Score vs Generation Tokens / Question — Qwen3.5-9B
DSPy
score
DSPy. base: direct 4.1k/3.3, k=5 7.9k/8.6, k=20 8.6k/9.6, k=20 NER 34.6k/29.4. cheatsheet: direct 3.9k/6.3, k=5 6.1k/14.4, k=20 7.1k/14.1, k=20 NER 29.9k/23.1.
generation tokens / question
OpenClaw
score
OpenClaw. base: direct 4.1k/2.3, k=5 4.6k/6.9, k=20 9.7k/15.9, k=20 NER 24.3k/17.6. cheatsheet: direct 3.8k/4.3, k=5 6.0k/8.6, k=20 9.1k/15.2, k=20 NER 20.1k/18.1.
generation tokens / question
Qwen3.5-9B Qwen3.5-9B + cheatsheet
Figure 6. Cheatsheet baseline — score vs. generation tokens per question for Qwen3.5-9B, base versus cheatsheet, on DSPy and OpenClaw. Points: direct, k=5, k=20, and k=20 no-early-return (NER).