Evolve-AGI worksheet

先把它当证据成熟度工作表，而不是领域指数。

Evolve-AGI Index 不是 AGI 能力分，而是一个 exploratory evidence-maturity worksheet，用来检查 AI Agent 自进化证据是否足够成熟。

这个数字只回答“证据是否足够进入下一轮复核”：benchmark 是否可复跑、闭环是否保留、迁移和治理是否说明。它不回答哪个系统更好，也不是 AGI 能力分。

系统证据队列 Benchmark 追踪研究优先级队列 GitHub 证据链方法说明

Worksheet formula

EAI = Σ(signal_score × signal_weight)

方法说明

Bench

Benchmark 表现

80.1/100

主指数只纳入 agent/code/web/app/algorithm-discovery 等自进化相关 benchmark family；AI-for-science adjacent evidence 单独展示，不进入主分。

权重 18% paper-drafts/appendix.tex + site/src/data/survey.ts + research/ranking-framework/radar-profiles.json

Loop

核心闭环强度

80.2/100

Top 自进化系统是否真的有可变对象、反馈信号、选择机制和保留机制。

权重 20% site/src/data/rankings.ts + research/ranking-framework/README.md

Evidence

证据链可信度

78.7/100

综合 D2 证据强度与项目报告覆盖率，防止只有口号没有 raw / report。

权重 18% analysis/github-project-data-analysis.json + projects/INDEX.md

Transfer

迁移与验证

65.2/100

改进是否能跨任务、跨环境或跨时间切片迁移，并有验证门约束。

权重 14% paper-drafts/ch5-evaluation.tex + site/src/data/rankings.ts

Access

可运行与可复用

58.8/100

系统是否有开源实现、文档、成本效率和实际采用价值。

权重 12% site/src/data/rankings.ts + site/src/data/starAnalysis.ts

Momentum

领域动量

68.9/100

结合证据分诊队列与 Star 活跃度，提示近期研究和工程关注度，不能单独当作价值排名。

权重 10% site/src/data/analysis.json + site/src/data/starAnalysis.ts

Governance

治理成熟度

65.1/100

看安全、成本、时间戳置信度和可审计边界是否跟上能力增长。

权重 8% site/src/data/rankings.ts + output/raw-github-timestamp-index.md

Reconstructed snapshots

这些点是回填快照，不是领域真实增长曲线。

Date	Worksheet score	Benchmark	Strict / broad repos	Reports	Source
2026-05-26 Archived baseline · backfilled	67.4	72.2	82 / 186	90	work/research/archived-analysis/github-project-data-analysis.json Backfilled with current formula over archived corpus counts.
2026-05-29 Site snapshot · backfilled	71.3	76.4	90 / 195	187	site/src/data/analysis.json Previous site snapshot before benchmark became a weighted signal.
2026-05-30 Benchmark-weighted	72.9	80.1	90 / 195	193	reports/evolve-agi-index-trend.json First benchmark-weighted snapshot before the public site data snapshot was synced.
2026-06-01 Corpus-synced	72.9	80.1	93 / 200	239	analysis/github-project-data-analysis.json + site/src/data/analysis.json Public site data snapshot synced to the latest GitHub analysis corpus.
2026-06-02 MPA evidence added	72.9	80.1	93 / 200	239	https://www.deepprinciple.com/papers/mpa.pdf + https://mp.weixin.qq.com/s/Do3sauQ8oSoRluaCptYe-g Added MPA / MIRA materials-property benchmark evidence; corpus counts unchanged.

Main-index benchmark evidence

主指数纳入的 benchmark evidence

这里仅展示进入主 worksheet 分数的 agent/code/web/app/algorithm-discovery benchmark family。MPA、MIRA 等 AI-for-science adjacent evidence 已在数据层分离，不进入主分。

Reflexion

HumanEval pass@1

91

Before80.0% GPT-4 baseline

After91.0%

Gain+11.0pp

Code-generation pass rate with language reflection memory.

paper-drafts/appendix.tex; site/src/data/survey.ts

Reflexion

ALFWorld success

97

Before77%

After97%

Gain+20pp

Embodied/web-style task success, used as cross-domain evidence.

site/src/data/survey.ts; research/ranking-framework/radar-profiles.json

SelfEvolve

HumanEval pass@1

86

Before74.39% ChatGPT baseline

After85.98%

Gain+11.59pp

Self-improvement loop on coding benchmark.

paper-drafts/appendix.tex

DGM

SWE-bench Verified

86

Before20.0%

After50.0%

Gain+30.0pp

Hard software-engineering benchmark; gain is weighted heavily.

paper-drafts/appendix.tex; research/ranking-framework/radar-profiles.json

SICA

SWE-bench Verified

88

Before17.0%

After53.0%

Gain+36.0pp

Self-modifying coding agent result in software repair.

paper-drafts/appendix.tex

DGM

Polyglot Coding

72

Before14.2%

After30.7%

Gain+16.5pp

Cross-language transfer evidence.

site/src/data/survey.ts; paper-drafts/appendix.tex

SAGE

LiveCodeBench

71

Beforebackbone model

After+8.9%

Gain+8.9%

Turn-based self-evolution reasoning improvement.

paper-drafts/appendix.tex

WebRL

WebArena-Lite

82

Before4.8%

After42.4%

Gain+37.6pp

Web-agent benchmark with large absolute improvement.

paper-drafts/appendix.tex

AgentEvolver

AppWorld

67

Before1.8%

After23.2%

Gain+21.4pp

App-agent environment improvement; still low absolute final score.

paper-drafts/appendix.tex

AlphaEvolve

Algorithm / infra discovery

89

Beforeprior best systems

After48 multiplications; Borg +0.7%; FlashAttention +23%

Gainnew record / infra gains

Non-pass-rate benchmark family normalized by impact and verification.

paper-drafts/appendix.tex; research/ranking-framework/radar-profiles.json

Evidence Snapshot

首页指数背后的语料规模

10 Benchmark evidence 80.1 weighted benchmark signal

93 Strict self-evolution 93 / 646 classified repos

200 Broad related field memory / skill / evaluation / harness included

239 Public reports 100% of analyzed projects

Evidence examples

闭环证据较强的代表系统，不是排名

Worksheet row 1

AlphaEvolve

56yr matrix multiplication record; Borg 0.7% compute recovery; 75% SOTA recovery on 50+ math problems

Tier A L4 worksheet score 7

Worksheet row 2

Darwin Godel Machine (DGM)

SWE-bench 20%→50%; Polyglot 14.2%→30.7%; archive-based agent evolution with lineage

Tier A L4 worksheet score 7

Worksheet row 3

FunSearch

New cap-set constructions; improved bin-packing heuristics; Nature publication

Tier A L4 worksheet score 7

Worksheet row 4

ADAS

Discovers agent architectures automatically; cross-domain transfer of discovered designs

Tier A L4 worksheet score 7

Worksheet row 5

OpenEvolve

Open-source AlphaEvolve implementation; evolutionary program search; accessible to researchers

Tier B+ L4 worksheet score 6

Research clues

近期研究线索，不是价值结论

Research clue 1

AgentEvolver

进化/搜索循环 → 评估器/打分器 → 智能体编排 → 训练/数据循环

Agent 进化框架 triage signal 83

Research clue 2

DARWIN

进化/搜索循环 → 反思记忆

安全策略进化 triage signal 77

Research clue 3

LLM-Self-Judge

进化/搜索循环 → 评估器/打分器 → 智能体编排 → 训练/数据循环

自评判训练 triage signal 76

Research clue 4

openevolve

进化/搜索循环 → 评估器/打分器

进化式代码优化 triage signal 71

Research clue 5

SCOPE

进化/搜索循环

上下文/Prompt 进化 triage signal 68