The Reasoning Sandwich is an optimization technique for large language model (LLM) task execution that strategically allocates computational reasoning resources across different phases of problem-solving. Rather than applying uniform reasoning effort throughout a task, the approach concentrates intensive reasoning during planning and verification stages while reducing reasoning complexity during intermediate code generation and execution phases. This selective resource allocation pattern reflects a sandwich-like structure: high reasoning → lower reasoning → high reasoning.

The Reasoning Sandwich technique addresses a fundamental challenge in LLM-based systems: reasoning tokens and computational cycles are expensive resources that should be allocated optimally based on task requirements. The method recognizes that different phases of complex task execution have varying reasoning demands. Planning phases benefit from intensive reasoning to develop robust strategies, while generation phases may benefit more from focused execution than from continuous deep reasoning. Verification and validation phases similarly demand high reasoning capacity to catch errors and validate solutions ¹⁾.

This approach contrasts with two alternative strategies: applying maximum reasoning effort uniformly throughout task execution, or relying on standard LLM approaches without reasoning optimization. The technique demonstrates that reasoning capacity allocation represents a critical optimization variable in complex task execution pipelines.

Benchmark testing conducted with LangChain's implementations provides quantitative evidence for the Reasoning Sandwich approach. The technique achieved 66.5% accuracy on Terminal Bench 2.0, a comprehensive benchmark for evaluating LLM performance on terminal-based tasks and code generation scenarios. This performance exceeded both alternative strategies: maxed-out reasoning across all phases achieved 53.9%, while standard approaches without strategic reasoning allocation produced lower scores ²⁾.

The 12.6 percentage point improvement over maxed-out reasoning (66.5% versus 53.9%) suggests that indiscriminate reasoning allocation may introduce inefficiencies, including increased latency, higher computational costs, and potentially degraded task performance through reasoning noise during generation phases where focused execution proves more effective.

The Reasoning Sandwich framework decomposes complex tasks into three primary phases:

Planning Phase: High reasoning effort focuses on understanding task requirements, decomposing problems into subtasks, identifying dependencies, and developing execution strategies. This phase benefits substantially from deep reasoning, as poor planning propagates errors throughout subsequent execution.

Code Generation and Building Phase: Reduced reasoning effort emphasizes efficient execution following established plans. This phase prioritizes adherence to planned approaches and rapid generation over continuous reasoning about alternatives, as ongoing reasoning may introduce inconsistencies with the planning phase decisions.

Verification and Validation Phase: High reasoning effort returns to validate generated solutions, identify errors, reason about edge cases, and confirm correctness. This phase requires intensive reasoning to catch logical flaws and ensure solution robustness before final output.

Implementing the Reasoning Sandwich technique requires controlling reasoning allocation parameters within LLM inference systems. Modern reasoning-enabled models support inference-time reasoning budgets measured in reasoning tokens or computational cycles. Practitioners must determine appropriate reasoning intensity levels for planning and verification phases, configure reduced reasoning baselines for generation phases, and establish phase transition boundaries.

The approach integrates with broader harness engineering practices, which encompass structured prompt engineering, tool integration, error handling, and execution monitoring. The Reasoning Sandwich complements other optimization techniques by providing a resource allocation framework specifically designed for reasoning-augmented systems.

The superior performance of the Reasoning Sandwich over uniform maximum reasoning suggests that task complexity and resource constraints should influence reasoning allocation strategies. Organizations deploying reasoning-capable models can optimize inference costs by reducing reasoning during generation phases while maintaining high reasoning capacity during planning and verification. This approach enables more efficient utilization of expensive reasoning tokens while achieving better task performance than undifferentiated reasoning allocation.

The technique also implies that model selection for complex tasks should consider phase-specific requirements rather than choosing models based solely on maximum reasoning capacity. A system designed for Reasoning Sandwich execution may achieve better cost-performance tradeoffs than systems optimizing for maximum uniform reasoning.

• Maxed Reasoning vs. Reasoning Sandwich
• Reasoning Models
• Reasoning Effort Levels
• Adaptive Reasoning / Cost-Aware Inference
• RAP: Reasoning via Planning with LLM as World Model