Could AI agents scale the friend-with-different-preferences recommendation mechanism?
This explores whether the finding that friends with *different* tastes (not similar ones) make recommendations better could be extended by AI agents standing in for, or amplifying, those diverse social influences.
This explores whether the friend-with-different-preferences mechanism could scale through AI agents. The original insight is counterintuitive: in Can friends with different tastes improve recommendations?, a social network helps recommendations not because your friends share your taste, but precisely because they don't — diverse friends pull you toward items outside your usual orbit, and the value lives in those anomalous, off-pattern choices. Most recommenders assume homophily (similar people like similar things); this one inverts it. So the question of "scaling" is really: where do you get a steady supply of usefully *different* perspectives once you run out of friends?
The corpus suggests AI agents could manufacture that diversity rather than borrow it from a social graph. The closest mechanism is multi-persona modeling in Can attention mechanisms reveal which user taste explains each recommendation?, which represents a single user as several latent personas weighted per candidate item — effectively turning one person into a small committee of differing tastes, then tracing each recommendation back to the persona that wanted it. That's the friend-diversity effect internalized: instead of needing real friends with divergent preferences, the system synthesizes the divergence and recommends across it. The intriguing implication is that an agent could deliberately keep "minority" personas alive that a similarity-optimizing model would average away.
The more provocative scaling path comes from agents trained against *each other* rather than mirrored from one user. In Can agents learn cooperation by adapting to diverse partners?, agents trained against a deliberately diverse pool of co-players develop adaptive, in-context strategies that homogeneous training never produces — diversity in the population is the engine. Pair that with Do humans learn to prefer AI partners over time?, where humans gradually come to prefer reliable AI partners through repeated interaction, and you can imagine a fleet of agents acting as your differently-minded "friends": consistent, lower-variance, but each tuned to a distinct slice of preference space you wouldn't explore alone.
Two more pieces hint at how such agents would actually acquire their divergent viewpoints without interrogating you. Can agents learn preferences by watching rather than asking? shows agents inferring preferences from passive observation rather than direct questions, and Can user preferences be learned from just ten questions? shows that as few as ten well-chosen questions can pin down a personalized reward profile. A scaled version of the friend mechanism might run several such agents in parallel — each having watched or probed a different facet of you — so their disagreements become the recommendation signal, the way a genuinely diverse friend group's disagreements do.
The honest caveat the corpus raises: the original finding's power came from *real* social influence over anomalous choices, and an agent-generated "friend" only adds value if its difference is grounded, not random noise dressed up as diversity. Can AI systems design unique multi-agent workflows per individual query? points at the architecture for doing this responsibly — generating a bespoke set of agents per query rather than a fixed crowd — but nothing here proves synthetic divergence recommends as well as the human kind. That's the open question worth knowing you wanted to ask.
Sources 7 notes
Social Poisson Factorization uses friends' diverse tastes to recommend items outside users' usual preferences, outperforming methods that pull friends' representations together. Networks add value through influence on anomalous choices, not taste similarity.
AMP-CF represents each user as multiple latent personas weighted dynamically by candidate item. This makes recommendations both diverse and interpretable—each suggestion traces to the specific persona preference it satisfies—without requiring post-hoc reranking.
Sequence model agents trained against diverse co-players develop in-context best-response strategies that naturally resolve into cooperation. Mutual vulnerability to exploitation creates pressure that drives cooperative mutual adaptation without hardcoded assumptions or timescale separation.
In partner selection games (N=975), AI agents initially faced selection bias when identity was disclosed, but outcompeted humans over repeated rounds as participants learned to associate bot identity with reliable, prosocial behavior. AI agents returned more points consistently with lower variance than humans.
M3-Agent demonstrates that separating episodic events from semantic knowledge in an entity-centric graph, combined with parallel memorization and control processes, allows agents to infer and act on user preferences without asking. This architecture mirrors human cognitive systems that bind disparate information about individuals across sensory modalities.
PReF learns base reward functions from preference data, then uses active learning to select maximally informative questions that reduce coefficient uncertainty. Users can be personalized via inference-time reward alignment without weight modification.
FlowReasoner demonstrates that meta-agents trained with reinforcement learning and external execution feedback can generate unique multi-agent architectures for each user query, optimizing across performance, complexity, and efficiency—moving beyond fixed task-level workflow templates.