Research map / Multi-agent systems and games / Human-AI and goal-conditioned
Simulating Human-like Daily Activities with Desire-driven Autonomy
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes D2A, an LLM agent that autonomously proposes and selects daily activities to satisfy multi-dimensional desire values grounded in the Theory of Needs, simulated in Concordia.
Problem
Current LLM agents need explicit task specifications or external rewards, which limits autonomy and behavioral diversity; the paper targets generating human-like daily activities driven by intrinsic desires.
Approach
A Value System tracks numerical desire dimensions (hunger, social connectivity, spiritual satisfaction, etc.) with decay and target values derived from profile traits; a Desire-driven Planner proposes N candidate activities, forecasts resulting desire states for each, and selects the one best satisfying desires.
Key results
- In pairwise GPT-4o win-rate comparisons, D2A is judged more human-like than ReAct, BabyAGI, and LLMob in indoor and outdoor settings.
- In the Random-8-Steps experiment, D2A shows significantly lower dissatisfaction means and standard deviations than the three baselines.
- GPT-4o agreed with human annotators on 100% of high-consensus pairs, 80% of moderate-preference pairs, and 66.7% of hard-to-distinguish pairs.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use desire-driven activity generation instead of ReAct, BabyAGI, or LLMob | D2A · Indoor daily activity sequences | Human-likeness (GPT-4o win rate) | Helped |
| Use desire-driven activity generation instead of the baseline agents | D2A · Outdoor environment | Human-likeness (GPT-4o win rate) | Helped |
| Use desire-driven activity selection rather than baseline activity-generation mechanisms | D2A · Random-8-Steps, indoor | Dissatisfaction | Helped |
| Use desire-driven activity selection | D2A · Fixed-12-Steps, indoor | Dissatisfaction relative to human-generated sequences | Mixed |
| Use desire-driven activity selection | D2A · Outdoor environment | Dissatisfaction | Helped |
| Remove qualitative value descriptions (ablation) | D2A · Desire-driven activity simulation | Rate of desire fulfillment | Hurt |
| Reduce planner width from 3 to 1 (D2A(W-1)) (ablation) | D2A · Desire-driven activity simulation | Dissatisfaction | Hurt |
| Increase planner width from 3 to 5 (D2A(W-5)) (ablation) | D2A · Desire-driven activity simulation | Dissatisfaction | Helped |
| Add Concordia's multi-step planning component (ablation) | D2A · Desire-driven activity simulation | Responsiveness to immediate desire states | Hurt |
Surprising findings
- Adding Concordia's multi-step planning component slightly reduced D2A's performance in desire satisfaction, which the authors attribute to rigid pre-defined hourly plans.
Limitations
- Desire dynamics are modeled as simple linear decay, which may miss organic variation in human needs.
- Desire dimensions are heuristically chosen based on environment features, limiting coverage of complex motivations.
- Evaluation relies on simulated environments with a single backbone for main results and GPT-4o as judge; human-likeness is judged pairwise rather than absolutely.
Experimental setup
Datasets: Concordia environments (indoor house, Central Park party)
Benchmarks: Concordia indoor environment Concordia outdoor Central Park party environment
Metrics: pairwise win rate (likelihood) dissatisfaction over time
Baselines: ReAct BabyAGI LLMob Human-controlled agent
Builds on: Concordia Tree of Thoughts ReAct BabyAGI LLMob
Closest papers by meaning
- Promptable Behaviors: Personalizing Multi-Objective Rewards from Human Preferences (CVPR 2024 · similarity 0.86)
- Scaling Autonomous Agents via Automatic Reward Modeling And Planning (ICLR 2025 · similarity 0.85)
- Motif: Intrinsic Motivation from Artificial Intelligence Feedback (ICLR 2024 · similarity 0.85)
- AutoManual: Constructing Instruction Manuals by LLM Agents via Interactive Environmental Learning (NeurIPS 2024 · similarity 0.85)
- Ground-Compose-Reinforce: Grounding Language in Agentic Behaviours using Limited Data (NeurIPS 2025 · similarity 0.84)
- Proactive Agent: Shifting LLM Agents from Reactive Responses to Active Assistance (ICLR 2025 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.