Skip to content
Back to full roadmap
topicadvanced

Self-Discover (Google)

Model discovers its own reasoning structure, then applies it. More flexible than CoT.

2 hours1 resources

Self-Discover (Google DeepMind, 2024): model selects the right reasoning module for a task (decomposition, CoT, analogy, etc.), then applies it.

3 stages: SELECT relevant reasoning modules → ADAPT to task → IMPLEMENT step-by-step structured plan.

32% better than CoT on complex reasoning, cheaper than ToT.

Resources(1)