Anchoring-Recovery Concession Framework: Enabling LLM as a Natural Language Translation Layer for Freight Negotiations

The research team proposes a dual-index Anchoring-Recovery framework targeting the challenges of freight negotiations in dynamic pricing environments. It ensures quote monotonicity via a spread-adaptive concession parameter β and an anchoring mechanism, uses LLM only as a natural language translation layer, and controls pricing decisions with deterministic formulas to enable large-scale concurrent negotiations with zero inference cost.

The freight brokerage industry conducts massive price negotiations daily. Dynamic pricing environments (fuel fluctuations, capacity changes, etc.) lead to constant adjustments of pricing objectives. Traditional frameworks assume fixed pricing objectives, which are disconnected from reality. Key challenges include: 1. The fixed β of classic time-dependent concession frameworks cannot adapt to real-time pricing objectives; 2. LLM negotiation agents have high costs, non-determinism, and prompt injection risks.

The framework includes two core mechanisms: 1. Spread-adaptive β parameter: Derive β from real-time spreads—narrow spreads lead to quick concessions to ensure deals, medium spreads balance robustness, and wide spreads aggressively pursue profits; 2. Anchoring-Recovery mechanism: When pricing objectives are updated, record the current lowest quote as an anchor, and the new concession curve recovers from this anchor to ensure quotes are non-decreasing.

The framework redefines the role of LLM: it does not participate in pricing decisions, only acts as a translation layer—converting system quotes and strategic intentions into natural language, and parsing carrier responses to extract key information. Advantages of the decoupled design: controllable cost (near-zero inference cost), guaranteed determinism (no LLM hallucinations), improved security (prompt injection does not affect pricing), and horizontal scalability to thousands of concurrent negotiations.

Validated based on 115,125 real negotiation data points: 1. Adaptive β increases deal closure rate for narrow spreads, and matches the best fixed β baseline for medium/wide spreads; 2. Compared to a 20-billion parameter LLM agent, it achieves similar deal closure rates and cost savings but with zero inference cost; 3. When facing LLM-driven opponents, it maintains cost savings levels and achieves higher deal closure rates.

Engineering value: Interpretable (decisions have mathematical basis), debuggable (precisely locate issues), maintainable (LLM decoupled from core logic), and compliance-friendly (meets regulatory requirements). Methodological insights: LLM is not a panacea—traditional algorithms enhanced with LLM are better; separation of duties (separate deterministic decisions from language interaction); need to balance cost and benefit.

Application scenario expansion: Dynamic pricing scenarios such as online advertising bidding, supply chain finance, energy trading, and procurement management. Limitations: Focuses on one-to-one negotiations, insufficient reflection of emotional factors and long-term relationships, simplified opponent modeling. Future directions: Expansion to multi-party games, reinforcement learning optimization of strategy parameters, and refined opponent modeling.