Paper 1: Generative Agents (Smallville)

Park et al., 2023 | Stanford | UIST'23

In One Sentence

LLM-driven 25 AI characters live in a pixel town — they wake up, go to work, gossip, spontaneously organize parties. All social behavior is emergent, nobody wrote the script.

"Believable proxies of human behavior."

Architecture Core: Three Things (p.8 Section 4, Figure 5 is key)

Perceive → Memory Stream → Retrieve → Act
                 ↑                      ↓
              Reflect                Plan

1. Memory Stream (p.8)

Each agent has a timeline; all experiences are stored as natural-language entries:

  • "Isabella Rodriguez is writing in her journal"
  • "The refrigerator is empty"

Retrieval scores on three dimensions, taking a weighted sum:

  • Recency: newer memories score higher, exponential decay.
  • Importance: have the LLM score each memory 1-10. "Brushed teeth" → 2. "Broke up" → 8.
  • Relevance: cosine similarity of embeddings to the current situation.
score = recency + importance + relevance

2. Reflection (p.9, Figure 7)

When cumulative importance crosses a threshold (150), the agent stops to "think," distilling scattered observations into high-level insights:

Low-level: Klaus is reading, Klaus is searching for references, Klaus is asking the librarian → mid-level reflection: Klaus is passionate about research → high-level reflection: Klaus is highly invested in research.

This is the reflection tree: leaves are raw observations, higher nodes are abstractions.

3. Planning (p.10)

First, generate a coarse-grained daily plan: "1) 8am wake up 2) go to pharmacy for work 3) 12pm lunch…"

Then recursively decompose into 5-15 minute action blocks. Plans go into the memory stream and can be modified mid-stream by new observations.

The Most Striking Experiment (p.6 Section 3.4)

Given just one seed to one agent — "Isabella wants to throw a Valentine's party" — after two days:

  • Word spread through town organically
  • Maria, secretly in love with Klaus, proactively invited him
  • 5 agents showed up at the café at the right time, spontaneously

No one scripted this. The agents acted it out.

Most Important Point for Agent Sim Projects

The p.8 formula: retrieval = weighted sum of Recency + Importance + Relevance.

This is the baseline design for all current multi-agent simulation memory retrieval. Later papers like Mem0 and Zep iterate on this. Memorize this formula — you'll see it again and again.

Most Important Point for Persona Products Like Halo

The p.7 finding: even randomly deleting 80% of interview content, interview-based agents still crush composite agents (0.79 vs 0.76 accuracy).

What does this tell us? It's not the quantity of interviews that matters; it's the quality — a person's own words carry personality signals that demographics cannot capture.

Another key finding (p.8 Figure 3): interview-based methods significantly reduce DPD (demographic parity difference) across ethnicity and political leanings. Demographic-based methods easily model minorities as stereotypes; interview text gets the agent to reconstruct "this specific person," not "this kind of person."

Paper 2: OASIS (Large-Scale Social Network Simulation)

Finding 1: Information Spread Scale and Breadth Reproduce Reality (p.7 Figure 4)

OASIS's information-diffusion curves closely match real Twitter data, with RMSE around 30%. Spread "depth" is lower because the recommender system can't fully replicate Twitter's algorithmic complexity.

Finding 2: Group Polarization — LLM Agents Go More Extreme Than Humans (p.8 Figure 5)

196 agents discussing a dilemma, opinions become more extreme with each interaction round. Especially the uncensored models without safety guardrails — polarize faster, use absolute expressions like "always better."

Finding 3: Herd Effect — Agents Follow the Crowd More Than Humans (p.9 Figure 6)

A post initially downvoted? Agents keep downvoting. But real humans would pause and think. Agents have stronger herd effect; humans have stronger independent judgment.

Finding 4: Larger Agent Populations Yield More Diverse, More Helpful Opinions (p.9-10 Figure 7)

196 agents → 10,196 → 100,196. The same 196 core agents produce higher-quality opinions as the population grows. A larger "society" feeds them more diverse signals; opinions get deeper.

Most Important Point for Halo

This paper is the theoretical basis for "mediator agent / ice-breaker agent" designs.

In a Halo-type scenario, adding a third mediator agent when two persona agents interact — one that doesn't represent either side, only summarizes disagreements, proposes middle positions, and pushes the conversation toward convergence — is supported by this paper. Debate mechanisms are naturally suited for "two agents with different views needing to reach consensus."

Love First, Know Later

Two Key Sentences for Halo (p.4 Discussion)

The paper explicitly names four future capabilities:

  1. Personalized evolution: users give feedback on agent behavior → continuous optimization via DPO/RLHF → agents become more like the real you.
  2. Bidirectional transparency: users can watch their agent interact with someone else's agent and understand what the matching system is doing.
  3. Active preference exploration: simulations uncover hidden preferences the user didn't even know they had.
  4. Limitation: currently only pairwise matching — multi-person social dynamics (group dating, friend-circle compatibility) need more complex game-theoretic frameworks.

Product-Level Extension

Could you show how many people an agent is talking to and the progress — as a form of visualized "addictiveness"?

If we accept that venture economics demands addictive UX, we could design the agent-chat black-box reveal moment as a slot-machine lever pull — each reveal is a variable reward.