合意と反対意見：グループ推薦における主観的嗜好の動的LLMモデリングConsensus vs. Dissent: Dynamic LLM Modeling of Subjective Preferences in Group Recommenders

Group recommender systems, which suggest a single item such as a restaurant, film, or travel itinerary to several people at once, face a problem that individual recommendation does not: the members rarely want the same thing. A new paper posted to arXiv, "Consensus and Dissent: Dynamic LLM Modeling of Subjective Preferences in Group Recommendation," proposes using large language models to represent both where group members agree and where they diverge, rather than collapsing their tastes into a single averaged signal. The work matters because group settings are common in practice yet remain harder to serve well than the individual case, and small improvements in how disagreement is handled can meaningfully change collective satisfaction.

The core issue the authors address is aggregation. Classic group recommendation relies on strategies borrowed from social choice, such as taking the average of members' predicted ratings, applying a "least misery" rule that prioritizes the least satisfied person, or a "most pleasure" rule that favors the most enthusiastic one. These heuristics are simple and interpretable, but they treat preferences as fixed numeric scores and tend to erase the nuance of why someone likes or dislikes an option. A member who is mildly negative about a horror film for one reason is treated the same as another who objects for a completely different reason, and the group's internal structure of agreement and conflict is lost.

According to the summary, the proposed framework instead uses an LLM to model members' subjective preferences dynamically, capturing consensus and dissent as distinct, evolving signals rather than a single aggregated value. The stated goal is to improve collective satisfaction beyond what simple preference aggregation achieves. The dynamic aspect appears intended to let the representation of a member's stance shift with context, so that the same person may weigh differently depending on the candidate item, the composition of the group, or the trade-offs at stake. By reasoning over natural-language descriptions of tastes, the model can, in principle, distinguish strong objections that should block a choice from mild preferences that can be accommodated.

This line of work sits at the intersection of two active research areas. The first is recommender systems built on collaborative filtering, the long-standing technique of predicting a user's interest by learning from patterns across many users and items, often through matrix factorization or, more recently, neural and graph-based models. The second is the rapid adoption of LLMs as reasoning components inside recommendation pipelines. Over the past few years, researchers have explored using models such as those in the GPT and open-weight LLaMA families to generate explanations, encode item and user profiles as text, perform zero-shot or few-shot ranking, and act as flexible scoring functions. Framing preferences in language rather than as opaque vectors is one of the main attractions, because it can make the resulting recommendations more explainable and easier to inspect.

The consensus-and-dissent framing also connects to the older idea in group recommendation that fairness and disagreement deserve explicit treatment. Prior systems have modeled group dynamics, member influence, and satisfaction balance, sometimes using attention mechanisms to weight members differently. What this paper appears to add is the use of an LLM's language understanding to represent those subjective and conflicting positions more richly, and to update them dynamically instead of fixing them in advance.

Several caveats are worth keeping in mind. As an arXiv preprint, the work may not yet have completed peer review, and the summary does not specify the datasets, baselines, or metrics used, so the magnitude of any improvement is unclear from the available description. LLM-based recommenders also carry known practical concerns, including inference cost and latency at scale, sensitivity to prompt design, and the risk of hallucinated or inconsistent preferences, all of which are relevant when decisions must be made quickly for many groups. Readers interested in the specifics should consult the full paper for its evaluation setup and ablations. Even so, the direction is a reasonable reflection of a broader industry trend toward integrating language models into recommendation systems, and its explicit focus on modeling dissent rather than smoothing it away is a notable framing.