[Question/Reproduction] Significant discrepancy in Sentient Scores compared to the paper (Abnormal results for GPT-4o and Qwen)

Hi there,

I am currently attempting to reproduce the results reported in the SAGE paper. Following the instructions, I used DeepSeek-V3 as the judge model to evaluate several models, including gemini-2.5-pro, gpt-4o, and qwen3-8b.

However, the experimental results I obtained are highly counter-intuitive and do not align with the benchmarks reported in the paper. Specifically, the score for GPT-4o is unreasonably low, while the score for the smaller model (Qwen3-8b) seems unexpectedly high.

Reproduction Results
Here are the Sentient Scores we obtained:

Suspected Cause
We suspect that this discrepancy might be caused by the lack of specific call_llm implementations (or configurations) for these models in the repository. Without specific handling, the default generation hyperparameters (such as temperature, top_p, etc.) might be causing instability in the model outputs, leading to skewed evaluation results.

Configuration Details
Below are the settings we used for the reproduction:

1. NPC LLM Configuration
We used anthropic/claude-3.5-sonnet via OpenRouter:

CUSTOM_API_CONFIG = {
    "api_key": "...",
    "base_url": "https://openrouter.ai/api/v1",
    "model_name": "anthropic/claude-3.5-sonnet",
    "max_tokens": 8192,
    "temperature": 0.7,
    "top_p": 0.8,
    "presence_penalty": 1.5
}

2. Judger LLM Configuration
We used deepseek-chat (DeepSeek-V3) as the judge:

# judger llm
response = client.chat.completions.create(
    model="deepseek/deepseek-chat", # which is deepseek v3 on openrouter
    messages=[{"role": "user", "content": prompt}],
    temperature=0.7,
)

Request
Could you please clarify the exact hyperparameters (temperature, top_p, etc.) used for the models and the judge in the paper? Alternatively, are there specific configurations required in the call_llm function that we might have missed?

Any guidance on reproducing the official results would be greatly appreciated.

Thanks!

Start by reviewing the repository's call_llm implementations and the reproduction instructions, then compare their model and judge configuration with the settings shown here. Check whether temperature, top_p, or other generation parameters are defined for GPT-4o and Qwen3-8b, and document the configuration needed to reproduce the paper's reported scores.