Complete AI Training
Sign inGet my AI kit

Your job's AI kit

Get your AI kit

Tell us who you are and what you do. We show you your kit right away and email you the link: skills, prompts, AI agents, MCP servers and courses for your job.

500+ jobs ready, and we make a kit for any other job. No payment needed to look.

Share

Video course · 9 chapters · 68 min · certificate

Self-Improving AI Agents: Open Problems & Future Directions

A research-level look at self-improving AI agents: recapping evolution, tool use and verifiers, then open problems in test-time scaling, meta-verification, multi-agent fine-tuning, self-proposed curricula, efficient inference and future research directions.

Become a memberAlready a member? Sign in

What you'll learn

  • Recap the self-improvement paradigms for agents
  • Explain the role of verifiers and rewards
  • Describe test-time scaling and diversity
  • Explain meta-verification and multi-agent fine-tuning
  • Understand self-proposed tasks and curriculum learning
  • Identify open problems in AI inference and efficiency

Chapters

9 chapters · 67:42
  1. 2:41 01Recap Members Course recap Self-improvement, evolutionary strategies, end-to-end workflows with tools, and agents as generalized LLMs.
  2. 2:19 02Verifiers Members Orchestration and verifiers Building workflows means orchestrating LLMs and verifiers, since you need a reward.
  3. 2:00 03Scaling Members Test-time scaling and diversity Improving with multiple samples and voting at test time, and using feedback at training time.
  4. 5:20 04Verification Members Meta-verification DeepSeek Math V2's meta-verification: critiquing critiques using multiple agents.
  5. 8:00 05Fine-tuning Members Multi-agent fine-tuning Fine-tuning multiple agents to keep diversity, tested on math with open-source models.
  6. 2:13 06Results Members Iterative proof improvement Proof scores climbing over iterations, and best-of-32 selection.
  7. 18:27 07Curriculum Members Self-proposed tasks and curricula Deciding what data or prompts to train on, including models that propose their own tasks.
  8. 14:00 08Efficiency Members Future inference and efficiency How AI workloads, model sizes and accelerators will evolve.
  9. 12:42 09Research Members Open questions Moving beyond single-turn chat: open research directions for agentic systems.