[Paper] Anchor-and-Resume Concession Under Dynamic Pricing for LLM-Augmented Freight Negotiation

Source: arXiv - 2604.20732v1

Overview

This paper tackles a real‑world pain point for freight broker platforms: negotiating carrier rates in a fast‑moving, dynamically priced market. Traditional concession models use a static “β” parameter that can’t keep up with mid‑conversation price shifts, while large language models (LLMs) add flexibility but at the cost of latency, nondeterminism, and security risks. The authors introduce a two‑index “anchor‑and‑resume” framework that adapts the concession curve to live market spreads and guarantees that offers never move backward, all while keeping the pricing logic deterministic and off‑loading only natural‑language translation to an LLM.

Key Contributions

• Spread‑derived adaptive β: Derives the concession shape parameter directly from the current price spread, automatically tuning aggressiveness per negotiation regime.
• Anchor‑and‑resume mechanism: Guarantees monotonic (non‑decreasing) offers even when the target price is revised mid‑dialogue, eliminating the “price retraction” bug of prior formulas.
• Deterministic pricing core: All price calculations are performed by a closed‑form formula; the LLM is used solely for translating between human‑readable language and the formula inputs/outputs.
• Scalable evaluation: Tested on 115 k real negotiations, showing comparable or better broker savings and agreement rates versus both fixed‑β baselines and a 20‑billion‑parameter LLM broker.
• Robustness to stochastic carriers: Maintains higher agreement rates and savings when negotiating against LLM‑powered carrier agents that introduce randomness.

Methodology

1. Margin‑Spread Mapping – For each load, the system computes the spread (difference between the carrier’s cost and the broker’s target). This spread feeds a simple function that outputs a β value tailored to the current market tightness (narrow, medium, wide).
2. Concession Formula – Using the adaptive β, the broker’s offer follows a classic time‑dependent concession curve (e.g., exponential decay) that determines how quickly the price moves toward the target.
3. Anchor‑and‑Resume Logic
   • Anchor: When a pricing shift occurs (e.g., the target is revised upward), the current offer is “anchored” as the new baseline.
   • Resume: Future concessions are computed relative to this anchor, ensuring the next offer is never lower than the anchored price.
4. LLM Integration (Optional) – An LLM receives the deterministic offer and produces a natural‑language message for the carrier (e.g., “We can meet you at $X per mile”). The reverse translation (carrier’s textual reply → numeric intent) is also handled by the LLM, but the core price never passes through the model.
5. Experimental Setup – The authors replayed 115 125 historic negotiations, comparing:
   • Fixed‑β baselines (static concession curves)
   • The proposed adaptive‑β + anchor‑and‑resume system
   • A full‑scale LLM broker (20B parameters)
   • Simulated stochastic carrier agents powered by LLMs

Results & Findings

• Regime‑aware behavior: In tight spreads the adaptive β pushes the broker to concede quickly, improving load coverage. In looser spreads it holds back, preserving margin.
• Monotonicity guarantee: No negotiation ever experienced a price “step back,” eliminating a common source of carrier frustration.
• Cost efficiency: By keeping the heavy lifting in a deterministic formula, the system can handle thousands of concurrent chats with negligible cloud‑compute expense.

Practical Implications

• For freight platforms: Deploy a lightweight pricing engine that reacts instantly to market volatility without the overhead of massive LLM inference.
• For developers building negotiation bots: Separate the decision logic (deterministic math) from the communication layer (LLM‑generated text) to achieve both transparency and conversational polish.
• Cost savings: Reduce inference spend dramatically—moving from multi‑hundred‑millisecond, GPU‑heavy LLM calls to sub‑millisecond CPU calculations.
• Compliance & auditability: Deterministic formulas provide a clear audit trail for pricing decisions, easing regulatory reporting and internal governance.
• Extensibility: The anchor‑and‑resume pattern can be applied to any dynamic‑pricing scenario (e.g., ride‑hailing surge pricing, ad‑bid negotiations) where target values shift mid‑interaction.

Limitations & Future Work

• Dependence on accurate spread estimation: The adaptive β assumes the spread is correctly measured; noisy cost data could mis‑tune concession aggressiveness.
• LLM translation quality: While the pricing core is deterministic, the natural‑language layer still inherits LLM quirks (e.g., occasional phrasing errors or prompt‑injection vulnerabilities).
• Real‑time market shocks: Sudden, extreme price spikes may require additional safeguards beyond the anchor‑and‑resume logic (e.g., hard caps).
• Future directions:
  • Incorporate reinforcement‑learning to fine‑tune β updates from live negotiation outcomes.
  • Explore hybrid models where a lightweight, domain‑specific LLM assists in edge‑case reasoning while the core remains formulaic.
  • Extend the framework to multi‑party negotiations (e.g., broker‑carrier‑shipper triads) and to other logistics verticals.

Authors

• Hoang Nguyen
• Lu Wang
• Marta Gaia Bras

Paper Information

• arXiv ID: 2604.20732v1
• Categories: cs.MA, cs.AI, cs.CL
• Published: April 22, 2026
• PDF: Download PDF