POPE

[Cc2021start]

Hello, author. Thank you for your contribution, which has provided us with a new perspective.

But I'd like to ask if you have any test results on the POPO indicator. Your method gives me an intuition that there might be negative optimization on the pope indicator?

[ZhangqiJiang07]

Hello, author. Thank you for your contribution, which has provided us with a new perspective.

But I'd like to ask if you have any test results on the POPO indicator. Your method gives me an intuition that there might be negative optimization on the pope indicator?

Hi, thanks for your feedback!

You are right that POPE uses a "Yes/No" VQA format instead of the object's noun, which is inconsistent with our paper's analysis approach.

The table below shows the average results on LLaVA-1.5-7B across the three splits (popular, random, and adversarial), conducted during rebuttal.

LLaVA-1.5	Accuracy	Precision	F1
Greedy	85.2	82.2	86.1
Ours	86.2	84.8	86.6

Hope this helps!

[Cc2021start]

Hello, author. Thank you for your contribution, which has provided us with a new perspective.
But I'd like to ask if you have any test results on the POPO indicator. Your method gives me an intuition that there might be negative optimization on the pope indicator?

Hi, thanks for your feedback!

You are right that POPE uses a "Yes/No" VQA format instead of the object's noun, which is inconsistent with our paper's analysis approach.

The table below shows the average results on LLaVA-1.5-7B across the three splits (popular, random, and adversarial), conducted during rebuttal.

LLaVA-1.5 Accuracy Precision F1
Greedy 85.2 82.2 86.1
Ours 86.2 84.8 86.6
Hope this helps!

Hello author, I can't reproduce the chair metric in my paper either. Of course, this is not the problem of your single paper. The chair metric in many papers cannot be reproduced, including vcd, icd, etc. I used your code and the Chairs were approximately 31%. Is this phenomenon normal? Is it because the chair metric is too sensitive to image selection or for some other reason?

[ZhangqiJiang07]

Hello, author. Thank you for your contribution, which has provided us with a new perspective.
But I'd like to ask if you have any test results on the POPO indicator. Your method gives me an intuition that there might be negative optimization on the pope indicator?

Hi, thanks for your feedback!
You are right that POPE uses a "Yes/No" VQA format instead of the object's noun, which is inconsistent with our paper's analysis approach.
The table below shows the average results on LLaVA-1.5-7B across the three splits (popular, random, and adversarial), conducted during rebuttal.
LLaVA-1.5 Accuracy Precision F1
Greedy 85.2 82.2 86.1
Ours 86.2 84.8 86.6
Hope this helps!

Hello author, I can't reproduce the chair metric in my paper either. Of course, this is not the problem of your single paper. The chair metric in many papers cannot be reproduced, including vcd, icd, etc. I used your code and the Chairs were approximately 31%. Is this phenomenon normal? Is it because the chair metric is too sensitive to image selection or for some other reason?

Yes, the CHAIR metric can be sensitive to image selection.

For full reproducibility, we've attached our detailed CHAIR results for LLaVA-1.5-7B. The file includes the MSCOCO image_ids used in our evaluation, which you can use to replicate our results.

llava_1.5_7B_CHAIR.json

[Cc2021start]

Hello, author. Thank you for your contribution, which has provided us with a new perspective.
But I'd like to ask if you have any test results on the POPO indicator. Your method gives me an intuition that there might be negative optimization on the pope indicator?

Hi, thanks for your feedback!
You are right that POPE uses a "Yes/No" VQA format instead of the object's noun, which is inconsistent with our paper's analysis approach.
The table below shows the average results on LLaVA-1.5-7B across the three splits (popular, random, and adversarial), conducted during rebuttal.
LLaVA-1.5 Accuracy Precision F1
Greedy 85.2 82.2 86.1
Ours 86.2 84.8 86.6
Hope this helps!

Hello author, I can't reproduce the chair metric in my paper either. Of course, this is not the problem of your single paper. The chair metric in many papers cannot be reproduced, including vcd, icd, etc. I used your code and the Chairs were approximately 31%. Is this phenomenon normal? Is it because the chair metric is too sensitive to image selection or for some other reason?

Yes, the CHAIR metric can be sensitive to image selection.

For full reproducibility, we've attached our detailed CHAIR results for LLaVA-1.5-7B. The file includes the MSCOCO image_ids used in our evaluation, which you can use to replicate our results.

llava_1.5_7B_CHAIR.json

Hello, author. I've observed the answers generated by your method and found that you have sacrificed a lot of the details of the answers?

Some photos are directly:

{"image_id": 131444, "caption": "The man is wearing a blue shirt and a brown tie."}

[ZhangqiJiang07]

As discussed in the ablation study for the balance factor $\alpha$ (Section 5), a trade-off exists: a lower $\alpha$ limits hallucination mitigation effectiveness, whereas a higher $\alpha$ can compromise descriptive richness. Since a fixed $\alpha$ controls the global intervention strength for every tested image, the scenario you mentioned may be expected for images with relatively simple or sparse content. In such cases, a high $\alpha$ might be overly suppressive. An adaptive, image-centric $\alpha$ could offer a more robust solution. Cc2021start ***@***.***> 于2025年10月9日周四 17:08写道：

…

*Cc2021start* left a comment (ZhangqiJiang07/middle_layers_indicating_hallucinations#3) <#3 (comment)> Hello, author. Thank you for your contribution, which has provided us with a new perspective. But I'd like to ask if you have any test results on the POPO indicator. Your method gives me an intuition that there might be negative optimization on the pope indicator? Hi, thanks for your feedback! You are right that POPE uses a "Yes/No" VQA format instead of the object's noun, which is inconsistent with our paper's analysis approach. The table below shows the average results on LLaVA-1.5-7B across the three splits (popular, random, and adversarial), conducted during rebuttal. LLaVA-1.5 Accuracy Precision F1 Greedy 85.2 82.2 86.1 Ours 86.2 84.8 86.6 Hope this helps! Hello author, I can't reproduce the chair metric in my paper either. Of course, this is not the problem of your single paper. The chair metric in many papers cannot be reproduced, including vcd, icd, etc. I used your code and the Chairs were approximately 31%. Is this phenomenon normal? Is it because the chair metric is too sensitive to image selection or for some other reason? Yes, the CHAIR metric can be sensitive to image selection. For full reproducibility, we've attached our detailed CHAIR results for LLaVA-1.5-7B. The file includes the MSCOCO image_ids used in our evaluation, which you can use to replicate our results. llava_1.5_7B_CHAIR.json <https://github.com/user-attachments/files/22589488/llava_1.5_7B_CHAIR.json> Hello, author. I've observed the answers generated by your method and found that you have sacrificed a lot of the details of the answers? Some photos are directly: {"image_id": 131444, "caption": "The man is wearing a blue shirt and a brown tie."} — Reply to this email directly, view it on GitHub <#3 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/A2QVVCYOBJ6LUM3ZF2S2Z3L3WYQZ3AVCNFSM6AAAAACHLAEDWWVHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMZTGOBUHEYDOOJZHE> . You are receiving this because you commented.Message ID: <ZhangqiJiang07/middle_layers_indicating_hallucinations/issues/3/3384907999 @github.com>