Reasoning Pattern Selection

Decision Flow

Need token efficiency above all?
  → R5 (ReWOO): 5× reduction vs ReAct; plan all tool calls upfront

Need mid-run adaptation to observations?
  → R4 (ReAct): adaptive tool use; each action informs the next

Multi-tool task needing self-debugging?
  → R13 (CodeAct): ~20pp accuracy gain over JSON tool calls

Hard open-ended problem, quality trumps cost?
  → R9 (Tree of Thoughts) or R10 (LATS)

Clear pass/fail criteria and retries are acceptable?
  → R7 (Reflexion): verbal self-critique across retries

Math or numerical computation?
  → R14 (Program of Thoughts): delegate to a deterministic executor

Parallel generation needed to reduce latency?
  → R12 (Skeleton-of-Thought): outline first, fill sections in parallel

Reusable reasoning templates exist for this task type?
  → R11 (Buffer of Thoughts): 12% cost of ToT/GoT

Multi-hop factual question?
  → R6 (Self-Ask): sub-question chains

Quick reasoning improvement with no examples?
  → R1 (Zero-Shot CoT): "think step by step"

Cost Guide

PatternLLM CallsRelative CostNotes
R1 Zero-Shot CoT1BaselineAdd "think step by step" only
R2 Few-Shot CoT1Low + example tokensStatic examples cache cleanly
R3 Plan-and-Solve2LowPlan + execute; two clean calls
R4 ReActN per stepMedium–HighScales with task complexity
R5 ReWOO2 total5$\times$ cheaper than R4All tool calls must be independent
R6 Self-Ask1 + N follow-upsMediumSub-question depth drives cost
R7 ReflexionN $\times$ retriesHighNeeds measurable success criterion
R8 Self-RefineN iterationsMediumIn-session; no separate judge
R9 ToTN (branching)Very HighUse when path genuinely unknown
R10 LATSN (tree search)HighestHighest quality; highest cost
R11 BoT1 + templateLowTemplates amortise across calls
R12 SoT1 + N parallelMediumLatency win via parallelism
R13 CodeActN (with execution)MediumSelf-debugging loop
R14 PoT1 + executionLowDeterministic computation free