TradeWar

From dialogue to executed trade

Six synthetic countries negotiate tariff and trade-share terms. A five-year episode contains 20 quarterly transitions and opens negotiation every four quarters. Proposals must pass party, product, direction and feasibility checks before signing changes the state; limited export capacity can still prevent promised delivery.

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

EHI measures terminal outcomes

The primary score is final-quarter Economic Health Index, displayed on the paper’s 0–100 point scale. It weights normalized GDP growth and trade balance at 0.20 each, and unemployment, inflation, export completion and import completion at 0.15 each. Agreement rate, GDP growth and target completion are separate diagnostics, not substitutes for EHI.

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

Matched agents and opponents

The baseline study runs 100 games with three-versus-three assignments and five episodes per ordered matchup. All LLM agents use GPT-4o-mini, temperature 0.3, max_tokens 1000 and a shared structured CoT interface. Random and RuleBased use no LLM. Rows retain the opponent pool or named opponent instead of treating framework differences as new base-model capabilities.

TradeWar paper v1 — protocols and publisher results

Keep sequential learning separate

Against fixed PromptLLM, each method runs 20 sequential episodes over three seeds. TACTIC persists settlement-driven guidance; TACTIC-static and MemoryLLM reset between episodes. The paper reports early and late three-episode means. Separate stress tests use 12 paired-mirror episodes per condition, so their baselines differ from the main cross-play table.

TradeWar paper v1 — protocols and publisher results

Interpret the simulator cautiously

EHI is a fixed benchmark aggregate, not welfare, a forecast or real-world policy advice. Country roles, score weights and delayed feedback favor particular strategies; the environment and TACTIC were introduced together. One LLM backbone and modest episode budgets do not establish a universal agent ranking.

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

What the public artifact actually includes

The author repository is reachable, but its Artifact Scope explicitly excludes TACTIC, LLM baselines, full experiment runners and generated result logs. It describes support for benchmark mechanics and non-LLM reference policies, not reproduction of all paper experiments. We preserve the paper’s numbers as publisher reports and link this boundary directly.

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

Main cross-play results

Table 2 aggregates the same 100-game opponent pool. Agreement, GDP and target metrics are retained alongside the primary final EHI.

No-shock baseline, 3v3 countries, five years/20 quarters. All LLM agents share GPT-4o-mini, temperature 0.3 and max_tokens 1000; Random and RuleBased use no LLM.

TradeWar paper v1 — protocols and publisher results

Cumulative module ablation

Table 3 adds modules cumulatively under the baseline cross-play protocol. The final Full TACTIC row repeats the main table’s TACTIC result and is not duplicated in primary observations.

Effects depend on module addition order. The zero-module row retains the shared CoT/action-stage scaffold; it is not a renamed PromptLLM baseline.

TradeWar paper v1 — protocols and publisher results

Opponent-conditioned cross-play

Figure 4 gives twenty off-diagonal pairings. The focal agent and opponent each control three countries; these conditional results are not overall scores.

Five episodes per ordered matchup. Exact numbers were read from the author heatmap; same-agent matchups are excluded.

TradeWar paper v1 — protocols and publisher results

Sequential-study endpoints

Section 5.2 explicitly reports these early/late three-episode means. The opponent and assignment remain fixed; these values do not share the cross-play opponent pool.

Three independent seeds, 20 episodes per method per seed, fixed PromptLLM opponent. TACTIC persists updates; TACTIC-static and MemoryLLM reset between episodes. Unlabelled curve coordinates are not estimated.

TradeWar paper v1 — protocols and publisher results

Relative drop under shocks

Section 5.2 reports relative EHI declines using each agent’s own matched no-shock stress-test baseline. A smaller percentage drop means less degradation, not a higher absolute EHI.

Twelve paired-mirror episodes per condition (six seeds and reversed assignments). Relative drops are a different metric; no absolute EHI is reconstructed from these percentages.

TradeWar paper v1 — protocols and publisher results