Simulating an online community at scale calls for language models that can stand in for many distinct users at once, not collapse them into a single generic voice. Prior work shows that a hypernetwork-driven language model can efficiently rebuild known users by storing a small per-user profile vector rather than separate fine-tuned models, one per user. This study explores whether those profile vectors support a harder task: generating behaviorally consistent new users who were never directly observed. Our frozen LLM uses adapter offsets emitted by a hypernetwork. We sample profiles within and beyond the real-user range, then score the generated replies on behavioral measures to determine whether synthetic users match the cohort patterns their profiles imply. We show that synthetic users reproduce behavioral patterns similar to those of real users in the training set: in short threads they stay close to their target profile, and across a longer multi-turn forum simulation they stay largely coherent and do not drift from it. Performance declines as profiles move farther from the distribution of users seen in the original data, suggesting that the method works best for synthetic users that recombine observed patterns rather than extrapolate beyond them.

Designing a generalist scientific agent capable of performing tasks in laboratory settings to assist researchers has become a key goal in recent Artificial Intelligence (AI) research. Unlike everyday tasks, scientific tasks are inherently more delicate and complex, requiring agents to possess a higher level of reasoning ability, structured and temporal understanding of their environment, and a strong emphasis on safety. Existing approaches often fail to address these multifaceted requirements. To tackle these challenges, we present DAVIS. Unlike traditional retrieval-augmented generation (RAG) approaches, DAVIS incorporates structured and temporal memory, which enables model-based planning. Additionally, DAVIS implements an agentic, multi-turn retrieval system, similar to a human’s inner monologue, allowing for a greater degree of reasoning over past experiences. DAVIS demonstrates substantially improved performance on the ScienceWorld benchmark comparing to previous approaches on 8 out of 9 elementary science subjects. In addition, DAVIS’s World Model demonstrates competitive performance on the famous HotpotQA and MusiqueQA dataset for multi-hop question answering. To the best of our knowledge, DAVIS is the first RAG agent to employ an interactive retrieval method in a RAG pipeline.