AI EVALUATION, IN CONTEXT.BY EQUALITY MACHINE
Benchmark Radar.Benchmarks. Model results. Original sources.
/

Model reports

TradeWar

TradeWar检验智能体能否把谈判达成的贸易协议转化为可执行条款，并在产能受限的结算后取得实际收益。

原论文官方仓库收录来源

AI Benchmark Radar / Equality Machine · 核对于 2026-11-02

对应2026年10月30日论文v1的作者报告期末EHI。交叉对战汇总、特定对手、消融和连续实验窗口采用不同协议。

从对话到实际贸易

六个合成国家就关税和贸易份额谈判。每轮模拟持续五年、含20次季度状态转移，每四个季度开放谈判。提案需通过参与方、商品、方向与可行性检查，签署后才改变状态；即便协议有效，出口产能不足仍可能导致交付落空。

TradeWar paper v1 — protocols and publisher results · TradeWar official repository

EHI衡量期末结果

主指标为最后一季度的经济健康指数，保留论文0–100分尺度。标准化GDP增长和贸易平衡各占0.20，失业、通胀、出口目标完成度和进口目标完成度各占0.15。协议达成率、GDP增长和目标完成率是另外的诊断指标，不能替代EHI。

TradeWar paper v1 — protocols and publisher results · TradeWar official EHI definition

代理与对手的匹配方式

基线研究共100局，采用三对三国家分配，每个有序对手组合运行五轮。全部LLM代理使用GPT-4o-mini、温度0.3、max_tokens 1000及统一的结构化CoT接口。Random和RuleBased不使用LLM。成绩保留对手池或具体对手，不把框架差异当作基础模型能力更新。

TradeWar paper v1 — protocols and publisher results

单独理解连续适应实验

面对固定PromptLLM对手，各方法在三个种子下连续运行20轮。TACTIC保留由结算结果更新的策略；TACTIC-static和MemoryLLM在轮间重置。论文报告最初和最后三个轮次的均值。冲击测试每种条件使用12轮镜像配对实验，因此基线不同于主要交叉对战表。

TradeWar paper v1 — protocols and publisher results

模拟器的适用边界

EHI是固定权重的评测汇总分，不是福利指标、预测或现实政策建议。国家角色、权重与延迟反馈会偏向特定策略，且环境和TACTIC同期提出。单一基础模型和有限实验轮数不足以建立通用代理排名。

TradeWar paper v1 — protocols and publisher results · TradeWar official EHI definition

公开材料实际覆盖范围

作者仓库可访问，但Artifact Scope明确排除了TACTIC、LLM基线、完整实验运行器及生成结果日志。其声明覆盖评测机制与非LLM参考策略，并不能复现论文全部实验。本页将数值保留为作者报告，同时直接链接这一公开范围说明。

TradeWar official artifact inclusion and exclusion scope · TradeWar paper v1 — protocols and publisher results

主要交叉对战结果

表2汇总同一100局对手池。除主指标期末EHI外，保留协议、GDP及目标完成诊断值。

主要交叉对战结果
代理达成协议%期末EHI×100GDP %目标完成%
Random1538.4-0.69.2
RuleBased3546.90.528.4
PromptLLM7557.11.343.6
MemoryLLM8060.22.154.8
TACTIC8568.84.071.3

无冲击基线，3v3国家、五年20季度。LLM代理统一使用GPT-4o-mini、温度0.3和max_tokens 1000；Random和RuleBased不使用LLM。

TradeWar paper v1 — protocols and publisher results

累计模块消融

表3在基线交叉对战协议下逐步加入模块。最后一行Full TACTIC与主表TACTIC重复，主成绩中不重复计数。

累计模块消融
配置期末EHI×100GDP %目标完成%
No adaptive modules51.60.832.1
+ Experience Memory55.31.438.7
+ Opponent Modeling59.01.947.4
TACTIC w/o Reflection (memory + opponent + planning)63.92.756.2
Full TACTIC (+ structured reflection)68.84.071.3

结果依赖模块加入顺序。零模块行仍保留统一CoT及动作阶段框架，并非把PromptLLM改名。

TradeWar paper v1 — protocols and publisher results

按对手划分的交叉对战

图4给出20个非对角线组合。被测代理和对手各控制三个国家；这些对手条件化结果不是整体成绩。

按对手划分的交叉对战
被测代理对手期末EHI×100（样本标准差）
RandomRuleBased47.2 (sample SD 4.1)
RandomPromptLLM41.5 (sample SD 2.8)
RandomMemoryLLM34.8 (sample SD 3.6)
RandomTACTIC30.1 (sample SD 2.3)
RuleBasedRandom54.9 (sample SD 3.2)
RuleBasedPromptLLM47.3 (sample SD 4.8)
RuleBasedMemoryLLM47.8 (sample SD 2.6)
RuleBasedTACTIC37.6 (sample SD 3.9)
PromptLLMRandom72.6 (sample SD 5.1)
PromptLLMRuleBased52.4 (sample SD 4.4)
PromptLLMMemoryLLM58.0 (sample SD 3.3)
PromptLLMTACTIC45.4 (sample SD 4.2)
MemoryLLMRandom62.8 (sample SD 2.9)
MemoryLLMRuleBased65.4 (sample SD 4.6)
MemoryLLMPromptLLM58.3 (sample SD 3.8)
MemoryLLMTACTIC54.3 (sample SD 2.4)
TACTICRandom75.9 (sample SD 3.1)
TACTICRuleBased63.1 (sample SD 4.0)
TACTICPromptLLM71.2 (sample SD 2.7)
TACTICMemoryLLM65.0 (sample SD 3.5)

每个有序组合五轮。数值直接读取作者热力图；不包含同一代理对战自身。

TradeWar paper v1 — protocols and publisher results

连续实验的阶段端点

第5.2节明确报告以下最初/最后三个轮次均值。对手和国家分配保持固定，不与交叉对战的对手池混用。

连续实验的阶段端点
代理轮次窗口平均期末EHI×100
TACTIC-static1-358.4
TACTIC1-358.2
MemoryLLM18-2062.5
TACTIC-static18-2065.2
TACTIC18-2073.2

三个独立种子，每方法每种子20轮，对手固定PromptLLM。TACTIC保留更新，TACTIC-static和MemoryLLM轮间重置；不估读曲线上的未标注坐标。

TradeWar paper v1 — protocols and publisher results

冲击下的相对降幅

第5.2节使用各代理在压力测试中自身匹配的无冲击基线，报告EHI相对下降比例。降幅较小代表退化较少，并不直接等于绝对EHI更高。

冲击下的相对降幅
代理技术限制：下降%能源危机：下降%
TACTIC16.725.3
MemoryLLM29.638.1
PromptLLM32.242.3

每种条件12轮镜像配对（六个种子与反向国家分配）。相对降幅是另一指标，不从这些百分数反推出绝对EHI。

TradeWar paper v1 — protocols and publisher results

原始发布

2026年11月2日核对的论文v1及作者直接链接的发布材料，成绩继续归属于评测发布者。

核对于 2026-11-02

已收录的评测资料

这条来源记录收录了 9 条数值成绩。来源区分了 9 个模型身份或配置。同一模型可能有多条记录；比较前请核对具体版本、工具和运行预算。 查看成绩及逐条证据 →

报告指标
Mean final Economic Health Index (EHI ×100)
原始单位
points
指标方向
数值越高越好

从这些资料开始

引用文档及日期（1）

评测信息与证据

来源
Model reports
发布者
未收录
评测发布日期
2026-10-30
采集日期
未收录
模态
未收录

本站保留来源报告的结果，不代表独立复测。请先核对任务版本、评测协议、分数单位和报告日期；不同运行条件不能直接合并为统一排名。

报告成绩 9

按来源中的观察顺序展示，顺序不代表排名。日期保留原始精度。

模型数值单位协议报告日期与口径来源
MemoryLLM60.2pointsPaper v1 Table 2; no-shock baseline cross-play, six countries in 3v3 assignments, five years/20 quarters, five episodes per ordered matchup against each other agent; 100 games total; balanced reversed country assignments; aggregate final EHI ×100. LLM agents use GPT-4o-mini, shared schema-constrained zero-shot CoT, temperature 0.3, max_tokens 1000; immutable API revision unreported. Random and RuleBased are non-LLM reference policies.2026-10-30 · document_publicationTradeWar paper v1 · mean final EHI (Tables 2–3)
PromptLLM57.1pointsPaper v1 Table 2; no-shock baseline cross-play, six countries in 3v3 assignments, five years/20 quarters, five episodes per ordered matchup against each other agent; 100 games total; balanced reversed country assignments; aggregate final EHI ×100. LLM agents use GPT-4o-mini, shared schema-constrained zero-shot CoT, temperature 0.3, max_tokens 1000; immutable API revision unreported. Random and RuleBased are non-LLM reference policies.2026-10-30 · document_publicationTradeWar paper v1 · mean final EHI (Tables 2–3)
Random38.4pointsPaper v1 Table 2; no-shock baseline cross-play, six countries in 3v3 assignments, five years/20 quarters, five episodes per ordered matchup against each other agent; 100 games total; balanced reversed country assignments; aggregate final EHI ×100. LLM agents use GPT-4o-mini, shared schema-constrained zero-shot CoT, temperature 0.3, max_tokens 1000; immutable API revision unreported. Random and RuleBased are non-LLM reference policies.2026-10-30 · document_publicationTradeWar paper v1 · mean final EHI (Tables 2–3)
RuleBased46.9pointsPaper v1 Table 2; no-shock baseline cross-play, six countries in 3v3 assignments, five years/20 quarters, five episodes per ordered matchup against each other agent; 100 games total; balanced reversed country assignments; aggregate final EHI ×100. LLM agents use GPT-4o-mini, shared schema-constrained zero-shot CoT, temperature 0.3, max_tokens 1000; immutable API revision unreported. Random and RuleBased are non-LLM reference policies.2026-10-30 · document_publicationTradeWar paper v1 · mean final EHI (Tables 2–3)
TACTIC68.8pointsPaper v1 Table 2; no-shock baseline cross-play, six countries in 3v3 assignments, five years/20 quarters, five episodes per ordered matchup against each other agent; 100 games total; balanced reversed country assignments; aggregate final EHI ×100. LLM agents use GPT-4o-mini, shared schema-constrained zero-shot CoT, temperature 0.3, max_tokens 1000; immutable API revision unreported. Random and RuleBased are non-LLM reference policies.2026-10-30 · document_publicationTradeWar paper v1 · mean final EHI (Tables 2–3)
TACTIC — + Experience Memory55.3pointsPaper v1 Table 3; cumulative TACTIC module ablation under the baseline cross-play opponent pool; six countries, 3v3, five years/20 quarters. LLM agents use GPT-4o-mini, shared schema-constrained zero-shot CoT, temperature 0.3, max_tokens 1000; immutable API revision unreported. Random and RuleBased are non-LLM reference policies.2026-10-30 · document_publicationTradeWar paper v1 · mean final EHI (Tables 2–3)
TACTIC — + Opponent Modeling59.0pointsPaper v1 Table 3; cumulative TACTIC module ablation under the baseline cross-play opponent pool; six countries, 3v3, five years/20 quarters. LLM agents use GPT-4o-mini, shared schema-constrained zero-shot CoT, temperature 0.3, max_tokens 1000; immutable API revision unreported. Random and RuleBased are non-LLM reference policies.2026-10-30 · document_publicationTradeWar paper v1 · mean final EHI (Tables 2–3)
TACTIC — No adaptive modules51.6pointsPaper v1 Table 3; cumulative TACTIC module ablation under the baseline cross-play opponent pool; six countries, 3v3, five years/20 quarters. LLM agents use GPT-4o-mini, shared schema-constrained zero-shot CoT, temperature 0.3, max_tokens 1000; immutable API revision unreported. Random and RuleBased are non-LLM reference policies.2026-10-30 · document_publicationTradeWar paper v1 · mean final EHI (Tables 2–3)
TACTIC — TACTIC w/o Reflection (memory + opponent + planning)63.9pointsPaper v1 Table 3; cumulative TACTIC module ablation under the baseline cross-play opponent pool; six countries, 3v3, five years/20 quarters. LLM agents use GPT-4o-mini, shared schema-constrained zero-shot CoT, temperature 0.3, max_tokens 1000; immutable API revision unreported. Random and RuleBased are non-LLM reference policies.2026-10-30 · document_publicationTradeWar paper v1 · mean final EHI (Tables 2–3)

相关评测探索

按共享资料或来源分类推荐，不代表同一评测或经审核的身份关联。

SpreadsheetBench 2来源分类相同 →Kimi Code Bench v2来源分类相同 →PostTrainBench Lite来源分类相同 →CC-Bench-V2 Repo Exploration来源分类相同 →Blueprint-Bench 2来源分类相同 →Management Consulting Tasks (Internal)来源分类相同 →SWE-MM来源分类相同 →CI Memories Coverage来源分类相同 →

继续阅读

工具与智能体评测金融评测

继续探索

按这条记录的来源标签检索目录；搜索匹配不代表相同评测。

agents →Model reports · 全部记录 →