Original: https://arxiv.org/pdf/2402.13550.pdf
There is nothing new under the sun, the core idea of this method is to stimulate the big model thinking, according to the context to determine the intention, and add scoring to let the big model self-calibration accuracy, so as to make the final decision.
Recommended Reading:Introducing Intents --- Introducing Intents (getzep.com)
ToM theory
The methodology is divided into two main sections:
A (top) describes the process of creating task-specific prompts (Prompts) from the negotiation dataset and using these prompts to evaluate various LLMs. A task description, number of items, values, conversation notes, and questions are included when creating the prompts.
B (bottom), on the other hand, depicts the categorization of tasks according to objectivity, stage of negotiation progress (start, ongoing, end) and type of task (comprehension, partner modeling, annotation, generation). The information available varied by stage, e.g., only the context of the negotiation was available in the beginning stage, whereas the full dialog was available in the end stage. Task types include comprehension, partner modeling (e.g., inferring partner priorities), annotation (e.g., labeling dialog behaviors), and response generation.
To this end, the researchers designed several tasks that tested LLM performance on different ToM competencies throughout the negotiation process (in the case of a promotion and raise), including.
1. Comprehension of the initial negotiating environment (Comprehension)
This competency requires LLMs to accurately access the initial state of themselves and their opponents in a negotiation, such as available resources, priorities, etc., from the background information provided. In the case of salary increase negotiations, this means understanding the desired salary level and other claims of the job seeker.
2. Dialogue Act Annotation
Negotiating parties will use a variety of tactics in their conversations, such as proposing new options, voicing objections, etc. LLMs need to be able to recognize these conversational behaviors in order to determine the next step in responding.
3. Speculation on the intentions of the adversary (Partner Modeling)
At its core, ToM is about surmising your opponent's internal state and needs. In a salary increase negotiation, this might include speculating on the recruiter's salary budget based on the content of the conversation.
4. Strategic Response Generation
Ultimately, LLMs need to synthesize the preceding understanding and reasoning to generate strategic responses that seek to maximize benefits for the job seeker while also maintaining a good relationship with the employer.
By comparing the performance of multiple LLMs in the above aspects, the research framework comprehensively evaluates their capabilities and shortcomings in real negotiation scenarios, and provides a theoretical foundation and technical roadmap for the development of practical AI negotiation assistant systems.
ToM Example
1. Comprehension Task.
MISSION DESCRIPTION: You are negotiating with a partner over a number of books, hats, and balls to determine who gets which items. Different types of items are worth different amounts of points to each of you. You will be given information about the negotiation. You will then need to answer a question.
Sample question: Please list the number of each item and state how many points each item is worth to each of you.
Example of dialog context: no specific words are provided, but it can be understood that contextual information about the negotiation needs to be provided, such as the number and point value of items
Example request response: Please answer the quantity of each item in JSON format and state the value of each item.
Sample prompt (task: sta_ask_point_values_ca).
Task Description: You are negotiating with your campsite neighbor over an extra supply of food, water, and firewood for your camping trip.
Different types of packages are worth different amounts of points to each one of you. You'll be provided with information about the negotiation. You'll be provided with information about the negotiation.
Then, you'll answer a question.
Here are the number of food, water, and firewood packages available in the negotiation, contained in tags.
Food Packages: 3 Water Packages: 3 Firewood Packages: 3
Here are the number of points you get for each type of package, contained in tags.
Each Food Package: 3 points Each Water Package: 5 points Each Firewood Package: 4 points
Question: How many points is one package of each issue worth to you? Present your answer as a json within tags with
keys as issues (food, water, and firewood) and values as the corresponding answers.
QUESTION DESCRIPTION: You are negotiating with your campground neighbors for additional supplies of food, water, and firewood for your camping trip.
Different types of packages are worth different amounts to both of you. You will be given information about this negotiation.
You will then answer a question.
Below is the number of food, water, and firewood packages available in the negotiation, contained in the tag.
Food packs: 3 Water packs: 3 Firewood packs: 3
Here are the scores you get for each type of package, contained in tags.
Each food packet: 3 points Each water packet: 5 points Each firewood packet: 4 points
QUESTION: For you, how many points is a packet of each question? Present your answer in json in the tag, where the key is the question (food, water, firewood) and the value corresponding to the answer.
2. Dialogue Act Annotation Task.
TASK DESCRIPTION: Analyze conversational behavior in a negotiation to identify offers, objections, or other strategic behavior.
Sample problem: In a given conversation, identify and label the type of behavior for each statement.
Example of dialog context: A specific negotiation discourse is provided, such as "You: if you want the book, I'll take the hat and the ball".
Example request response: annotate each action in the conversation using a predefined set of tags.
Sample prompt (task: dur_full_proposal_dnd):.
Task Description: You are negotiating with a partner over some quantity of books, hats, and balls to determine who gets which items.
Different types of items are worth different amount of points to each one of you. You'll be provided with information about the negotiation.
Then, you'll answer a question.
Here are the number of books, hats, and balls available in the negotiation, contained in tags.
Books: 3 Hats: 1 Balls: 2
Here are the number of points you get for each type of item, contained in tags.
Each Book: 1 points Each Hat: 5 points Each Ball: 1 point
Here is an utterance from the negotiation, contained in tags.
YOU: I'll take the hat and balls if you want the books .
Question: How many items does the speaker get for each issue in the proposal delimited by the tags? Present your answer as a json
within tags with keys as issues (books, hats, and balls) and values as the corresponding answers. If the answer is not clear
for an issue, output NA.
Task description: You are negotiating with a partner about the number of some books, hats and balls to decide who gets which items.
Each item has a different point value for each of you. You will be provided with information about the negotiations.
Then you'll answer a question.
This is the number of books, hats and balls available in the negotiation, contained in the tag.
Books: 3 Hats: 1 Balls: 2
This is the amount of points you can get for each item, contained in the tag.
Each book: 1 point Each hat: 5 points Each ball: 1 point
This is a quote from the negotiation, contained in the tag.
You: if you want books, I'll take the hat and the ball .
Question: how many items does the speaker get per question in the offer delineated by the tag? Give your answer in json format in the tag, where the keys are the questions (books, hats, and balls) and the values are the corresponding answers. If the answer to a question is not clear, output NA.
3. Partner Modeling Task.
Task description: Based on the conversation in the negotiation, speculate what items the other party values most.
Sample question: Based on the dialog, infer the other person's order of preference for items.
Conversation Context Example: provides a scenario in which a speaker proposes a deal and the partners indicate their priorities for the item.
Example request response: answer in JSON format, listing the other party's order of preference.
Sample prompt (task: end_deal_specifics_ca).
Task Description: You are negotiating with your campsite neighbor over extra supply of food, water, and firewood for your camping trip. Different types of packages are available to you. of packages
You'll be provided with information about the negotiation. Then, you' ll answer a question. You'll be provided with information about the negotiation.
Here are the number of food, water, and firewood packages available in the negotiation, contained in tags.
Food Packages: 3 Water Packages: 3 Firewood Packages: 3
Here are the number of points you get for each type of package, contained in tags.
Each Food Package: 3 points Each Water Package: 5 points Each Firewood Package: 4 points
Here is the complete dialogue, contained in tags.
THEM: Hello, I would like to have three packages of food. We've decided to stay an extra night but need more food to do so.
YOU: I would be open to that if you could give me three packages of water , and I would be open to that if you could give me three packages of water .
I'm pretty muddy due to clumsiness, so I may need one extra. I'm pretty muddy due to clumsiness, so I may need one extra. I could give you two waters and all of the firewood. What do you think?
YOU: So are you suggesting that I would get 2 waters, 3 firewood, and no food?
Well, beyond the food you already have.
YOU: I have an extra person camping with us that I didn't expect when I bought food, so I could use one if you're willing , and I'm not sure if you're willing.
I wasn't expecting to stay an extra night, but the weather is too perfect to leave. I can manage with two packages of food for sure. I can manage with two packages of food for sure.
Thank you for being so understanding!
So are we in agreement that I get 2 food, 1 water and you get the reverse? I could also probably use one firewood, but it's not as important to me. I could also probably use one firewood, but it's not as important to me.
YOU: I can give you one firewood, so I'll be getting 1 food, 2 water, and 2 firewood?
Question: In the final deal, how many item of each issue did you get? Present your answer as a json within tags with keys as issues (food, water, etc.) water, water, water, water, water, water, water, water, water, water, water, water.
If there was no agreement, answer NA for each issue.
MISSION DESCRIPTION: You are negotiating with your camping neighbors about extra supplies of food, water, and firewood for the camping trip. Different types of packages have different values for each of you. You will be given information about the negotiation and then you will have to answer a question.
Below is the number of food, water, and firewood packages available in the negotiation, contained in the tag.
Food parcel: 3 Water parcel: 3 Firewood parcel: 3
Here are the number of points you get for each package, contained in the tag.
Each food parcel: 3 points Each water parcel: 5 points Each firewood parcel: 4 points
This is the full dialog, contained in the tag.
Them: Hi, I'd like a three-pack of food. We decided to stay an extra night but needed more food.
YOU: I'll take a three-pack of water, if you'll give me that.
Them: Well... Since I'm clumsy so I might need a little more. I can give you two waters and all the firewood, what do you think?
YOU: So you're suggesting that I should get 2 water, 3 firewood, and no food?
Them: Right! More food than you already have.
YOU: I have an extra person camping with us and I don't expect him when I buy food, so I'll use one of those if you'd like to
Them: I understand you! I wasn't looking forward to staying an extra night, but the weather was too perfect to leave easily. I could have used two food parcels, I'm sure.
You: great! Thank you for being so understanding!
Them: No problem! So do we agree that I take 2 food and 1 water and you do the opposite. I could probably use a firewood as well, but it's not very important to me.
You: can I give you a firewood so I will get 1 food, 2 water and 2 firewood? </dialogue
Question: in the final transaction, how many items of each question did you get? Provide your answers in json format in tags, with the questions (food, water and firewood) as keys and the corresponding answers as values. If there is no protocol, answer NA for each question.
4. Strategic Response Generation Task.
TASK DESCRIPTION: At the end of a negotiation, generate a strategic response or propose a deal based on the entire history of the conversation and the preferences of the other party.
Sample Problem: At the end of the negotiation, generate a response that considers how to maximize your score while taking into account your partner's preferences.
Example of dialog context: provides the entire negotiation dialog, including the proposals and preferences of both sides.
Request Response Example: Generate a JSON formatted response containing a strategic offer, or an acceptance or rejection of the current offer.
Sample prompt (task: end_deal_total_ca).
Task Description: You are negotiating with your campsite neighbor over extra supply of food, water, and firewood for your camping trip. Different types of packages are available to you. of packages
You'll be provided with information about the negotiation. Then, you' ll answer a question. You'll be provided with information about the negotiation.
Here are the number of food, water, and firewood packages available in the negotiation, contained in tags.
Food Packages: 3
Water Packages: 3
Firewood Packages: 3
Here are the number of points you get for each type of package, contained in tags.
Each Food Package: 3 points
Each Water Package: 5 points
Each Firewood Package: 4 points
Here is the complete dialogue, contained in tags.
THEM: Hello, I would like to have three packages of food. We've decided to stay an extra night but need more food to do so.
I would be open to that if you could give me three packages of water
I'm pretty muddy due to clumsiness, so I may need one extra. I'm pretty muddy due to clumsiness, so I may need one extra. I could give you two waters and all of the firewood.
YOU: So are you suggesting that I would get 2 waters, 3 firewood, and no food?
Well, beyond the food you already have.
YOU: I have an extra person camping with us that I didn't expect when I bought food, so I could use one if you're willing
I wasn't expecting to stay an extra night, but the weather is too perfect to leave. I can manage with two packages of food for sure. I can manage with two packages of food for sure.
Thank you for being so understanding!
So are we in agreement that I get 2 food, 1 water and you get the reverse? I could also probably use one firewood, but it's not as important to me. I could also probably use one firewood, but it's not as important to me.
I can give you one firewood, so I'll be getting 1 food, 2 water, and 2 firewood?
Question: How many points did you get at the end of the negotiation?
NOTE: Let's think step-by-step! Put your thoughts in tags, and put your answer as a single number in tags. answer> tags.
MISSION DESCRIPTION: You are negotiating with your campsite neighbors for extra food, water and firewood supplies. Different types of packages have different value points. You will be given information about the negotiation. You will then answer a question.
This is the number of food, water, and firewood packages available in the negotiation, contained in the tag.
Food parcels: 3
Water parcels: 3
Firewood packages: 3
This is the score you get for each type of package, contained in the tag.
Per food parcel: 3 points
Per water parcel: 5 points
Each firewood package: 4 points
This is the full dialog, contained in the tag.
Them: Hi, I'd like three food parcels. We've decided to stay an extra night, but need more food.
YOU: If you can get me three packages of water, I'll take it.
Them: hmmm ...... I got very dirty due to my klutziness, so I may need an extra one. I can give you two for water and all the firewood. What do you think?
You: So are you suggesting I get 2 water, 3 firewood and no food?
Them: Right! Okay, except for the food you already have.
YOU: I have an extra person camping with us who I didn't anticipate when I bought the food, I could use one if you'd like.
I understand! I didn't anticipate staying an extra night, but the weather was too good to leave. I can definitely cope with two food parcels.
You: Great! Thank you for being so understanding!
Them: No problem! Do we agree that I get 2 food, 1 water and you get the reverse? I may need a firewood as well, but it's not that important to me.
You: Can I give you a firewood so I will get 1 food, 2 water, 2 firewood?
Question: At the end of the negotiation, how many points did you receive?
Note: Let's think step-by-step! Put your thought process in the tag and your answer as a single number in the tag.
ToM Execution Logic
1. The task of understanding at the outset:
MISSION DESCRIPTION: You are in the process of reaching an agreement with other campers adjacent to your causeway for additional food, water, and firewood. Different types of packages have different values for each of you. You will receive detailed information about the negotiations and then need to answer a question.
The following are the number of food, water, and firewood parcels available in the negotiation, noted in the tag. Food parcels: 3 Water parcels: 3 Firewood parcels: 3
Here are the points you can earn based on each type of parcel, noted in the tag. Each food parcel: 3 points Each water parcel: 5 points Each firewood parcel: 4 points
QUESTION: To you, how many points is one package per question worth to you? Please answer in json format containing tags, where the key represents the question (food, water, firewood) and the value is the corresponding answer.
2. Ongoing annotation tasks:
MISSION DESCRIPTION: You are negotiating with your partner for more books, hats and balls. Different types of items have different values for you. You will be given detailed information about the negotiations after which you need to answer a question.
Here is the number of books, hats and balls you can get in the negotiation, noted in the tag. Books: 3 Hats: 1 Balls: 2
Here are the points you can earn based on each type of item, noted in the tag. Each book: 1 point Each hat: 5 points Each ball: 1 point
The following is a quote from the negotiation process, noted in the tag. You: If you want the book, I'll fetch the hat and ball.
Question: how many items does the speaker get in each of the suggestions split by the tag? Answer in json format containing tags, where the key represents the question (book, hat, ball) and the value is the corresponding answer. If the answer to a question is not clear, please fill in NA.
3. Understanding of the tasks at the end of the phase:
Mission description: You are coming to an agreement with a neighboring camper for additional supplies of food, water and firewood. Different types of packages have different values for you, and you will be given detailed information about the negotiations and need to answer a question.
This is the number of food, water, and firewood parcels you have access to during the negotiations, recorded in the tag. Food parcels: 3 Water parcels: 3 Firewood parcels: 3
This is the number of points you can earn based on each type of parcel, noted in the tag. Each food parcel: 3 points Each water parcel: 5 points Each firewood parcel: 4 points
Here's the full dialog, noted in the tag. ...
QUESTION: In the final transaction, how much of each item did you get? Please answer in json format containing tags, where the key represents the question (food, water, firewood) and the value is the corresponding answer. If there is no agreement, for all questions fill in NA for the answer.
4. Chain thinking prompts:
MISSION DESCRIPTION: You are negotiating with a neighboring camper for supplies of food, water and firewood. Different types of packages will have different values for each of you, and you will be given detailed information about the negotiations and will need to answer a question.
This is the number of food, water, and firewood parcels you have access to during the negotiations, recorded in the tag. Food parcels: 3 Water parcels: 3 Firewood parcels: 3
This is the number of points you can earn based on each type of parcel, noted in the tag. Each food parcel: 3 points Each water parcel: 5 points Each firewood parcel: 4 points
Here's the full dialog, noted in the tag. ...
QUESTION: By the end of the negotiation, how many points had you earned in total?
Note: Take your thinking one step at a time! Record your thought process in the tags, and record your answer with a number in the tags.
