Multi‑turn negotiation remains a hard problem for large language model (LLM) agents. Existing benchmarks mostly measure agreement rates and ignore whether negotiated commitments generate value after execution. To bridge this gap, we introduce TradeWar, a multi‑country simulation benchmark. It translates dialogue into concrete tariff and trade‑share policies, enforces capacity and network constraints at settlement, and evaluates outcomes with an Economic Health Index (EHI). To tackle this execution‑grounded task we propose TACTIC (Trade‑aware Adaptive Consolidation and Theory‑of‑mind for International Commerce). TACTIC iteratively refines its negotiation strategy across episodes without updating the base LLM weights, leveraging prior experience, opponent behavior, and past settlement outcomes. In experiments, zero‑shot chain‑of‑thought (CoT) prompting (PromptLLM) achieves a 75% agreement rate but only a 57.1 final EHI (out of 100), whereas TACTIC reaches 68.8. Over 20 sequential episodes and three seeds, TACTIC’s EHI improves from 58.2 to 73.2; an ablation without cross‑episode updates scores 65.2. Under technology‑control and energy‑crisis shocks, TACTIC degrades less than PromptLLM and MemoryLLM. These findings indicate that CoT deliberation and interaction memory alone do not fully address delayed execution feedback, while settlement‑grounded adaptation markedly enhances long‑horizon negotiation.
Review