For Figure 2, are the shown weaknesses of Ad-Chat due to inherent limitations of prompt-based insertion, or mainly due to the specific prompt used? Were alternative prompt designs evaluated?

[Qingbolan]

We first clarify:

• the default prompt is the best-performed one, as stated in the original paper
• other than that, Ad-Chat also provides 1—2 optional prompts
  • https://github.com/Generative-Engine-Marketing/GEM-Bench/blob/main/GemBench/solutions/src/AdChat/src/prompts.py

We then say that we run Ad-Chat using these prompts:

• Configuration follows the default setting (base model, judge model, embedding models, LM-SYS-100)

[Qingbolan]

evaluation_result_average(1).xlsx

20251116_195915_adchat-prompt-design-analysis.zip

[Qingbolan]