证.
跑基准要在 credential + GPU 确保之后.
PERSONA_ONTOLOGY 10 名是 heuristic — factor_affinity vector 是基于
传记 / 哲学史的人为初始值. 计划用 CE-23 PersonaCorpusLoader 换成基于语料的,
但目前是经验法则.
Governance skeleton 的 wire-in 未完 — 向 Quarantined Memory 的
实际写入 委托给另一 module. 完成还要 1-2 会话.
N=64 派生群体未在实机运行 — 本会话只到 module + test 着地.
end-to-end 群体 GA loop 的实机 run 在下一会话.
CE-23 LLM extractor 未实现 — 只着地了 keyword fallback. 经 LLM 的
thought pattern 抽取要在 credential 恢复后.
AlphaStar League mode (E.5) 未着手 — 在 credential / judge LLM 恢复后.
Debate mode (E.6) 也未着手 — 同上.

11. Mermaid — v0.E 全貌

12. 期望值 — 接下来要做的

v0.7 Rust 加速: docs/rust_hotspot_v0E_addendum.md 的 RUST-15..18.
v0.E E.5 (League mode) — AlphaStar 风的 Main / Exploiter / League Exploiter.
v0.E E.6 (Debate mode) — Irving 2018 风的 argument / counter-argument +
human/LLM judge. human / LLM judge 整合是下一步明显的方向.
lleval bridge v0.1.0a2 — 实现派生 Genome → ProviderSpec mapper.
CE-19/23 LLM extractor — 从 Raptor RAD 语料自动抽取 persona.
群体进化 end-to-end 实机 run — N=64 派生跑 30 世代 → 测量 diversity
metrics / collusion 检测率 / governance trigger 数.

13. 2026-05-22 追记 — Rust 加速 RUST-15/16/17 落地

在一次会话中落地了 goal_release_ready_v0E_rust 附录中的 3 个 kernel.
作为连载核心文章反映最新成果:

13.1 着地 3 kernel

ID	功能	hot path	5x gate 结果
RUST-15 persona_dissimilarity_pairwise	NxN pair 的 Jaccard + L2 + 合成	PersonaOverlapPenalty.apply	avg x12.71 (N=64 时 x17.07)
RUST-16 collusion_score_kernel	NxN peer matrix 的 variance / symmetry / concentration	CoevolutionGovernance.evaluate_generation	avg x66.70 (N=8 时 x115.04)
RUST-17 novelty_score_batch	群体 N × archive A 的 L2 + top-k mean	NoveltyScorer.novelty_batch	avg x5.01 (A=50 时 x9.55, A=1000 时 x1.72)

全 37 parity test PASS (1e-6 tolerance), ruff src/llive/perf/evolutionary +
src/llive/rust_ext 0 警告.

13.2 冲击性的 honest disclosure — 「Rust 化 = 快」是谎言

RUST-15 单次调用 Rust 反而更慢 (x0.80, FAIL). 因 FFI overhead 输给了
Python set 操作. 做成 batch (N×N pair 一次 FFI call) 后才伸到 x12.71. 即使
同算法 · 同 Rust kernel, 结果也因 FFI 边界怎么划 而相差数量级.

也观察到反例: RUST-16 单次也以 x66.70 完胜. numpy 的 np.nanvar /
np.corrcoef 在 小 NxN (N 小于 100) 时 Python overhead 占主导, 200μs+/call.
Rust 的简单 C 循环 (numpy zero-copy 接收) 是 2μs/call.

还有边界线: RUST-17 随 archive 大小结果反转. A=50 为 x9.55, 但 A=1000 时
numpy BLAS vectorized 追上, 缩到 x1.72.

13.3 5 模式判定表 (本会话语言化)

Python 路径的特性	Rust 化的单次 ROI	实例
A 纯 Python 循环 (不用 numpy) 的 1-pair	单次 FAIL, 必须 batch	RUST-15 (x0.80 → batch x12.71)
B numpy 大 array (超过 1000) vectorized	不涨 (numpy 内部 BLAS)	(尚无对应 kernel)
C numpy 小 NxN (小于 100) 多用 API	单次也 10-100x	RUST-16 (x66.70)
D numpy 中规模 BLAS 1 函数	在边界线上: 小尺寸 Rust 完胜, 大尺寸被追上	RUST-17 (A=50 x9.55 → A=1000 x1.72)
E 冷数据边界 (dict / 字符串)	overhead 大, 必须 batch	—

详细表在 docs/perf_comparison/2026-05-22_kernel_implementation_comparison.md.

13.4 Cython 路径出局 (无 build chain)

在 scratch 比较中写了 Cython kernel 想做 3 way 比较, 但 Windows MSVC
build tools 缺失 + mingw 与 MSVC Python incompatible 导致无法 build.
这是「能等价地写出数值计算」并不足以做语言选择的实例:
能否确立 build chain 是必要条件. source 保存在 scratch/cython_collusion/,
以便在 Linux/WSL 重试.

13.5 RUST-17b 追记 (2026-05-22 同日): rayon 并行 + quickselect 让全 A 5x 通过

RUST-17 baseline 在 archive 大 (A=200/1000) 时 gate FAIL, 但同日内
作为 RUST-17b 用 2 个手段重新实现:

rayon par_iter 把 N=64 群体循环 8-core 并行化 + py.allow_threads
release GIL
Vec::select_nth_unstable_by (Hoare quickselect, O(A) avg) 做 top-k
partial sort — 替换 O(A log A) full sort

结果:

archive	RUST-17 (naive)	RUST-17b	改善率
A=50	x9.55	x12.83	+34%
A=200	x3.76 (FAIL)	x8.71 (PASS)	+132%
A=1000	x1.72 (FAIL)	x6.41 (PASS)	+273%
avg	x5.01	x9.32	+86%

把判定表 (D)「numpy 中规模 batch」update 为「在边界线上 → 可经并行化挽回」.
不仅「naive 双重循环会输」, 还展示了「经 rayon + algorithmic 改善转为完胜」.

std::simd 仅 nightly stable 不可 → 加上还能再 2-3x. RUST-17c 候选.

13.6 接下来要做的 (截至 2026-05-22 已计划)

PyBind11 + C/C++ ctypes 路径的 3 kernel scratch 比较 (已投入 queue).
RUST-17c — std::simd (切到 Rust nightly) 做 SIMD 4-lane 化.
月度 re-measure — 因 env drift / numpy minor up / Rust nightly 等
结果会变, 周期执行 (已投入 queue).
callers 切换 — 把 PersonaOverlapPenalty.apply / NoveltyScorer.novelty_batch /
CoevolutionGovernance 切到 rust_ext 路径的 PR.

14. References

Hillis, W. D. (1990). Coevolving parasites improve simulated evolution. Physica D.
Mouret, J.-B. & Clune, J. (2015). Illuminating search spaces by mapping elites. arXiv:1504.04909.
Lehman, J. & Stanley, K. (2008/2011). Novelty Search.
Stanley, K. & Miikkulainen, R. (2002). NEAT. Evolutionary Computation.
Deb, K. et al. (2002). NSGA-II. IEEE Trans Evol Comp.
Vinyals, O. et al. (2019). Grandmaster level in StarCraft II (AlphaStar). Nature.
完整列表将在 v0.6.0a1 发布时随 references.bib 一同提供.

Series Navigation

한국어

llive 완전 해설 (5) — "집단이 학습하는 AI": v0.B/C/D/E 파생 집단 진화 총괄

콘셉트 hook: 1개의 AI가 똑똑해지는 것이 아니라, 64개의 AI가 세대를
돌리며 서로 평가하고, 거짓 합의는 Approval Bus가 멈춘다 — 그것이 llive의
v0.E다. 2026-05-21 marathon에서 그 아키텍처가 303건 test + ruff 0 경고 +
governance skeleton 착지까지 갖춰졌다. Hillis 1990부터 AlphaStar 2019까지
30년의 계보를 1개의 OSS에 압축한 결과다.

본 글은 연재 #24의 핵심이다. v0.B (Genome / EvolutionLoop) → v0.C (subprocess
분리) → v0.D (self-adaptive + meta mutation) → v0.E (peer evaluation +
persona + governance)의 4단계를 한 편에 총괄한다.

0. 연재에서의 위치 — 본 연재의 핵심

#24-00 series index
#24-01 4층 메모리          ← 「개체 안의 기억」
#24-02 사고 인자 × COG-MESH ← 「개체 안의 사고 축」
#24-03 구조 진화 × TRIZ × Z3 ← 「개체 내의 구조 재작성」
#24-04 B-series           ← 「개체 내의 수렴 (빠른 소뇌)」
#24-05 EvolutionLoop      ← 「개체 간의 탐색 (느린 대뇌)」 ★ 본 글
#24-06 LLM backend         ← 「개체를 움직이는 관」
#24-07 governance         ← 「개체 간 결정의 audit」
#24-08 lleval              ← 「개체를 재는 안경」

#24-05는 전체의 등뼈다. v0.B/C/D/E로 「파생 집단 그 자체」를 만든다. 다른
글은 그 위에 얹히는 기능이다. 이것은 연재의 핵심 — 다른 모든 장의 기능이 얹히는
기반이다.

1. 왜 집단 진화인가 — Hillis의 경고

W. D. Hillis (1990)가 보인 것은 「평가자와 피평가자가 동시에 진화한다」면
fitness landscape가 지수적으로 더 흥미로워진다는 것이다. Red Queen Effect로
집단 전체의 질이 자주적으로 상승한다. 단일 best를 계속 고르면 국소
최적에 빠진다.

llive는 이것을 LLM에 가져왔다. 파생 집단 N=64가 서로 평가하고, 평가 결과가
fitness, fitness가 다음 세대의 selection이 된다. 그러면:

「평가자의 질」 자체가 세대와 함께 상승한다
단일 best가 전체를 지배할 수 없다
「전 파생이 거짓 고득점을 서로 매기는」 공모가 발생할 수 있다 (CE-06에서 검출)

2. v0.B — Genome / EvolutionLoop / 병렬 scheduler

v0.B core는 GA 고전이다. 착지 module은 Genome, Selection, Crossover,
Mutation, scheduler:

Genome (실수 vector + bounds + labels) + Individual + Population.
TournamentSelection / RouletteSelection / ElitismSelection.
UniformCrossover / BlendCrossover / SegmentCrossover.
GaussianMutation / ResetMutation / ChainedMutation.
EvolutionLoop (EvolutionConfig + EvolutionResult).
병렬 scheduler 3종: serial_scheduler / MultiprocessingScheduler / AsyncioScheduler.

이것만으로 「집단 → 평가 → 선별 → 교배 → 돌연변이 → 다음 세대」의 루프가 돈다.

3. v0.C — subprocess 분리 + 파생 실주행

LLM 추론은 파생 개체 1개당 OS process 1개로 완전 분리하고 싶다. 이유는:

LLM이 무겁다 → 메모리 leak / GIL 경합을 물리적으로 분리
1 파생이 죽어도 나머지는 생존
OS-level timeout / SIGKILL로 fault isolation

VariantSubprocessScheduler (subprocess_scheduler.py) — subprocess.run +
ThreadPool 병렬 + timeout + retries + cleanup. 이로써 variant_runner.py
스크립트를 파생 1개체로 기동할 수 있다.

4. v0.D — 자기 참조 mutation (Schwefel σSA-ES + meta mutation)

v0.D core는 「mutation rate 그 자체를 진화시킨다」이다.

SelfAdaptiveGaussianMutation (Schwefel σSA-ES, log-normal σ update).
Genome에 σ vector를 묻고, mutation이 σ도 재작성한다.
MetaMutation (strategy_id를 genome에, 집단 내에서 4 전략 병주).
pack_self_adaptive_bounds / pack_meta_strategy_bounds — 38/20/39 dim화.

이로써 「어떤 mutation 전략이 지금의 문제에 효과적인가」 자체가 세대를
넘어 학습된다.

5. v0.E — peer evaluation + persona ontology + governance

v0.E core. CE-01..34를 포함한다. 주요 module은 다음과 같다:

5.1 평가 (CE-01..05)

PeerEvaluationMatrix — N×N 채점 행렬. 공모 검출 3 지표
(score_variance / symmetry / concentration). Mermaid 시각화.
PeerFitnessAdapter — EvolutionLoop.scheduler와 호환.
EvaluationStyleGenome — 파생에 「신랄 / 관대 / 정밀 / 속도」의
evaluation persona dim을 묻는다.

5.2 다양성 보호 (CE-24..29)

latin_hypercube_population — 공간 균등 초기 집단 (scipy.stats.qmc).
NoveltyScorer — k-NN, Lehman-Stanley 2008/2011.
DiversityPreservingBreedFilter — novelty rejection + resample.
DiversityMonitor — diversity_l2 / spread / median + 임계값 alarm.

5.3 Quality Diversity (CE-25 / CE-26, 금일 착지)

PersonaOverlapPenalty — fitness 축에 persona dissimilarity의 집단 평균을 가산.
MAPElitesGrid — Mouret & Clune 2015의 4축 버전 (persona 2 × thought_factor 2).
각 cell에 최대 fitness 개체를 저장.

5.4 Historical persona (CE-19..23)

PERSONA_ONTOLOGY 10명 (오카 기요시 / 그로텐디크 / 파인만 / 갈루아 /
폰 노이만 / 뉴턴 / 칸트 / 소크라테스 / 노자 / 손자).
PersonaComposition (3 policy: exclusive / mix / moderator).
PersonaCompositionMutation (CE-21).
persona_dissimilarity — Jaccard + L2 of factor_affinity.
PersonaImportAlgorithm (CE-20, 금일 착지) — 파생 간 persona 부분 채용.
PersonaSurvivalAnalysis (CE-22, 금일 착지) — 어떤 persona 조합이
세대를 살아남았는지 통계.
PersonaCorpusLoader (CE-23, 금일 착지 skeleton) — Raptor RAD에서
자동 추출.

5.5 집단 조합 기구 (CE-30..34)

MutualScorePairSelector (CE-30, mating.py) — assortative mating,
softmax sampling.
NSGA2Selection (CE-31, nsga2.py) — Pareto front + crowding distance.
Speciation (CE-32, speciation.py) — NEAT 식 종 분류.
IslandModel (CE-33, island_model.py) — ring/fully/star 3 topology +
best/random/worst migration.
LexicaseSelection (CE-34, mating.py) — Helmuth 2014, case-by-case 순위.

5.6 Governance (CE-06..08, 금일 착지 E.4)

CollusionDetector (CE-06) — is_suspected_collusion를 threshold
dataclass로 wrap.
CoevolutionGovernance (CE-07) — 공모 의심 → ApprovalBus.request 발화.
collusion_risk_score (CE-08) — TonicRiskMonitor.tick에 투입하는
state → [0, 1] risk.
GovernanceReport (frozen).

6. 숫자로 본 금일 (2026-05-21) 착지

지표	값
evolutionary module 수 (금일 종료 시)	29 (+5)
금일 추가 test 케이스	130 (41 + 28 + 26 + 16 + 19)
ruff `src/llive/perf/evolutionary` 경고	0 (-7)
금일 착지 module	5 (`quality_diversity / coevolution_governance / persona_import / persona_survival / persona_corpus_loader`)
CE-IDs 커버율	34 / 34 ID 전 커버 (skeleton 포함)
CHANGELOG `[0.6.0a1]` 섹션	E.17 / E.12 / E.4 sections + 41행 추가
docs/release/v0.6.0a1_PR_PLAN.md	신규 — 5 PR 분할 계획
docs/rust_hotspot_v0E_addendum.md	신규 — RUST-15..18 spec
연재 #24 글 (본 세션 draft)	7편 (#24-02 / 03 / 04 / 05 / 06 / 07 / 08)

7. 선행 연구 9건 (본 글의 뼈대를 만든다)

Hillis, W. D. (1990). Coevolving parasites improve simulated evolution. Physica D.
Mouret, J.-B. & Clune, J. (2015). Illuminating search spaces by mapping elites. arXiv:1504.04909.
Lehman, J. & Stanley, K. (2008/2011). Novelty Search.
Stanley, K. & Miikkulainen, R. (2002). NEAT. Evolutionary Computation.
Deb, K. et al. (2002). NSGA-II. IEEE Trans Evol Comp.
Cohoon, J. (1987). Island Model GA.
Goldberg, D. & Richardson, J. (1987). Fitness sharing.
Helmuth, T. et al. (2014). Lexicase Selection.
AlphaStar (Vinyals et al. 2019). League / Exploiter / Main Pool.

8. 삼중 줄무늬 — 사고 인자 / persona / TRIZ의 3층 동거

사용자가 언어화한 concept. 파생 개체 내에서 3층이 동거한다:

layer 1: 10 사고 인자 vector (factor_structurize / ... / factor_reality_link)
layer 2: persona composition (Newton + Galois의 hybrid 등)
layer 3: TRIZ 40 원리 + ARIZ 사고 프로세스

이 3 layer가 동시 병주한다. 1 파생 개체가 「Galois 풍 + 다관점 중시 + TRIZ
Segmentation을 선호」처럼 multi-dimensional한 개성을 가진다. E.17
quality-diversity의 MAP-Elites grid는 이 3 layer의 교차점을 grid화하는 최초의
기구다.

9. Rust addendum (#24-04와 #24-05를 잇는다)

docs/rust_hotspot_v0E_addendum.md (금일 신규)에서 RUST-15 .. 18을 spec화:

RUST-15: persona_dissimilarity Rust화 (5x gate)
RUST-16: collusion_score (peer matrix metrics) Rust화
RUST-17: NoveltyScorer L2 + top-k batch Rust화
RUST-NEW-B: MAPElites bin + submit batch Rust화
RUST-18: parity test harness 확장

이것은 B-series의 Python 최적화와 집단 진화의 Rust 최적화가
직교함을 보여준다: B-series는 추론 hot path (Python 그대로 28%), 집단 진화는
N=64 파생의 집계계 hot path (Rust화로 5-15x 노림).

10. honest disclosure

「v0.E의 효과」는 벤치 미취득 — module은 전 PASS지만 「30세대에서
baseline보다 30% diversity 유지」 같은 가설 H10 / H11은 미검증.
벤치를 돌리는 것은 credential + GPU 확보 후.
PERSONA_ONTOLOGY 10명은 heuristic — factor_affinity vector는 전기 /
철학사 기반의 인위적 초기값. CE-23 PersonaCorpusLoader로 코퍼스 기반으로
교체 예정이지만 현재는 경험칙.
Governance skeleton은 wire-in 미완 — Quarantined Memory로의
실제 쓰기는 별도 module에 위임. 완성까지 1-2 세션.
N=64 파생 집단은 실기 미실행 — 본 세션은 module + test 착지까지.
end-to-end 집단 GA loop의 실기 run은 다음 세션.
CE-23 LLM extractor는 미구현 — keyword fallback만 착지. LLM 경유의
thought pattern 추출은 credential 복구 후.
AlphaStar League mode (E.5)는 미착수 — credential / judge LLM 복구 후.
Debate mode (E.6)도 미착수 — 동상.

11. Mermaid — v0.E 전체상

12. 기댓값 — 다음에 오는 것

v0.7 Rust 고속화: docs/rust_hotspot_v0E_addendum.md의 RUST-15..18.
v0.E E.5 (League mode) — AlphaStar 풍의 Main / Exploiter / League Exploiter.
v0.E E.6 (Debate mode) — Irving 2018 풍의 argument / counter-argument +
human/LLM judge. human / LLM judge 통합이 다음의 명확한 한 수.
lleval bridge v0.1.0a2 — 파생 Genome → ProviderSpec mapper의 구현.
CE-19/23 LLM extractor — Raptor RAD 코퍼스에서 persona 자동 추출.
집단 진화 end-to-end 실기 run — N=64 파생으로 30세대 → diversity
metrics / collusion 검지율 / governance trigger 수를 계측.

13. 2026-05-22 추기 — Rust 고속화 RUST-15/16/17 착지

goal_release_ready_v0E_rust addendum의 3 kernel을 1 세션에 착지.
연재 핵심 글로서 최신 성과를 반영:

13.1 착지 3 kernel

ID	기능	hot path	5x gate 결과
RUST-15 persona_dissimilarity_pairwise	NxN pair의 Jaccard + L2 + 합성	PersonaOverlapPenalty.apply	avg x12.71 (N=64에서 x17.07)
RUST-16 collusion_score_kernel	NxN peer matrix의 variance / symmetry / concentration	CoevolutionGovernance.evaluate_generation	avg x66.70 (N=8에서 x115.04)
RUST-17 novelty_score_batch	집단 N × archive A의 L2 + top-k mean	NoveltyScorer.novelty_batch	avg x5.01 (A=50에서 x9.55, A=1000에서 x1.72)

전 37 parity test PASS (1e-6 tolerance), ruff src/llive/perf/evolutionary +
src/llive/rust_ext 0 경고.

13.2 충격의 honest disclosure — 「Rust화 = 빠름」은 거짓

RUST-15 단발 호출은 Rust 쪽이 더 느리다 (x0.80, FAIL). FFI overhead로
Python set 조작에 진다. batch (N×N pair를 1 FFI call)로 만들고 나서야
x12.71까지 늘어난다. 같은 알고리즘 · 같은 Rust kernel이라도 FFI 경계를 어떻게
긋는가로 결과가 자릿수만큼 다르다.

역례도 관찰: RUST-16은 단발에서도 x66.70으로 압승. numpy의 np.nanvar /
np.corrcoef는 작은 NxN (N이 100 미만)에서 Python overhead가 지배적이라
200μs+/call. Rust의 단순 C 루프 (numpy zero-copy 수령)는 2μs/call.

그리고 경계선: RUST-17은 archive 크기로 결과가 반전. A=50에서 x9.55지만
A=1000에서 numpy BLAS vectorized가 따라잡아 x1.72까지 줄어든다.

13.3 5 패턴 판정표 (본 세션에서 언어화)

Python 경로의 특성	Rust화의 단발 ROI	실례
A 순수 Python 루프 (numpy 미사용)의 1-pair	단발 FAIL, batch 필수	RUST-15 (x0.80 → batch x12.71)
B numpy 큰 array (1000 초과) vectorized	늘지 않음 (numpy 내부 BLAS)	(해당 kernel 아직 없음)
C numpy 작은 NxN (100 미만) API 다용	단발에서도 10-100x	RUST-16 (x66.70)
D numpy 중규모 BLAS 1 함수	경계선 위: 작은 크기 Rust 압승, 큰 크기에서 따라잡힘	RUST-17 (A=50 x9.55 → A=1000 x1.72)
E 차가운 데이터 경계 (dict / 문자열)	overhead 큼, batch 필수	—

상세표는 docs/perf_comparison/2026-05-22_kernel_implementation_comparison.md.

13.4 Cython 경로의 탈락 (build chain 부재)

scratch 비교에서 Cython kernel을 써서 3 way 비교를 시도했지만 Windows MSVC
build tools 부재 + mingw가 MSVC Python과 incompatible로 build 불가.
이것은 「수치 계산을 동등하게 쓸 수 있다」만으로는 언어 선택에 충분하지 않은
실례: build chain을 확립할 수 있는가가 필수 조건. source는
scratch/cython_collusion/에 저장해 Linux/WSL에서 재시도할 수 있는 형태로.

13.5 RUST-17b 추기 (2026-05-22 같은 날): rayon 병렬 + quickselect로 전 A 5x clear

RUST-17 baseline은 archive 큼 (A=200/1000)에서 gate FAIL이었지만, 같은 날 안에
RUST-17b로서 2 수단으로 재구현:

rayon par_iter로 N=64 집단 루프를 8-core 병렬화 + py.allow_threads로
GIL release
Vec::select_nth_unstable_by (Hoare quickselect, O(A) avg)로 top-k
partial sort — O(A log A) full sort를 치환

결과:

archive	RUST-17 (naive)	RUST-17b	개선율
A=50	x9.55	x12.83	+34%
A=200	x3.76 (FAIL)	x8.71 (PASS)	+132%
A=1000	x1.72 (FAIL)	x6.41 (PASS)	+273%
avg	x5.01	x9.32	+86%

판정표 (D)「numpy 중규모 batch」를 「경계선 위 → 병렬화로 만회 가능」으로
update. 「naive 이중 루프는 진다」뿐만 아니라 「rayon + algorithmic 개선으로
압승으로 전환된다」가 제시되었다.

std::simd는 nightly만이고 stable 불가 → 넣으면 추가로 2-3x. RUST-17c 후보.

13.6 다음에 오는 것 (2026-05-22 시점에서 계획 완료)

PyBind11 + C/C++ ctypes 경로의 3 kernel scratch 비교 (queue 투입 완료).
RUST-17c — std::simd (Rust nightly로 전환)로 SIMD 4-lane화.
월간 re-measure — env drift / numpy minor up / Rust nightly 등으로
결과가 움직이므로 주기 실행 (queue 투입 완료).
callers 전환 — PersonaOverlapPenalty.apply / NoveltyScorer.novelty_batch /
CoevolutionGovernance를 rust_ext 경로로 전환하는 PR.

14. References

Hillis, W. D. (1990). Coevolving parasites improve simulated evolution. Physica D.
Mouret, J.-B. & Clune, J. (2015). Illuminating search spaces by mapping elites. arXiv:1504.04909.
Lehman, J. & Stanley, K. (2008/2011). Novelty Search.
Stanley, K. & Miikkulainen, R. (2002). NEAT. Evolutionary Computation.
Deb, K. et al. (2002). NSGA-II. IEEE Trans Evol Comp.
Vinyals, O. et al. (2019). Grandmaster level in StarCraft II (AlphaStar). Nature.
완전한 목록은 v0.6.0a1 릴리스 시 references.bib에 동봉할 예정.

Series Navigation

You get articles that match your needs
You can efficiently read back useful information
You can use dark theme

What you can do with signing up