Recursive Language Model 推理任务设计与评估框架

You are an expert in Recursive Language Models (RLMs) and iterative reasoning systems.

Design a task execution framework using RLM principles for the following scenario:

Task Description

• Problem domain: [e.g., mathematical proof, code debugging, research synthesis]
• Complexity level: [e.g., requires 3-7 recursive refinement steps]
• Quality threshold: [e.g., 95% accuracy on validation set]

Framework Requirements

1. Recursion Strategy

• Define the base case (when to stop recursion)
• Design the recursive decomposition (how to break the problem into sub-problems)
• Specify the merge/synthesis function (how to combine sub-results)
• Set maximum recursion depth and timeout limits

2. Sandbox Isolation

• Each recursion level runs in an isolated sandbox
• Define input/output contracts between levels
• Implement state checkpointing for recovery

3. Quality Gates

• Self-evaluation criteria at each recursion step
• Confidence scoring mechanism
• Automatic retry with modified parameters on failure

4. Optimization

• Token budget allocation across recursion levels
• Early termination when confidence exceeds threshold
• Caching strategy for repeated sub-problems

Output Format

Provide:

1. Pseudocode for the recursive execution engine
2. Prompt templates for each recursion level
3. Evaluation rubric with scoring criteria
4. Example walkthrough with a concrete problem
5. Cost estimation formula based on recursion depth