TradeWar: A Benchmark for Execution-Grounded Trade Negotiation with Adaptive LLM Agents
Organizations: College of Computer Science, Beijing Information Science and Technology University, Beijing, China
Abstract
Multi-turn negotiation remains a challenging problem for large language model (LLM) agents. While LLMs can conduct natural-language dialogue, existing benchmarks predominantly assess only agreement rates, overlooking whether negotiated commitments actually create value once executed. To address this gap, we introduce TradeWar, a multi-country simulation benchmark that translates negotiation dialogues into concrete tariff and trade-share policies, enforces capacity and network constraints during settlement, and evaluates outcomes via an Economic Health Index (EHI). To solve this execution-grounded task, we propose TACTIC (Trade-aware Adaptive Consolidation and Theory-of-mind for International Commerce), an agent that iteratively refines its negotiation strategy across episodes-without updating base LLM weights-by leveraging prior experience, opponent behavior, and past settlement outcomes. Despite using zero-shot chain-of-thought (CoT) prompting, PromptLLM attains a 75% agreement rate but only 57.1 final EHI (*100), whereas TACTIC reaches 68.8. Across 20 sequential episodes and three seeds, TACTIC's final EHI improves from 58.2 to 73.2, compared with 65.2 for an ablation without cross-episode updates. Under technology-controls and energy-crisis shocks, TACTIC also exhibits smaller performance degradation than PromptLLM and MemoryLLM. These results indicate that CoT deliberation and interaction memory alone do not fully address delayed execution feedback, whereas settlement-grounded adaptation improves long-horizon negotiation in TradeWar.
Figures & tables
| Benchmark family | Multi-party | Executable | Delayed post-exec. feedback | Primary endpoint |
| Bilateral bargaining ( Lewis et al., 2017 ) | — | — | — | Deal / immediate utility |
| Multi-issue negotiation ( Chan et al., 2024 ; Kwon et al., 2025 ; Abdelnabi et al., 2024 ) | — | — | — | Agreement quality |
| Economic games / agent markets ( Shapira et al., 2024 ; Liu et al., 2026b ; Sugiura et al., 2026 ; Feng et al., 2026 ; Bansal et al., 2025 ) | Payoff / profit / welfare | |||
| Diplomacy-style ( Bakhtin et al., 2022 ; Guan et al., 2024 ; Xu et al., 2025 ; Duffy et al., 2026 ) | ✓ | — | — | Match / board outcome |
| TradeWar (ours) | ✓ | ✓ | ✓ | Post-execution EHI |
| Agent | Agree. (%) | EHI | GDP (%) | Target (%) |
| Random | 15 | 38.4 | -0.6 | 9.2 |
| RuleBased | 35 | 46.9 | 0.5 | 28.4 |
| PromptLLM | 75 | 57.1 | 1.3 | 43.6 |
| MemoryLLM | 80 | 60.2 | 2.1 | 54.8 |
| TACTIC | 85 | 68.8 | 4.0 | 71.3 |
| Configuration | EHI | GDP (%) | Target (%) |
| No adaptive modules | 51.6 | 0.8 | 32.1 |
| + Experience Memory | 55.3 | 1.4 | 38.7 |
| + Opponent Modeling | 59.0 | 1.9 | 47.4 |
| TACTIC w/o Reflection (memory + opponent + planning) | 63.9 | 2.7 | 56.2 |
| Full TACTIC (+ structured reflection) | 68.8 | 4.0 | 71.3 |
Appendix figures & tables4 assets
Supplementary material from the paper’s appendix.
Appendix
| Component | Specification |
| Agents | Six country agents with heterogeneous private constraints, public indicators, and product-level trade targets. |
| Temporal unit | Quarterly transitions over multi-year episodes; formal negotiation opens once every four quarters. |
| Interaction protocol | Bilateral invitations, agenda exchange, alternating-turn bargaining, and a separate signing stage; bargaining is capped at 50 rounds. |
| Action categories | Natural-language negotiation messages, unilateral policy actions, and executable bilateral tariff/share agreements. |
| Agreement schema | Parties, product, importer/exporter direction, term type, magnitude, and duration. |
| Validation | Party consistency, edge existence, supported term type, directional consistency, and graph/share feasibility. |
Automated Negotiation
Automated negotiation studies computational agents that bargain over prices, preferences, and agreements in human, market, and multi-agent settings.
See all 49 papers