LATS (Tree/Beam Search Over Solutions)

What Problem It Solves

If “reasoning” is a search space, you can improve results via:

• expand candidate solutions
• score them
• keep the best and iterate

Core Flow

flowchart TD
  S["Seed"] --> E["Expand candidates"]
  E --> Sc["Score candidates"]
  Sc --> K["Keep top-K (beam)"]
  K -->|repeat| E
  K --> O["Best"]

Evolution Path

• Complements: Plan-based methods (Plan & Solve, PER)
• Often combined with: strong evaluator (rubrics, unit tests, tools)

Repo Reference

• Code: src/agent_patterns_lab/patterns/lats.py
• Example: examples/54_lats.py
• Tests: tests/test_lats.py