A negotiation method for fine-tuning a large language model with LoRA

Abstract

The present application relates to the technical field of natural language processing, and particularly relates to a negotiation method of fusing LoRA fine-tuning large language model, aiming at the pain point of e-commerce bargaining, specifically realizing the structured perception of 6 intents and 8 types of emotions through LoRA fine-tuning large model, converting discrete emotions into continuous benchmark concession rate by combining fuzzy reasoning, designing a deterministic pricing strategy with triple rigidity constraint, matching a random disturbance anti-probe mechanism triggered by low price, and generating consistent natural language reply with numerical hard constraint and emotional soft constraint. The present application solves the problems of unstable perception, abnormal pricing and weak anti-probe of the existing method, improves the robustness and interaction experience of the intelligent negotiation system, is lightweight and easy to adapt, provides a feasible solution for the landing of e-commerce intelligent bargaining, and promotes the fusion application of related technologies in e-commerce scenarios.

Description

Technical Field

[0001] This invention relates to the field of natural language processing technology, specifically to a negotiation method that integrates LoRA fine-tuning of a large language model. Background Technology

[0002] As the scale of online transactions expands, the frequency and complexity of negotiation interactions increase significantly. In existing technologies, negotiation systems typically develop along two main lines.

[0003] Firstly, negotiation methods based on rules or game theory generate offers and responses through preset utility functions, concession strategies, or rule templates. While these methods offer a degree of controllability, their ability to understand user expressions in open-domain dialogues is limited, especially in cases of emotional expressions, implicit intentions, and cross-round strategy evolution. This can easily lead to insufficient understanding of the dialogue and rigid strategy responses.

[0004] Secondly, while negotiation methods based on neural networks or large language models can achieve good results in natural language understanding and generation, they are prone to numerical instability, price overreach, and strategy drift when it comes to numerical pricing decisions. Furthermore, they lack stable and effective defensive constraints when facing malicious price suppression or repeated probing, resulting in insufficient protection of system revenue. Some methods attempt to introduce emotional factors to adjust dialogue style, but emotional expression often lacks a unified and controllable mapping with numerical strategies, making it difficult to achieve emotionally consistent strategy adjustments while ensuring negotiation bottom lines and stable convergence. Summary of the Invention

[0005] The purpose of this invention is to provide a negotiation method that integrates LoRA fine-tuning of a large language model, constructs an emotion-driven fuzzy reasoning mechanism in e-commerce negotiation scenarios, realizes a controllable mapping from discrete emotional states to continuous benchmark concession rates, and solves the problem of no unified mapping between emotion and numerical strategies in the prior art.

[0006] To achieve the above objectives, this invention provides a negotiation method that integrates LoRA fine-tuning of a large language model, comprising the following steps:

[0007] Step 1: Data preparation and LoRA fine-tuning;

[0008] Step 2: Neural perception layer processing to complete the structured perception of user status and price parsing during online negotiation;

[0009] Step 3: Symbolic decision layer processing, which transforms discrete emotional states into continuous, interpretable baseline concession rates through fuzzy logic reasoning;

[0010] Step 4: Deterministic pricing layer processing, which transforms the benchmark concession rate into a precise and secure new offer from the seller;

[0011] Step 5: Introduce an adaptive anti-probing mechanism for robust anti-probing modulation;

[0012] Step 6: Collaborative generation layer processing to generate natural language responses and quotes.

[0013] Optionally, during the data preparation process in step 1, a negotiation corpus based on real e-commerce dialogue scenarios is constructed. The negotiation corpus contains two types of supervised labeled data subsets: intent recognition corpus and sentiment recognition corpus. The intent label space is defined as 6 categories, namely greeting, bargaining, product inquiry, logistics inquiry, transaction, and negotiation failure; the sentiment label space is defined as 8 categories, namely gratitude, happiness, grievance, disappointment, anxiety, fear, anger, and fury.

[0014] During LoRA fine-tuning, low-rank matrix fine-tuning is used, and cross-entropy loss function is used for optimization during the training phase.

[0015] Optionally, the execution process of step 2 includes the following steps:

[0016] Step 2.1: Concatenate the current user's utterance with the historical dialogue context and input it into the large language model fine-tuned by LoRA to obtain the intent prediction result and sentiment prediction result;

[0017] Step 2.2: Analyze the buyer's current bid. If the analysis result is valid, use it directly; otherwise, revert to the previous bid.

[0018] Step 2.3: Output the structured perception vector and write it into the dialogue state.

[0019] Optionally, after completing the structured perception in step 2, the system executes branch control based on the current round of negotiation intent; when the current round of negotiation intent is bargaining and there is a legitimate buyer's offer, steps 3 to 5 are triggered to perform concession reasoning, offer update, and anti-probing modulation; when the current round of negotiation intent is greeting, product consultation, or logistics consultation, steps 3 to 5 are skipped and the system directly proceeds to step 6 to generate an explanatory response; when the current round of negotiation intent is to reach a deal or the negotiation fails, the system enters the end state response branch.

[0020] Optionally, when the current round of negotiations is intended to bargain and there is a legitimate buyer's offer, an emotion-driven fuzzy rule base is constructed in step 3. The Mamdani inference model is used to activate and aggregate multiple rules in parallel to obtain an output fuzzy set. Finally, the centroid method is used to defuzzify the output fuzzy set to obtain a unique continuous benchmark concession rate.

[0021] Optionally, in step 4, the benchmark concession rate is converted into the seller's new offer using a deterministic formula with rigid constraints, as shown in the following formula:

[0022]

[0023] in, This is the seller's bottom price. For the seller's offer in round t, For the buyer's offer in round t, c t The current round's baseline concession rate is γ, and the anti-probing disturbance coefficient is given in step 5. When the anti-probing mechanism is not triggered, γ = 0, and the current round's baseline concession rate is c. t Regular concession updates to control.

[0024] Optionally, in step 5, when the trigger condition that the buyer's bid in the most recent two rounds of the system cache is lower than the seller's bottom price for two consecutive rounds is met, it is determined to be a probing behavior and an anti-probing disturbance coefficient γ is generated.

[0025] Optionally, step 6 employs a collaborative generation mechanism that combines numerical hard constraints and emotional soft constraints. Specifically, it routes responses based on the negotiation intent output by the neural perception layer to fill the quotation hard constraint slots; at the same time, it injects the emotional category as a style control variable and outputs a natural language response through the generator Gen(·).

[0026] This invention provides a negotiation method integrating a LoRA fine-tuned large language model, addressing the pain points of e-commerce negotiation. Specifically, it achieves structured perception of 6 intentions and 8 types of emotions through LoRA fine-tuning of the large model, transforms discrete emotions into continuous baseline concession rates using fuzzy reasoning, designs a deterministic pricing strategy with triple rigid constraints, and combines it with a low-price triggered random perturbation anti-probing mechanism. Consistent natural language responses are generated through numerical hard constraints and emotional soft constraints. This invention solves the problems of unstable perception, irregular pricing, and weak anti-probing in existing methods, improving the robustness and interactive experience of intelligent negotiation systems. It is also lightweight and easily adaptable, providing a feasible solution for the implementation of intelligent negotiation in e-commerce and promoting the integrated application of related technologies in e-commerce scenarios. Attached Figure Description

[0027] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0028] Figure 1 This is a schematic diagram of the model structure in a specific embodiment of the negotiation method that integrates LoRA fine-tuning of a large language model according to the present invention.

[0029] Figure 2This is a schematic diagram of the membership functions of five language levels of concession in a negotiation method that integrates LoRA fine-tuning of a large language model, as presented in this invention. Detailed Implementation

[0030] Embodiments of the present invention are described in detail below, examples of which are illustrated in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain the present invention, and should not be construed as limiting the present invention.

[0031] This invention provides a negotiation method that integrates LoRA fine-tuning of a large language model, comprising the following steps:

[0032] Step 1: Data preparation and LoRA fine-tuning;

[0033] Step 2: Neural perception layer processing to complete the structured perception of user status and price parsing during online negotiation;

[0034] Step 3: Symbolic decision layer processing, which transforms discrete emotional states into continuous, interpretable baseline concession rates through fuzzy logic reasoning;

[0035] Step 4: Deterministic pricing layer processing, which transforms the benchmark concession rate into a precise and secure new offer from the seller;

[0036] Step 5: Introduce an adaptive anti-probing mechanism for robust anti-probing modulation;

[0037] Step 6: Collaborative generation layer processing to generate natural language responses and quotes.

[0038] Please see Figure 1 and Figure 2 The following description, in conjunction with specific embodiments and execution steps, provides further details:

[0039] To avoid ambiguity, the following notation will be used consistently: the seller's reserve price will be denoted as... The seller's offer in round t is denoted as The bid offered by the buyer in round t is denoted as The new offer from the seller in round t+1 is denoted as The baseline concession rate output by the symbolic inference layer is denoted as c. t The anti-probe disturbance coefficient is denoted as γ.

[0040] Step 1: Data Preparation and LoRA Fine-tuning of the Large Model

[0041] To enable large language models to discriminate pragmatic behavior and emotional states in e-commerce negotiation contexts, a negotiation corpus D based on real e-commerce dialogue scenarios is constructed. The negotiation corpus D contains two types of supervised labeled data subsets: intent recognition corpus D... I A total of 3165 sentiment recognition corpus D E A total of 8399 entries.

[0042] The intent label space I is defined as having 6 categories: greeting, bargaining, product inquiry, logistics inquiry, transaction completed, and negotiation failed. The emotion label space E is defined as having 8 categories: gratitude, happiness, grievance, disappointment, anxiety, fear, anger, and fury. These label spaces constrain the discrete set of categories output by the model and provide structured state variables for subsequent symbolic reasoning.

[0043] To achieve domain adaptation without significantly increasing model storage and computational overhead, Low-Rank Adaptation (LoRA) is employed for efficient parameter fine-tuning of the pre-trained model. Let the weight matrix of any linear transformation layer of the pre-trained model be:

[0044] W ∈ R d×k (1)

[0045] During fine-tuning, W is kept frozen and only the incremental parameters consisting of two low-rank matrices are learned to approximate the weight update. Introducing A ∈ R d×r With B ∈ R r×k , where r If min(d, k) is a low-rank dimension, then:

[0046] = AB (2)

[0047] To adjust the adaptation range, a scaling factor α is introduced, and the fine-tuned weights are expressed as follows:

[0048] (3)

[0049] During the training phase, the cross-entropy loss function is used to optimize the LoRA parameters A and B, ensuring that the model's predicted distributions of intent and sentiment labels match the true label distributions, thereby minimizing classification error. The loss function is expressed as:

[0050] (4)

[0051] Where N is the number of training samples, y i The indicator variable for the true label. i This represents the predicted probability output by the model.

[0052] Step 2: Processing of the neurosensory layer

[0053] During online negotiations, the current round of user discourse will be used to... t Context C of dialogue with history t The concatenated data is then input into a large language model fine-tuned by LoRA. This is to obtain structured perception results that can be used for subsequent symbolic reasoning. Specifically, the model outputs posterior probability distributions in the intent label space I and the sentiment label space E, and takes the largest posterior class as the intent prediction result and sentiment prediction result for the current round, respectively.

[0054] (5)

[0055] Among them, I t Indicates the category of intent in the current round of negotiations, e t The structured label indicates the current sentiment category and is used to trigger subsequent fuzzy rule bases and adjust the strength of concession inference.

[0056] Meanwhile, to support deterministic price updates and prevent probing, this step also analyzes the buyer's current round of bids from user discourse. When a valid numerical price quote can be extracted from the text, let... The result is used for parsing; if the current bid cannot be parsed or does not meet the preset numerical validity constraints, the previous buyer's bid is used as the fallback value to maintain the continuity of the interactive state:

[0057] (6)

[0058] Here, Parse(·) represents the price parsing function, preferably including three parts: numerical extraction, unit normalization, and validity verification. Let... To extract and normalize candidate quotes from the text, their validity can be determined as follows:

[0059] (7)

[0060] in and For configurable business price boundaries. If Valid( If ) = 1, then let = Otherwise, trigger the rollback rule. ← This is to maintain the continuity of the interaction state. Finally, this step outputs a structured perception vector.

[0061] o t = (8)

[0062] This information is then written into the current round of dialogue state for use by the symbolic decision layer and the deterministic pricing layer.

[0063] Furthermore, the system is based on the intent of the current round of negotiations. Execute branch control. When When bargaining is needed and a legitimate buyer offers exist in the current round, steps 3 to 5 are triggered for concession reasoning, price updates, and anti-probing adjustments; when For greetings, product inquiries, and logistics inquiries, skip steps 3 to 5 and proceed directly to step 6 to generate an explanatory reply; when When a deal is not reached or negotiations fail, the system will enter the corresponding end status response branch.

[0064] Step 3: Symbolic Decision Layer Processing

[0065] Given that the current round of negotiations aims at bargaining and a legitimate buyer's offer exists, this step employs a symbolic reasoning mechanism based on fuzzy logic to generate the current round's baseline concession rate C in order to transform the discrete emotional state output by the neural perception layer into an interpretable and continuously controllable concession decision. t First, the domain of concession is defined as the normalized interval x ∈ [0,100], and the concession linguistic variable L = {VS, S, M, L, VL} is constructed, corresponding to the concession levels of Very Small, Small, Medium, Large, and Very Large, respectively. To maintain smooth transition and inference stability while covering the entire domain, this embodiment uses a hybrid membership function of Sigmoid, Gaussian, and Generalized Bell forms. The membership functions for the five linguistic items are defined as follows:

[0066] (9)

[0067] (10)

[0068] (11)

[0069] (12)

[0070] (13)

[0071] After constructing the fuzzy variables and membership functions, an emotion-driven fuzzy rule base is built to represent the user's emotional state E. t The degree of concessions compared to the previous round d t-1 The joint mapping represents the proposed concession level for the current round (language output). t. Among them, E t ∈ E represents the 8 types of emotion labels output by the neural perception layer, d t-1 ∈ L represents the linguistic representation of the concession level in the previous round. The rule base uses the IF–THEN format, with typical rules as follows:

[0072] IF Et = Angry AND d t-1 = M THEN t = L (14)

[0073] To ensure the reproducibility and full disclosure of symbolic reasoning, the fuzzy rule base R in this embodiment consists of several IF-THEN rules, with rule inputs covering the combination space of user emotional states and historical concession states. As a preferred implementation, the rule base satisfies the following coverage constraints and construction principles:

[0074] (1) Coverage constraint: The rule base must contain at least a coverage of the E × L combination space, that is, for any E t ∈ E and d t-1 For any L, there exists at least one rule that provides a corresponding suggested concession level. t ∈ L; when computational resources are limited, it can also cover the high-frequency subset of its business, but it should ensure that key areas with negative sentiment and low concession status have regular coverage.

[0075] (2) Monotonicity principle: Under a fixed emotional state Et, it is recommended to give a concession level. t varies with the degree of historical concessions d t-1 The increase is monotonically unchanging; at a fixed historical concession level d t-1 Below, we recommend a concession level. The t-value remains monotonically constant as the intensity of negative emotion increases. This monotonicity constraint is used to suppress cross-turn policy jumps and improve reasoning consistency.

[0076] (3) Rule family representation: Rules can be represented as a family of mapping functions

[0077] t = f(Et, d) t-1 ; R), f : E × L → L (15)

[0078] R is a set of rule parameters that can be calibrated based on business data or expert experience, thereby enabling adaptation to different e-commerce categories and price ranges without changing the inference structure.

[0079] The rule is used to increase the suggested concession level in the current round to mitigate conflict and increase the probability of success when the user shows strong negative emotions and the historical concession level is moderate; when the emotions are relatively positive or the previous round concession was small, a lower concession level is output to protect gains and maintain strategy stability.

[0080] In the rule reasoning phase, the Mamdani reasoning model is used to activate and aggregate multiple rules in parallel. For rule set A... k → B k When the input is x, the membership function of the aggregated output fuzzy set C is defined as:

[0081] (16)

[0082] To obtain a unique, continuous concession decision, this embodiment uses the centroid method to defuzzify the output fuzzy set, thus obtaining the current round's baseline concession rate c. t :

[0083] (17)

[0084] Finally, this step outputs c. t It is also written into the dialogue state for subsequent deterministic quote updates and collaborative generation processes.

[0085] Step 4: Deterministic Quotation Layer Processing

[0086] The baseline concession rate c of the symbolic inference layer output. t To convert to accurate and safe numerical decision-making and avoid out-of-bounds and drift issues in generative models during numerical updates, this step employs a deterministic bid update strategy with rigid constraints to generate the seller's bid in round t+1. Let the seller's reserve price be... The seller's offer in round t is The buyer's offer in round t is The seller's next offer will be calculated using the following formula:

[0087] (18)

[0088] Where γ is the anti-probing disturbance coefficient, the value of which is given in step 5; when the anti-probing mechanism is not triggered, γ = 0, at which time equation (18) is determined by the current round's benchmark concession rate c. t Regular concession update of control

[0089] The aforementioned deterministic pricing strategy achieves boundary safety and trajectory consistency through multiple operators. The outermost maximum operator max{ , Forced constraints No less than the seller's bottom price At the same time, not lower than the bid offered by the buyers in this round. This is to avoid the price falling below the bottom line or making irrational concessions to buyers. The inner minimum operator is min{ The , ·} clause is used to restrict the quoted price from being higher than the previous seller's quote. This prevents price rebounds or abnormal reverse updates, thus ensuring the interpretability and stable convergence of the price sequence. Finally, this step outputs... It is also written into the dialogue state, providing a definite and auditable numerical basis for subsequent collaborative generation layers.

[0090] Step 5: Robust Modulation Against Probing

[0091] To curb uncooperative behavior such as buyers repeatedly offering low prices to probe sellers' bottom line and inducing the system to make continuous concessions, this step introduces an adaptive anti-probing mechanism during the deterministic price update process. The system maintains caches of the two most recent rounds of buyer offers. And based on the bottom line threshold Anomaly detection is performed. When the trigger condition of falling below the baseline for two consecutive rounds is met, it is judged as a probing behavior and a disturbance coefficient γ is generated.

[0092] (19)

[0093] Where ξ is a random variable sampled from a uniform distribution U(0, 1). If the triggering condition is not met, then let γ = 0, indicating that anti-probing modulation is not enabled. The disturbance coefficient γ is used as a defense term in equation (18) for calculation, making the effective concession term (c t —γ) exhibits a random suppression or reverse offsetting effect when the trial is triggered, thereby reducing the predictability of buyers inferring the bottom line or guiding the concession trajectory through fixed-pattern low-price bids. Since the introduction of γ does not change the bottom line constraint and monotonic constraint, the bid is still strictly limited to the safe range. Therefore, this mechanism can improve the robustness and profit protection capability of the system in non-cooperative interactions while maintaining the controllability of numerical decision-making.

[0094] Step 6: Collaborative Generation Layer Processing

[0095] To generate natural, coherent, and emotionally consistent negotiation responses while ensuring strict control over numerical decision-making, this step employs a collaborative generation mechanism combining "hard numerical constraints and soft emotional constraints." Specifically, based on the negotiation intent (It) output by the neural perception layer, response routing is performed. A response framework matching the current pragmatic behavior is selected from a preset template set or a set of generated instructions. The next round of seller's offer output by the deterministic offer layer is then used. Write the required numerical value into the slots to ensure that the quote appearing in the response text is completely consistent with the numerical decision made in step 4. Accordingly, the generated prompt can be formally represented as:

[0096] πt = Prompt (It , et , p s t+1 , Ct) (20)

[0097] Where π t To generate a hint or template instance for the current round, C t Represents dialogue context information, e t This is a signal for the emotional category.

[0098] To achieve an interactive experience that aligns with emotional sentiment, this step further categorizes the emotion into e. t As a style control variable, it is injected into the generation process to adjust the intensity of wording, level of politeness, and appeasement strategies, without changing the already determined numerical price. Therefore, the output of the collaborative generation layer can be expressed as:

[0099] rt = Gen(πt) (21)

[0100] Where r t Gen(·) represents the natural language response output by the seller in round t. Gen(·) represents the generator, which can be a text generation module or a template filling module based on a large language model.

[0101] Through the above design, this invention achieves decoupling and synergy between strategic decision-making and verbal expression: numerical quotations are generated by deterministic formulas and written into the response in a hard-constraint manner, ensuring safety and consistency; emotional signals only adjust the expression style, improving communication effectiveness and reducing conflict intensity. The final output is a natural language response r. t and the corresponding numerical price quote To complete the current round of negotiation interactions.

[0102] In summary, compared with the prior art, the present invention has the following advantages:

[0103] (1) By fine-tuning the large language model through LoRA, the structured output of intent and emotion is realized, thereby improving the stability of state perception for complex natural language input.

[0104] (2) By mapping the emotional state to a continuous concession rate through fuzzy reasoning, smooth adjustment is achieved and cross-round policy jumps are reduced, thereby improving policy interpretability.

[0105] (3) By introducing bottom line constraints, counterparty pricing constraints and monotonic constraints through deterministic pricing strategies, we can avoid pricing out of bounds and numerical instability, and ensure the safe range of pricing.

[0106] (4) By using the anti-probing mechanism triggered by two consecutive rounds of price drops below the floor price and introducing uniform random disturbances, the probability of the opponent inferring the floor price or inducing excessive concessions through repeated low-price probing is reduced, thereby improving robustness and profit protection capabilities.

[0107] (5) Inject pricing as a hard constraint and emotion as a style constraint into the generation module to make the natural language response consistent with the numerical strategy and improve the interactive experience.

[0108] The above description discloses only one preferred embodiment of the present invention, and should not be construed as limiting the scope of the present invention. Those skilled in the art will understand that all or part of the processes of the above embodiments can be implemented, and equivalent changes made in accordance with the claims of the present invention are still within the scope of the invention.

Claims

1. A negotiation method integrating LoRA fine-tuning of a large language model, characterized in that, Includes the following steps: Step 1: Data preparation and LoRA fine-tuning; Step 2: Neural perception layer processing to complete the structured perception of user status and price parsing during online negotiation; Step 3: Symbolic decision layer processing, which transforms discrete emotional states into continuous, interpretable baseline concession rates through fuzzy logic reasoning; Step 4: Deterministic pricing layer processing, which transforms the benchmark concession rate into a precise and secure new offer from the seller; Step 5: Introduce an adaptive anti-probing mechanism for robust anti-probing modulation; Step 6: Collaborative generation layer processing to generate natural language responses and quotes.

2. The negotiation method integrating LoRA fine-tuning of a large language model as described in claim 1, characterized in that, In the data preparation process of step 1, a negotiation corpus based on real e-commerce dialogue scenarios is constructed. The negotiation corpus contains two types of supervised labeled data subsets: intent recognition corpus and sentiment recognition corpus. The intent label space is defined as 6 categories, namely greeting, bargaining, product inquiry, logistics inquiry, transaction, and negotiation failure; the sentiment label space is defined as 8 categories, namely gratitude, happiness, grievance, disappointment, anxiety, fear, anger, and fury. During LoRA fine-tuning, low-rank matrix fine-tuning is used, and cross-entropy loss function is used for optimization during the training phase.

3. The negotiation method integrating LoRA fine-tuning of a large language model as described in claim 2, characterized in that, The execution process of step 2 includes the following steps: Step 2.1: Concatenate the current user's utterance with the historical dialogue context and input it into the large language model fine-tuned by LoRA to obtain the intent prediction result and sentiment prediction result; Step 2.2: Analyze the buyer's current bid. If the analysis result is valid, use it directly; otherwise, revert to the previous bid. Step 2.3: Output the structured perception vector and write it into the dialogue state.

4. The negotiation method integrating LoRA fine-tuning of a large language model as described in claim 3, characterized in that, In step 3, an emotion-driven fuzzy rule base is constructed. The Mamdani inference model is used to activate and aggregate multiple rules in parallel to obtain the output fuzzy set. Finally, the centroid method is used to defuzzify the output fuzzy set to obtain a unique continuous benchmark concession rate.

5. The negotiation method integrating LoRA fine-tuning of a large language model as described in claim 4, characterized in that, In step 4, the benchmark concession rate is converted into the seller's new offer using a deterministic formula with rigid constraints. The formula is expressed as follows: in, This is the seller's bottom price. For the seller's offer in round t, For the buyer's offer in round t, c t The current round's baseline concession rate is γ, and the anti-probing disturbance coefficient is given in step 5. When the anti-probing mechanism is not triggered, γ = 0, and the current round's baseline concession rate is c. t Regular concession updates to control.

6. The negotiation method integrating LoRA fine-tuning of a large language model as described in claim 5, characterized in that, In step 5, when the trigger condition that the buyer's bid in the most recent two rounds of the system cache is lower than the seller's bottom price for two consecutive rounds is met, it is determined to be a probing behavior and an anti-probing disturbance coefficient γ is generated.

7. The negotiation method integrating LoRA fine-tuning of a large language model as described in claim 6, characterized in that, Step 6 employs a collaborative generation mechanism that combines numerical hard constraints with emotional soft constraints. Specifically, it routes responses based on the negotiation intent output by the neural perception layer to fill the quotation hard constraint slots. Simultaneously, it injects the emotional category as a style control variable and outputs natural language responses through the generator Gen(·).

Images