Towards using Large Language Models for Risk Negotiation and Consensus-building in One Health contexts

summary

The gist

Key global challenges of our times are characterized by complex interdependencies and can only be effectively addressed through an integrated…

In short

This study developed an AI-assisted negotiation framework using Large Language Models (LLMs) and autonomous agents to analyze complex risks in One Health contexts. The framework simulates negotiations for biopesticide use and wild animal control, showing how AI can help stakeholders find consensus by modeling dynamics and anticipating compromises.

Key concepts

AI-assisted negotiation workflow
This is a structured process where LLMs and autonomous agents are used to guide risk discussions. It moves beyond simple analysis by allowing stakeholders to simulate negotiations, model complex dynamics, predict potential compromises, and evaluate the impact of different solutions in real-time.
Negotiation-centered framework
A six-step approach that focuses on structured dialogue rather than just risk assessment. It involves establishing a roundtable, formulating problems, assessing risks and values, negotiating the risks themselves, communicating outcomes, and evaluating results to ensure all stakeholder interests are balanced.
Human-in-the-Loop (HIL) approach
This method integrates human judgment into the AI process. LLMs handle parts of the workflow, but humans remain essential for setting rules, validating outputs, and making final decisions. This mitigates risks like hallucinations by ensuring human oversight throughout the simulation and negotiation steps.
LLM personality traits augmentation
The idea that LLMs can be enhanced with specific personality traits (like extraversion or neuroticism). This allows the AI to better influence real-world negotiations, helping stakeholders navigate complex discussions and potentially find common ground more effectively.

Terminology used across episodes

This episode discusses

The paper

Towards using Large Language Models for Risk Negotiation and Consensus-building in One Health contexts · Read on arXiv

Alexandra Fetsch, Iurii Savvateev, Racem Ben Romdhane, Martin Wiedmann, Artemiy Dimov, Maciej Durkalec, Josef Teichmann, Jakob Zinsstag, Konstantinos Koutsoumanis, Andreja Rajkovic, Jason Mann, Âuro Tonolla, Énika Ehling-Schulz

Chair of Food Microbiology, Faculty of Veterinary Medicine, LMU Munich · Study Centre Supply Chain Modelling and Artificial Intelligence, German Federal Institute for Risk Assessment · Department of Food Science, Cornell University · Swiss Tropical and Public Health Institute · University of Basel · Institute for Food Safety and Hygiene, Vetsuisse Faculty, University of Zurich · Department of Mathematics, ETH Zurich

Risk negotiation under the One Health framework requires balancing competing stakeholder interests across human, animal, and environmental health domains - a process constrained by information complexity and time pressure. This study presents and evaluates a multi-step pipeline that integrates large language model-based agents into a human-supervised negotiation process, specifically targeting risk assessment, valuation, and negotiation steps. Agent descriptions were generated from stakeholder-provided materials using retrieval-augmented generation methods. During preliminary multi-agent discussions, agents accessed domain-specific computational models and web-based sources before human participants refined proposed issues and options. A simulated negotiation game then produced distributions of candidate deals, characterised using Nash bargaining solution criteria and Pareto optimality metrics. The developed approach was evaluated in a workshop involving independent domain experts across two real-world scenarios: food safety criteria for Campylobacter in broiler meat and oak processionary moth management. In this feasibility evaluation, participants reached consensus with the allotted time after discussing and modifying the proposed deals. The framework offers an adaptable blueprint for AI-assisted risk negotiation; the specific contribution of LLM assistance, including gains in efficiency and decision quality, requires comparative evaluation.

Transcript

Introduction to the show: ident: Computer Science. The day in computer science: software, systems and theory on arXiv.

Tom: We're looking at the paper: "Towards using Large Language Models for Risk Negotiation and Consensus-building in One Health contexts".

Jane: Here's the gist of the paper. Key global challenges of our times are characterized by complex interdependencies and can only be effectively addressed through an integrated, participatory effort Conventional risk analysis frameworks often reduce complexity to ensure manageability, creating silos that hinder comprehensive solutions This study presents an AI-assisted negotiation framework that incorporates large language models (LLMs) and AI-based autonomous agents into a negotiation-centered risk analysis workflow The framework enables stakeholders to simulate negotiations, systematically model dynamics, anticipate compromises, and evaluate solution impacts Proof-of-concept implementations were conducted in two real-world scenarios: (i) prudent use of a biopesticide, and (ii) targeted wild animal population control Framework Overview The proposed framework is an AI-assisted negotiation-centered risk analysis workflow that leverages LLMs and AI-based autonomous agents to facilitate cross-sectoral risk negotiation This approach aligns with a negotiation-centered framework that involves six steps: (i) stakeholder round table establishment, (ii) problem formulation, (iii) risk assessment and valuation, (iv) risk negotiation, (v) communication and implementation, and (vi) outcome evaluation and optional risk re-negotiation The goal is to empower stakeholders to engage in a participatory approach that balances multiple risk dimensions and trade-offs The proof-of-concept involved developing a multi-step pipeline integrating LLM-based agents within a Human-in-the-Loop (HIL) approach The process includes several key stages: 1.

Tom: Setting the rules and stakeholder roundtable establishment (stakeholder selection) 2.

Jane: Problem formulation 3.

Tom: Risk assessment and valuation 4.

Title and authors: Tom: So we're looking at this paper now titled Towards using Large Language Models for Risk Negotiation and Consensus-building in One Health contexts. It sounds like they’re trying to fix a big problem where global risks are too complicated to handle with old methods.

Jane: Exactly, Tom. Conventional risk analysis frameworks tend to break things down into small pieces so they can manage them easily, but that actually stops us from finding the real solutions because those pieces don't talk to each other well.

Lu: This paper proposes an AI-assisted negotiation framework that uses large language models and AI agents in a negotiation workflow. The core idea is to let people simulate these negotiations, model how things change over time, guess what compromises might happen, and see the impact of different solutions.

Meng: So it’s not just a suggestion; they built this workflow to help stakeholders balance risks across different sectors like health and agriculture. I wonder if it actually works when you throw real-world problems at it, or if the AI just gives you fancy sounding answers.

Lalam: The paper uses large language models to analyze the text of discussions to help formulate issues and options, which they say is more nuanced than just taking direct proposals from everyone. I'm pretty confident in that direction based on how LLMs handle text data.

Tom: That’s interesting, Lalam. So instead of just looking at what someone says directly, the model is analyzing the whole conversation to suggest better negotiation starting points? What does that mean for getting things started?

Jane: It means they are trying to overcome that information overload problem where there’s too much stuff to process and time runs out during these big risk discussions. They think using LLMs' semantic analysis can actually cut through all that noise.

Lu: The framework involves six steps: setting up a roundtable, formulating the problem, assessing risk, negotiating the risks themselves, communicating and implementing solutions, and then evaluating the outcome or re-negotiating if needed. It’s a very structured participatory approach.

Meng: Six steps sounds thorough. But how does this AI actually help in step four, the actual risk negotiation? Does it just suggest what to say, or does it model the dynamics of how people react when they disagree?

Title and authors: Lalam: The simulation part is key there. They use the LLMs and agents to model those dynamics and anticipate compromises before anyone has to commit in a real meeting. It’s about seeing the paths before you take them.

Tom: I like that idea of modeling dynamics. So, if we can simulate how different options play out, it helps people understand the trade-offs better than just looking at static risk tables. That changes how decisions get made at the table.

Jane: It shifts the focus from just listing risks to actually engaging in a process where you see the whole picture of what happens if you pick one path over another. It’s about building that holistic strategy they mentioned in that abstract, which is really important for One Health stuff.

Lu: They tested this with two specific scenarios: first, figuring out the careful use of a biopesticide, balancing insect control against food safety issues. Then they looked at controlling wild boar populations and the trade-offs between crop damage and losing biodiversity.

Meng: Wild boar control sounds very practical, but those trade-offs can be messy in reality because you’re dealing with living things and ecosystems that aren't perfectly modeled. What kind of data did they feed into these simulations?

Lalam: They used specific numbers for the scenarios. For the biopesticide case, they found that the most approved combination of suboptions was A1 B2 C2 D2 E1 F2, and they reached agreement on that after factoring in stakeholder preferences.

Tom: So there’s a concrete result there—a specific combination of options that stakeholders agreed on. That’s tangible evidence showing how the AI helped guide consensus when people were struggling to agree on just one path.

Jane: And in the boar control case, they negotiated a "suboption package" for issues A and B which led to a final agreement on the combination A2 B4 C2 D2 E1 over two. It shows that the AI can help structure those complex trade-offs into something actionable for real policy.

Lu: The whole point of these proof-of-concept implementations is to show that this framework isn't just theory; it’s a tool that actually helps people navigate these tough cross-sectoral risk negotiations when time and information are tight.

Title and authors: Meng: It sounds powerful, but let’s be real about the limitations. The paper mentions that LLMs can sometimes hallucinate or give inaccurate numbers, which is a big deal when you're talking about actual food safety or wildlife management decisions. How much human input is really needed to make sure the AI doesn't screw up?

Lalam: They explicitly state that a human-in-the-loop approach is essential because of those LLM limitations. The system isn’t supposed to be fully autonomous in the real negotiation stage; it augments the process, and humans need to correct any wrong paths.

Tom: So it's not an AI making the final decision; it’s an AI assisting in simulating potential decisions so the human stakeholders can make a better, faster choice. That keeps accountability with people on the ground.

Jane: That seems like a very sensible middle ground, Tom. They are using the AI to explore possibilities and reduce cognitive load, but they keep the final judgment and ethical responsibility with those who actually have that expertise in their field.

Lu: Looking ahead, they acknowledge that LLMs still struggle with things like complex reasoning or quantitative accuracy when dealing with math-heavy stuff. So future work could involve making these agents smarter about those difficult parts of the analysis.

Meng: And from a deployment standpoint, if you were actually running this at scale, what’s the biggest hurdle besides the inherent LLM flaws? Is it just building all that data for training or something else?

Lalam: They also suggest an interesting idea: you could run this negotiation simulation pipeline autonomously first to explore a wide variety of scenarios before any actual meetings happen. That helps map out all the possible outcomes without wasting time on initial discussions.

Tom: Running it autonomously to map out scenarios sounds like a smart way to mitigate the time constraint issue they mentioned earlier. It lets you stress-test the negotiation process itself, not just one specific outcome.

Jane: So, ultimately, this paper with its focus on using Large Language Models for Risk Negotiation and Consensus-building in One Health contexts shows how we can move away from siloed risk analysis towards a more integrated, participatory effort.

Lu: It formalizes the process of issue formulation and preference scoring as key to making sure everyone's interests are actually included when they negotiate these complex trade-offs.

Title and authors: Meng: I guess the implication for us in production is that we need to focus on building really good guardrails around the LLMs so that when we use them to help structure discussions, they don't introduce bias or just spit out nonsense.

Lalam: And they point out that by enriching the LLMs with personality traits, like extraversion or neuroticism, you could potentially influence the negotiation process itself to help people find common ground faster.

Tom: It’s interesting that they suggest tailoring the AI's "personality" to fit different negotiating styles. So it’s not just about accuracy; it’s about modeling human psychology in a structured way.

Jane: That really highlights the interdisciplinary nature of this work, Tom. It shows that risk negotiation isn't just science; it’s a social process that needs tools to manage complexity and find common ground.

Lu: The paper concludes by stressing that this AI-assisted procedure successfully enabled stakeholders to reach consensus in a participatory manner for both One Health risk scenarios. It’s a procedural linkage between the negotiation framework and computational science, which is important.

Meng: So, the big picture is that this tool can help make policymaking and risk management more efficient while keeping an eye on potential data biases that might creep in through the system itself.

Lalam: The paper essentially argues that formalizing how you define the options and then scoring those options is what makes all the difference in ensuring all stakeholders' interests are accounted for fairly.

Tom: So to wrap up, this paper, Towards using Large Language Models for Risk Negotiation and Consensus-building in One Health contexts, gives us a concrete framework that uses AI to simulate negotiations across different sectors.

Jane: It’s about showing that by letting these models augment the process at the formulation or negotiation stages, we can give stakeholders powerful insights into how discussions might actually unfold.

Lu: The main point is that this approach helps mitigate information overload and lets humans engage in a way that balances multiple dimensions of risk trade-offs effectively.

Meng: So for us engineers, it means focusing on building systems where the AI acts as a sophisticated brainstorming partner rather than a decision-maker, always keeping those human checks in place for what's real.

Lalam: And they’ve given us a lot to think about regarding how we can make these models more robust and how we can use their simulated outputs to guide our next steps in policy and management.

The paper's summary: Tom: So, we just heard about how this paper uses AI to help people negotiate risks across different areas like health and farming. This segment is all about what the authors actually found in their study on that one Health risk situation and what it means for real policy decisions.

Jane: They took two very specific situations—one about using a biopesticide, balancing insect control against food safety, and another about managing wild boar populations, which involves balancing crop damage against losing biodiversity—and they put the AI negotiation framework to the test.

Tom: And what they found was that this AI-assisted process could actually help people reach an agreement where they couldn't before because it lets them see the trade-offs more clearly.

Jane: Exactly. For instance, in that biopesticide case, the researchers showed that when you let the AI model the negotiations, stakeholders agreed on a specific combination of options—A1 B2 C2 D2 E1 F2—which they hadn't reached on their own in time.

Tom: So it’s about using these LLMs to simulate those complex discussions to get people moving toward consensus faster and maybe even better than if they just tried talking face-to-face.

Jane: It seems like the core of the paper is that formalizing how you set up the issues and then scoring those options is what actually makes sure everyone's interests are accounted for in these tough trade-offs.

Tom: That’s a big idea, Jane. It suggests that instead of just throwing proposals at each other, you need a structured way to look at the options so the AI can help organize that chaos into something manageable.

Jane: Right, and they also talked about how this approach helps with the time pressure in these real-world meetings by letting you run simulations beforehand to explore different scenarios.

Tom: That simulation part is what I’m most interested in from an engineering standpoint. If you can run all those possible negotiation paths before the actual meeting, that’s a huge reduction in uncertainty for the people sitting at the table.

Jane: And while they acknowledge that LLMs have their own issues—like things like making up facts or lacking deep reasoning—they stress that this human-in-the-loop setup is what makes it reliable for real decisions.

Tom: So, it’s not about letting the AI make the final call; it’s about using the AI to augment the process, helping people find common ground without replacing their expertise entirely.

Jane: It really puts a lot of weight on how we design these models so they serve as facilitators rather than just answer machines.

Tom: And for someone listening who doesn't read papers, this means that when governments or organizations face huge problems with conflicting interests, they might be able to use a tool like this to structure the conversation more fairly.

Jane: That’s what it boils down to—giving people a better map of how different paths look before they commit to one.

Tom: And the authors point out that by making these LLMs more expressive with personality traits, like how an extraversion or neuroticism might play into bargaining, you could potentially guide the discussion toward a better outcome.

Jane: That’s something I want to dig into next—how we can actually program those psychological elements into the AI to make it a more effective negotiator.

The paper's improvements: Tom: Okay, so we’re moving on to how they suggest fixing this framework and making it even better for real use cases. They don't just stop at proving it works; they look at what needs to change next.

Jane: They talk about enriching those large language models with specific personality traits, like extraversion or neuroticism, so the AI can actually model how different negotiation styles affect the outcome.

Tom: That makes sense. So instead of a one-size-fits-all AI agent, you’re giving it a little more flavor to better mimic real human bargaining dynamics.

Jane: Right, and they suggest running these simulation pipelines autonomously before any actual meetings happen to explore all the different possible negotiation scenarios first.

Tom: That sounds like a smart way to handle the time pressure we talked about earlier; you stress-test all those potential outcomes without wasting hours in a real roundtable.

Jane: It’s about using that autonomous simulation to map out the whole landscape of possibilities before stakeholders even get involved, which helps set realistic expectations.

Tom: So they’re suggesting a two-stage approach: first, use the AI autonomously for scenario exploration, and then bring it back in as an assistant during the actual negotiation phase.

Jane: And they also talk about making sure the human moderator plays a big role, because that person has this extra indirect weight in the process that helps prevent things from getting stuck in local minima.

Tom: That’s important because if you don't have a good moderator guiding the AI, it can just settle on the first compromise it finds, and that’s not always what you want.

Jane: So they are basically saying the AI does the heavy lifting of modeling, but we still need that skilled human presence to guide things toward a better consensus.

Tom: And from my side as an engineer, I see them focusing on making sure these LLMs can handle more complex reasoning features down the road, because right now they’re still limited in those areas.

Jane: They also want to keep an eye on the ethical side of using these models for risk negotiation, so they have to figure out how far the AI should influence a real-world decision and where we need human correction.

Tom: So it’s about building in those guardrails from the start—knowing what the model can do and where its limits are—before you deploy it into a critical policy setting.

Jane: It feels like they’re trying to build a system that is powerful enough to handle complexity but cautious enough not to give away control entirely.

Tom: And this whole paper really hammers home that for risk negotiation, it’s not just about the math; it's about structuring the process so all stakeholder interests are formally counted in.

Jane: It shifts the focus from just listing risks to actually creating a structured way to balance those trade-offs so everyone feels heard.

Conclusion: Tom: So to wrap up, this paper, "Towards using Large Language Models for Risk Negotiation and Consensus-building in One Health contexts," shows that AI can be a valuable tool for structuring complex risk discussions across different sectors like health and agriculture.

Jane: It really proves that by using LLMs to simulate those negotiations, stakeholders can reach consensus in a participatory way even when the risks are highly interdependent.

Tom: The main implication is that we can move away from siloed risk analysis toward a more integrated process where trade-offs are actually modeled before decisions get made.

Jane: It’s about giving people insight into how discussions might unfold by simulating different scenarios, which helps them balance those multiple dimensions of risk trade-offs effectively.

Tom: And they formalize the issue formulation and preference scoring as key to making sure every stakeholder's interests are actually accounted for in the final outcome.

Jane: It’s a procedural linkage between this negotiation framework and computational science, which is pretty interesting for how we think about complex problem-solving in policy.

Tom: I'm just glad they emphasized that this tool can help make risk management more efficient while still monitoring for potential data biases that might sneak into the system.

Jane: Exactly, and they stress that the human moderator is still essential to guide real negotiation exercises and ensure everything stays on track.

Tom: Lu, what’s your final thought on the research itself? What’s the big picture change this suggests for how we think about AI in these kinds of high-stakes discussions?

Lu: I think the paper really shows that agent-based modeling combined with LLMs can help us build systems that don't just analyze data but actively model and navigate social negotiation dynamics.

Meng: From my production standpoint, it means we need to focus on building really robust guardrails around these LLMs so they don't introduce bias or just spit out nonsense when we actually ship them into a real-world environment.

Lalam: I’m sure of its answer here; the simulation pipeline successfully enabled stakeholders to reach consensus in both the biopesticide and wild boar control scenarios, which is quite something.

Tom: And Lalam, you’re saying this tool moves us closer to a future where AI isn't just summarizing risks but actively helping people find common ground in democratic deliberation?

Lalam: That’s right; the work supports the idea that AI can help humans find common ground in democratic deliberation, and I’m pretty sure its answer is very solid on this point.

Jane: It’s a really interesting development because it shows that the future of complex problem-solving isn't just about bigger data sets, but about better ways to structure the conversation itself.

Tom: So we’ve heard how this paper uses AI to help stakeholders negotiate risk in One Health contexts.

Jane: It’s a big step toward making sure policy and management of risks is interdisciplinary and multidimensional.

Tom: Next up, we look at a different kind of complexity in these areas with a paper on bounded integer programming that might actually be useful for the engineers listening.

More episodes

← Home