💬 NLP

Consensus vs. Dissent: Dynamic LLM Modeling of Subjective Preferences in Group Recommenders

This paper demonstrates that fine-tuned Large Language Models can effectively mimic human sensitivity to group preference distributions to dynamically select optimal aggregation strategies, thereby maximizing satisfaction, fairness, and consensus in group recommender systems.

Original authors: Cedric Waterschoot, Nava Tintarev, Francesco Barile

Published 2026-07-14

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Cedric Waterschoot, Nava Tintarev, Francesco Barile

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you and four friends are trying to decide where to eat dinner. You all have different tastes: one person loves spicy food, another hates it, and a third just wants the cheapest option. In the world of computer science, this is called a Group Recommender System (GRS). For a long time, computers have tried to solve this by using rigid math formulas. Some formulas say, "Let's just average everyone's vote." Others say, "Let's pick the place that makes the least unhappy person happy." These formulas are like a robot chef who follows a recipe perfectly but has no idea if the food actually tastes good to the specific group sitting at the table.

The problem is that these robots can't "feel" the group's mood. They can't tell if a group is mostly in agreement (a consensus) or if they are split into two fighting factions (a coalition). To fix this, researchers Cedric Waterschoot, Nava Tintarev, and Francesco Barile decided to teach a computer to act like a human judge.

The Robot Judge Gets a Brain Upgrade

The team started with a "teacher" robot, a massive AI called DeepSeek-V3.1, which is so smart it can generate diverse reasoning. They fed this teacher a huge pile of data from a previous study where 1,152 real humans had rated group recommendations on how fair, satisfying, and agreed-upon they felt.

Instead of just memorizing the scores, the researchers asked the teacher to explain why it gave those scores. It was like asking a teacher to write a detailed essay on why a student got an A, rather than just writing "A" on the paper. They used the teacher to generate 5,318 examples of synthetic reasoning based on the existing human scores. This created a special training set where the AI learned to mimic the logic behind human opinions, not just the opinions themselves.

Then, they took two smaller, open-source robots (Llama-3.1 and OLMo-3) and taught them using these essays. These new robots, named Judgmental Llama and Judgmental Olmo, learned to mimic human feelings. Before this training, the robots were terrible judges; they gave very narrow, boring answers that didn't match real human opinions. After the training, they became much better at guessing how a group would feel, matching the spread of human opinions almost perfectly.

The Big Test: Can the Robot Pick the Best Strategy?

The researchers built a system where, for every group of friends, the computer didn't just pick one recipe. Instead, it ran six different math formulas (like "Average," "Least Misery," and "Majority Vote") to generate six different dinner suggestions.

Then, the Judgmental Olmo robot (which turned out to be the best student) stepped in. It looked at all six suggestions and asked itself: "If I were a human in this specific group, which one would I think is the fairest and most satisfying?" It picked the winner based on its human-like feelings.

To see if this actually worked, the team ran a real-life experiment with 284 human participants. They showed these people group scenarios (like a group of five friends with different tastes) and asked them to rate the recommendations.

The Results: The Robot Judge Wins (Sometimes)

The results were exciting but specific. The system that used the Judgmental Olmo to dynamically pick the best strategy scored the highest for perceived fairness and group consensus. It also scored the second-highest for satisfaction, just behind one specific math formula called "Fairness" (FAI).

However, the paper makes it very clear that the robot isn't a magic wand that works the same way every time. The robot's success depended heavily on the type of group.

For groups that were all similar (uniform): The robot knew to pick strategies that made everyone happy.
For groups that were split (coalitional) or had a minority: The robot realized that the usual "average" math formulas often failed. It switched gears and picked different strategies that handled the conflict better.

The data showed a statistically significant interaction between the group type and the strategy used. In other words, the robot's ability to adapt was what made it better than the old, static methods. For example, in "coalitional" groups (where two subgroups disagree), the robot's choice was perceived as significantly fairer than the "Majority Vote" or "Approval Voting" strategies.

What the Robot Can't Do (Yet)

The paper is careful to point out what this system isn't yet.

It's not a breakthrough for every problem: The researchers explicitly state that their system is currently limited to specific scenarios (like choosing restaurants) and relies on data from previous studies. They don't claim it works for any group recommendation problem in the world yet.
It's not instant: Because the robot has to run multiple "thought experiments" (generating five assessments per recommendation) to be sure, it takes time. In a real-time app, this could mean a delay of several minutes, which is too slow for a hungry group of friends.
It's not perfect: The base robots (before training) were biased and narrow. The training helped, but the researchers admit that relying on a "teacher" model to generate the training data might introduce a dependency. They suggest that future work needs to train on actual human reasoning, not just AI-simulated reasoning.

The Bottom Line

This paper suggests that we can build better group recommendation systems by teaching AI to understand the nuance of human feelings, rather than just crunching numbers. By letting a trained AI judge pick the best math formula for the specific group it's dealing with, we get recommendations that feel fairer and more agreed-upon.

But, the authors warn us not to get too excited just yet. The system is still a simulation and a prototype. It needs to get faster (by using smaller, "tiny" robots) and needs to be tested on more types of groups before we can trust it to pick our dinner without us. For now, it's a very promising step toward computers that understand that sometimes, the best math answer isn't the one that makes the most people happy, but the one that makes the right people happy.

Technical Summary: Consensus vs. Dissent: Dynamic LLM Modeling of Subjective Preferences in Group Recommenders

Problem Statement
Group Recommender Systems (GRS) face the inherent challenge of reconciling conflicting individual preferences into a single group outcome. Traditional approaches rely on static, social choice-based aggregation strategies (e.g., Average, Least Misery, Most Pleasure) to bridge individual ratings and group outcomes. While these methods are computationally efficient and transparent, they lack adaptability to nuanced group dynamics, such as the distribution of preferences (e.g., uniform consensus vs. coalitional dissent). A significant gap exists between the evaluation of these systems—where human studies assess perceived fairness, satisfaction, and consensus—and their implementation. Current systems lack a scalable mechanism to predict how a specific group will feel about a recommendation in real-time, preventing the dynamic selection of the optimal aggregation strategy based on subjective human metrics.

Methodology
The authors propose a pipeline that integrates fine-tuned Large Language Models (LLMs) as "judges" to dynamically select the best aggregation strategy for a given group scenario. The methodology consists of three primary phases:

Data Preparation and Knowledge Distillation:
- Source: The process begins with a human ground truth dataset ( $n=1,152$ assessments) from previous user studies (Barile [4], Barile et al. [5]), where participants rated group recommendations on fairness, satisfaction, and consensus using a 7-point Likert scale.
- Synthetic Expansion: To address data scarcity for training, the authors employed Knowledge Distillation. A high-capacity teacher model (DeepSeek-V3.1) was used to generate synthetic Chain-of-Thought (CoT) reasoning for the existing human ground truth. This expanded the training set from 890 instances to 5,318 samples, capturing diverse justifications for identical scores while preserving the original human scores.
- Fine-Tuning: Two open-source models, Llama-3.1-8B-Instruct and OLMo-3-7B-Instruct-SFT, were fine-tuned on this augmented dataset using Parameter-Efficient Fine-Tuning (LoRA). A custom label masking function was applied to ensure the models learned to replicate the reasoning process and scores rather than the prompt structure.
LLM Evaluation:
- The fine-tuned models ("Judgmental Llama" and "Judgmental Olmo") were evaluated against the base models and human ground truth using the Wasserstein Distance (Earth Mover's Distance) to measure distributional alignment.
- Judgmental Olmo emerged as the superior model, achieving the lowest Wasserstein Distance across all three metrics (fairness, satisfaction, consensus) and replicating the variance ( $\sigma > 1.55$ ) found in human responses.
Dynamic GRS Pipeline:
- The pipeline generates recommendations for a group using six distinct social choice-based strategies: Additive Utilitarian (ADD), Approval Voting (APP), Fairness (FAI), Least Misery (LMS), Majority (MAJ), and Most Pleasure (MPL).
- For each potential outcome, the fine-tuned LLM generates multiple assessments (simulating a committee of five evaluators) to predict fairness, satisfaction, and consensus.
- The system averages these scores and dynamically selects the aggregation strategy that maximizes the predicted human-like evaluation for that specific group configuration.
User Study Validation:
- A user study ( $n=284$ ) was conducted to compare the dynamic LLM-based selection against the six static strategies.
- The study utilized a mixed-design ANOVA with group configuration (uniform, divergent, minority, coalitional) as a within-subject factor and aggregation strategy as a between-subject factor.

Key Contributions

Distilled Reasoning Dataset: Creation of a dataset grounded in human assessments of fairness, satisfaction, and consensus, augmented with synthetic reasoning generated by DeepSeek-V3.1.
Judgmental Models: Introduction of "Judgmental Llama" and "Judgmental Olmo," models specifically fine-tuned to simulate group assessments and align with human ground truth distributions.
Dynamic Pipeline: Development of an end-to-end GRS pipeline that employs these models to dynamically select aggregation strategies based on predicted human evaluations rather than static rules.
Empirical Validation: A user study demonstrating that the dynamic approach outperforms static baselines in perceived fairness and consensus, with significant interactions observed based on group configuration.

Results

Model Alignment: Fine-tuning significantly reduced the Wasserstein Distance between model outputs and human ground truth. Judgmental Olmo achieved the best alignment (WD: 0.70 for fairness, 0.68 for satisfaction, 0.82 for consensus) compared to base models, which exhibited high distances and narrow output distributions.
User Study Performance: The LLM-based dynamic selection achieved the highest mean scores for perceived fairness (0.82) and perceived consensus (0.81), and the second-highest for satisfaction (0.68), outperforming all static strategies (ADD, APP, FAI, LMS, MAJ, MPL).
Interaction Effects: Statistical analysis revealed significant interaction effects between group configuration and aggregation strategy ( $p < .001$ $p < .001$ ). The dynamic method was particularly effective in complex scenarios:
- In coalitional groups, the LLM significantly outperformed APP and ADD strategies regarding fairness and consensus.
- In minority groups, the LLM selection provided significantly higher satisfaction than MAJ and MPL strategies.
- The system correctly identified that "Most Pleasure" (MPL) was optimal only for uniform groups, while other strategies were better suited for divergent or coalitional groups.

Significance and Claims
The paper claims that its methodology successfully bridges the gap between static mathematical aggregation and the nuanced, subjective perceptions of human groups. By leveraging LLMs fine-tuned on human ground truth, the system can adapt aggregation strategies to specific within-group preference distributions (e.g., handling dissent in coalitional groups vs. consensus in uniform groups).

The authors assert that this approach offers a scalable, real-time alternative to traditional user studies for evaluating group recommendations. The findings support the hypothesis that dynamically adapting aggregation strategies based on LLM-generated human-like judgments leads to recommendations that are perceived as fairer and more consensual than those produced by fixed, static strategies. The work highlights that quality fine-tuning data is essential for aligning LLMs with human judgment, as base models alone fail to capture the necessary nuance and diversity of human responses.