ProPerSim: Personalized Simulation Benchmark

• ProPerSim is a simulation framework that models user–assistant interactions as sequential decision tasks in realistic home environments.
• Its ProPerAssistant architecture uses retrieval-augmented generation and direct preference optimization to refine recommendation timing and content.
• The benchmark shows significant, persona-specific improvements in key metrics like frequency, timing, and communication safety over a 14-day simulation.

ProPerSim refers to a simulation framework and benchmark for developing AI assistants that are simultaneously proactive and personalized in their interactions with users. The framework formalizes the user–assistant interaction as a sequential decision task in realistic home scenarios, leveraging rich simulated personas and feedback-driven adaptation to optimize assistant behavior. ProPerSim forms the foundation for the ProPerAssistant architecture, which integrates retrieval-augmented memory and explicit preference learning in a loop of continual online adaptation (Kim et al., 26 Sep 2025).

1. Formal Task Definition and Simulation Structure

ProPerSim models the assistant–user dyad as a time-sequenced decision process situated in a simulated environment $E$ (a multi-room house with objects, sensors, and spatial layouts). The user's day is represented as a set of actions

$\{(A_i,\Range_i)\}_{i=1}^N \;=\; \mathcal{U}(E,P,S),$

where $A_i$ is an atomic user action, $\Range_i$ its duration, $P$ encodes the user's persona with Big Five traits and additional attributes, and $S$ encompasses the user’s internal state, including memory, plans, and affective context.

At each discrete time step $T$ (default: 2.5 min), the assistant policy $\mathcal{A}_\theta$ observes the current user activity $A_t$ and its internal memory $S^{(a)}_t$. It must choose—out of a candidate set (including “No Recommendation”)—a recommendation $\{(A_i,\Range_i)\}_{i=1}^N \;=\; \mathcal{U}(E,P,S),$0, which is immediately scored by a rich, persona-aligned rubric: $\{(A_i,\Range_i)\}_{i=1}^N \;=\; \mathcal{U}(E,P,S),$1 The simulation objective is to maximize cumulative expected score, integrating both the content and timing of interventions.

2. User Agent Modeling and Persona Diversity

Each user agent in ProPerSim is parameterized by a persona vector $\{(A_i,\Range_i)\}_{i=1}^N \;=\; \mathcal{U}(E,P,S),$2 comprising:

• Binary values for the Big Five: Extraversion, Agreeableness, Openness, Conscientiousness, Neuroticism.
• Six free-text slots: age, background, interests, lifestyle, daily-plan constraints, long-term goals.

Persona generation is constrained to span diverse regions of trait–lifestyle space, with 32 concrete exemplars validated for coherence and coverage. The user agent executes an adaptive, hourly schedule synthesized by GPT-4o, subject to dynamic replanning. Feedback is structured via a four-dimensional rubric distilled from an initial ten metrics: Personal Preference, Frequency, Timing, Communication & Safety. For each action, the agent, leveraging both immediate and compressed long-term memory ($\{(A_i,\Range_i)\}_{i=1}^N \;=\; \mathcal{U}(E,P,S),$3), assigns 0–4 points reflecting alignment with requirements.

3. Recommendation Architecture and Preference Alignment

ProPerAssistant, the canonical instantiation within ProPerSim, integrates two central mechanisms:

• Retrieval-Augmented Generation: The assistant maintains a structured memory of recent $\{(A_i,\Range_i)\}_{i=1}^N \;=\; \mathcal{U}(E,P,S),$4 tuples, with segmented short-range (last 10 min) and longer-range (up to several hours) memory. At each step, the prompt to the model includes the 5 most semantically similar past instances, identified via embedding-based retrieval, as well as summaries generated with GPT-4o-mini.
• Direct Preference Optimization (DPO): Candidate recommendations are scored by the user agent rubric at every interval. Preference pairs are extracted and form the basis of per-session fine-tuning via DPO: $\{(A_i,\Range_i)\}_{i=1}^N \;=\; \mathcal{U}(E,P,S),$5 where $\{(A_i,\Range_i)\}_{i=1}^N \;=\; \mathcal{U}(E,P,S),$6 is the logistic sigmoid and $\{(A_i,\Range_i)\}_{i=1}^N \;=\; \mathcal{U}(E,P,S),$7 is the buffer of logged preferences. Daily, 200 buffer samples are randomly selected for each QLoRA fine-tuning step on a LLaMA-3.3-70B backbone quantized to 4 bits.

This adaptation process enables the assistant to refine both “when” (timing/frequency) and “what” (personalized content) aspects of its recommendations over a simulated 14-day horizon.

4. Evaluation Metrics and Baseline Comparisons

Performance in ProPerSim is tracked by the daily mean recommendation score,

$\{(A_i,\Range_i)\}_{i=1}^N \;=\; \mathcal{U}(E,P,S),$8

and disaggregated along the four core rubric criteria. Three baselines are considered:

• No Memory: The assistant sees only current $\{(A_i,\Range_i)\}_{i=1}^N \;=\; \mathcal{U}(E,P,S),$9.
• AR Memory: The policy prompt includes previous $A_i$0, but model parameters are static.
• ARS Memory: Previous $A_i$1 are included; still no parameter adaptation.

ProPerAssistant exhibits a notable trajectory from 2.2/4 average initial score (Day 1) to 3.3/4 after 14 simulation days, surpassing ARS Memory (plateauing at 2.6), AR Memory (2.3), and No Memory (2.1). Recommendation rates dynamically adapt downward, from ~24 to ~6 per hour in response to user feedback, reflecting learned proactivity moderation and alignment to individual preference for suggestion frequency (Kim et al., 26 Sep 2025).

5. Adaptation Dynamics and Persona-Specific Outcomes

Persona-level analyses reveal that the assistant’s adaptive gains are heterogeneous:

• Users low in Extraversion or Neuroticism experienced the most substantial improvements, indicating robust adaptation to more routine or less impulsive schedules.
• High-Openness personas saw smaller gains, a plausible implication being that current rubrics do not incentivize novelty or diversity in suggestions.
• The most challenging persona—requiring precisely timed, multi-criteria recommendations—remained suboptimal, suggesting a need for more sophisticated, context-hierarchical adaptation frameworks.

On the rubric dimension level, improvements were particularly marked in Frequency, Timing, and Communication & Safety, echoing the system’s ability to moderate intrusiveness and improve interaction competence.

6. Limitations and Forward Directions

Several structural and methodological limitations are identified:

• The simulated environment presupposes noiseless, immediate feedback, whereas real-world deployment would need robustness to delayed, incomplete, or privacy-constrained signals.
• The DPO learning regime may be sensitive to batch size and replay buffer construction; further hyperparameter optimization and potentially hybrid RL frameworks (e.g., actor–critic) are suggested for improving adaptation granularity.
• Handling of complex, multi-dimensional persona constraints may require hierarchical memory organization or explicit novelty-seeking objectives.

Extensions under consideration include multi-user households, temporal field evolution, memory privacy, and integration with actor–critic reinforcement learning for finer control of recommendation timing and content.

7. Significance and Benchmark Composition

ProPerSim, as presented by Sileo et al. (Kim et al., 26 Sep 2025), constitutes a fully specified benchmark and simulation infrastructure for next-generation AI assistant research. Its principal contributions are:

• A formalized, replicable environment for testing recommendation policies with explicit, articulated user personas.
• A robust feedback loop connecting recommendation quality, user context, and model adaptation through a preference-aligned learning mechanism.
• Detailed rubrics and evaluation metrics capturing both proactivity and personalized alignment, enabling rigorous ablation and comparative studies.

This suggests ProPerSim is positioned as a foundational platform for systematic research on proactive and personalized assistants, as well as a proving ground for adaptive policy architectures in simulated, detailed user-centric environments.