Results are only presented with niche LLM of doubao-1.5

[Qingbolan]

Fix LaaJ LLM model to GPT-4.1-mini, try the following possible candidates as base LLMs:

• Qwen3: 8b, 14b, 30b-a3b
• Gemini: 2.0-flash-lite, 2.5-flash-lite
• repeat time: one time first for all five models, and preserve two base models (one from Qwen and one from Gemini) for repeating three times
• datasets: first try LM-SYS-100

[Qingbolan]

• Qwen3-8B and Qwen3-14B could not be included in our experiments because their context windows are too small to support the full input required by the baseline and our three proposed methods. As a result, these models cannot execute comparable runs.

[Qingbolan]

[Qingbolan]

[Qingbolan]

rebuttle01.xlsx