Week 12: Research Frontiers

Open problems, future directions, and cutting-edge research

Week 12 of 12

Learning Objectives

  • Define embodied agent, generative agent, and world model
  • Explain key open research problems in agent AI
  • Identify research opportunities in specific domains
  • Compare different approaches to agent safety and alignment
  • Assess feasibility and impact of proposed research directions
  • Design a research proposal for advancing agent capabilities

Topics Covered

  • Field evolution timeline (2022-2025)
  • Open research problems (planning, world models, memory)
  • Agent safety challenges (alignment, scalable oversight)
  • World models and embodied agents
  • Generative agents for simulated societies

Resources

Jupyter Notebooks

Open Generative Agents Demo in Colab Generative Agents Demo

Required Readings

PaperAuthorsYearLink
Voyager: An Open-Ended Embodied Agent Wang et al. 2023 arXiv
Generative Agents: Interactive Simulacra Park et al. 2023 arXiv
Constitutional AI: Harmlessness from AI Feedback Bai et al. 2022 arXiv

Reading Guide: Research Frontiers

4-5 hours Generative Agents Voyager Open problems

Exploration of Generative Agents, Voyager, and open research problems

Primary Paper

Generative Agents: Interactive Simulacra of Human Behavior
Park, J. S., O'Brien, J. C., et al. (2023)
arXiv arXiv

Secondary Papers

  • Voyager: Open-Ended Embodied Agent with LLMs - Wang, G., Xie, Y., et al. (2023) arXiv
  • Constitutional AI: Harmlessness from AI Feedback - Bai, Y., Kadavath, S., et al. (2022) arXiv
  • Language Agents: A Critical Perspective - Sumers, T. R., et al. (2024) arXiv

Exercise: Research Proposal

100 Points 8-10 hours Expert

Develop a research proposal for agent advancement

Learning Objectives

  • Create: Design novel agent approach
  • Evaluate: Critically assess current limitations
  • Synthesize: Propose research directions

Tasks

TaskPointsDescription
Literature Review 30 Survey current research
Research Proposal 45 Propose novel contribution
Feasibility Analysis 25 Assess implementation feasibility

Key Concepts

Field Evolution:

  • 2022: Foundation (CoT prompting, RLHF alignment)
  • 2023: Emergence (ReAct, Reflexion, Generative Agents)
  • 2024: Production (Claude Computer Use, GraphRAG)
  • 2025+: World models, embodied agents, long-horizon planning

Open Research Problems:

  • Capability Gaps: Long-horizon planning, world modeling, compositional generalization
  • Safety/Alignment: Scalable oversight, goal stability, corrigibility
  • Infrastructure: Evaluation benchmarks, scalable memory

Generative Agents: Simulated personas maintaining identity, memories, plans, and relationships. Enable computational social science experiments.

World Models: Internal representations of environment dynamics enabling mental simulation before acting.

Final Project Presentations

Students present their research projects on agentic AI topics.

Exercise

Write a research proposal for an open problem in agentic AI:

  1. Identify a meaningful, tractable problem
  2. Review related work and current approaches
  3. Propose a novel approach with clear methodology
  4. Design evaluation methodology and success metrics
  5. Assess feasibility and potential impact

Discussion Questions

  1. What are the most important open problems in agentic AI?
  2. How do we ensure agent safety as capabilities scale?
  3. What governance structures are needed for autonomous agents?
  4. Will agents with persistent world models emerge in the next 5 years?
  5. What role will multi-agent emergence play in future systems?

Final Project Directions

  • Novel agent architecture for a specific domain
  • Improved evaluation methodology for agent systems
  • Safety or alignment technique for autonomous agents
  • Multi-agent coordination mechanism

Additional Resources

Discussion & Questions

Join the Conversation

Have questions about this week's material? Want to discuss concepts with fellow students?



Back to top