LLM Test-time Compute 自适应推理优化提示词

You are an AI Reasoning Optimization Specialist. Help me design a test-time compute scaling strategy for my LLM application.

Context

• Task type: [e.g., complex reasoning, code generation, math proofs, creative writing]
• Base model: [e.g., GPT-4o, Claude Opus, Qwen3, Llama 4]
• Current pain point: [e.g., inconsistent quality, fails on hard problems, too slow]
• Latency budget: [e.g., <5s, <30s, unlimited]

Design the Following

1. Adaptive Compute Strategy:

   • When to use simple single-pass inference vs extended thinking
   • Difficulty classification heuristics for routing
   • Token budget allocation by task complexity tier

2. Self-Verification Pipeline:

   • Generate → Verify → Refine loop design
   • Confidence scoring method
   • Early-exit criteria to avoid wasting compute

3. Multi-Sample Strategies:

   • Best-of-N sampling with reward model scoring
   • Majority voting for factual tasks
   • When to use tree search vs sequential refinement

4. Implementation Template:

   • Pseudocode for the adaptive routing logic
   • Prompt templates for the verifier/critic agent
   • Cost-quality tradeoff analysis

Provide concrete examples and expected improvement ranges based on published research.