Does the reward_dim parameter actually do anything?

[KZF-kzf]

I’ve noticed that whether I set it to dim1, dim2, or overall_detail, the output values remain exactly the same.
my code

import torch
from EditReward import EditRewardInferencer

CHECKPOINT_PATH = "EditReward-MiMo-VL-7B-SFT-2508"
CONFIG_PATH = "EditReward/config/EditReward-MiMo-VL-7B-SFT-2508.yaml"

_inferencer = None


def _get_inferencer(dim):
    global _inferencer
    if _inferencer is None:
        _inferencer = EditRewardInferencer(
            config_path=CONFIG_PATH,
            checkpoint_path=CHECKPOINT_PATH,
            device="cuda" if torch.cuda.is_available() else "cpu",
            reward_dim=dim,
        )
    return _inferencer


def compute_editreward_score(src_path: str, tgt_path: str, instruction: str, dim) -> float:
    inferencer = _get_inferencer(dim)
    with torch.no_grad():
        rewards = inferencer.reward(
            prompts=[instruction],
            image_src=[src_path],
            image_paths=[tgt_path]
        )
    print(rewards)
    print(dim)
    return rewards[0][0].item()


if __name__ == "__main__":
    img_src = "22_source.png"
    img_tgt = "22.png"
    prompt = "在古董车打开的行李箱里添加一只坐着的小棕狗。"
    score = compute_editreward_score(img_src, img_tgt, prompt, dim="dim1")
    print(f"EditReward result: {score:.4f}")

[KemingWu]

Thanks for reporting this.

What you are seeing is expected for the current released EditReward-MiMo-VL-7B-SFT-2508 checkpoint.

A few clarifications:

• This checkpoint is configured with rm_head_type="ranknet_multi_head" and pooling_strategy="mean".
• In that mode, the inference code always builds both dim1 and dim2 inputs internally, and then fuses the two heads into a single final output.
• So for the released multi-head checkpoint, reward_dim does not switch the returned score to an individual dimension-specific score.
• overall_detail is also not exposed as a separate third head in the current public inference path.

There is also one extra issue in your example: _inferencer is cached globally, so after the first initialization, later calls with a different dim will still reuse the same inferencer instance. However, even if you recreate the inferencer each time, the returned score will still be the same for this checkpoint because of the multi-head fusion described above.

One more note: the returned tensor has shape [B, 2], but these two values are not [dim1, dim2]. For this checkpoint, they correspond to the fused reward statistics used by the uncertainty formulation, where rewards[..., 0] is the fused mean score and rewards[..., 1] is the log-uncertainty term.

So in short: with the current released MiMo checkpoint, getting the same value for dim1, dim2, and overall_detail is expected behavior.

If you need per-dimension outputs, we would need to expose the individual heads explicitly in the inference API rather than returning only the fused score.