- Rating
- 2
- Confidence
- 3
- Informativeness
- 3(+1)
- Presentation
- 3(+1)
- Soundness
- 2
Read review
Paper summary
The paper proposes OmniDebias for unified understanding/generation models (Janus-Pro, BAGEL). For each sensitive attribute, it takes counterfactual activation differences (e.g. male CEO vs female CEO) across many occupation contexts. It treats each context’s deviation from the averaged difference as bias. It then removes semantic-content directions with a reversed FairPCA step and keeps the top singular vectors. These per-attribute biases are merged across attributes and modalities by nested SVD into one shared subspace. That subspace is either projected out of the weights (PE) or penalised in the hidden states during fine-tuning (FT). Experiments cover T2T, T2I and I2T on gender, race and age, with SES held out as an unseen attribute.
Critical strengths
-
Using one shared bias subspace instead of a separate target for each attribute is a reasonable idea for models that handle several modalities. The PE variant is a cheap, label-free, one-time weight edit. The observation in Fig. 2, that the gender direction rotates with the occupation, is a nice motivating analysis.
-
The evaluation covers all three directions (T2T/T2I/I2T) on two models. The method is compared against prompt, PE and FT baselines adapted to this setting.
Critical weaknesses
The idea is clean, but the main claim, removing bias while keeping factual knowledge, isn’t really tested. Utility is only checked with things like GPT-2 PPL, and even that gets worse. The examples in Fig. 11 actually look like knowledge is being removed. On top of that, a lot of the gains are just 0.01 with no variance reported, so it’s hard to tell if the method really works. Details as follows:
-
The gains are small, and the claim of consistent improvement across all 11 settings is not well supported.
- Many T2T changes are 0.01 on a two-decimal metric. For example, Ours-PE on Janus moves gender from 0.16 to 0.15 and race from 0.24 to 0.23.
- There are no seeds, standard deviations or significance tests, and results stay far from the oracle.
- On I2T, Ours-FT is clearly worse than CSD and CDA-FT on gender and race.
- The OOD claim rests on a single attribute (SES). SES is plausibly correlated with race and with the stereotype dimensions already targeted. The T2I SES numbers are also hard to interpret.
-
"Preserving factual knowledge" is the core claim but is not really tested.
- Utility is measured only with GPT-2 PPL, MUSIQ and CLIPScore, and even these degrade. On Janus, Ours-PE raises PPL from 7.74 to 9.01 (T2T) and from 7.70 to 9.87 (I2T). Ours-FT lowers MUSIQ from 71.2 to 69.4 and CLIPScore from 0.29 to 0.26.
- There are no general capability benchmarks for understanding or generation after the edit.
- The children-as-firefighters cases in Fig. 11 suggest that legitimate age–occupation knowledge is being removed.
- The assumption that deviation from the mean difference equals bias is never validated. If the contexts are skewed, the mean itself can carry bias. Some context dependence may also be legitimate.
Questions for authors
- Are the contexts used to extract Φ the same WinoBias professions used for evaluation? If so, please report results on held-out contexts. This matters a lot for how I read the gains.
- Can you report mean ± std over several seeds or prompt samples for Tables 1–2?
- How does the model do on standard understanding/generation benchmarks after PE and FT?
- How does this compare with linear concept erasure methods such as hard-debias (Bolukbasi et al.), INLP or LEACE, built from the same counterfactual pairs? None of these are discussed, and they seem methodologically close.
- Which weight matrices and layers are edited in Eq. 7, and what input data is used for x in Eq. 8?
- CDA-PE gets MUSIQ 23.2 on BAGEL, which looks like a broken adaptation. Were the baselines tuned?
3 comments
Response to Reviewer prUR (Part 2/2)
General Capability (W2/ Q3)
To further evaluate the impact of OmniDebias on general model capabilities, we conduct additional experiments on MMLU (5-shot), TruthfulQA, and COCO (T2I). All results are reported on Janus using OmniDebias-PE with the same hyperparameter settings as in the original manuscript.
| Dataset | MMLU Micro Acc (↑) | TruthfulQA Acc (↑) | COCO ClipScore (↑) | COCO MUSIQ (↑) |
|---|---|---|---|---|
| Vanilla | 0.494 | 0.285 | 0.31 | 69.19 |
| Ours | 0.479 | 0.275 | 0.30 | 68.81 |
As shown above, OmniDebias maintains comparable performance to the vanilla model across standard understanding and generation benchmarks, with only minor performance reductions. These results further support that our bias mitigation preserves general model capabilities without substantial utility degradation.
Comparison with Related Works (Q4)
We thank the reviewer for highlighting these relevant works and will include them in the revised manuscript. Similar to our existing baselines (e.g., SDID and FairImagen), Hard Debias, INLP, and LEACE also follow an attribute-centric paradigm, targeting attribute-related representations without explicitly disentangling stereotypical bias from legitimate factual knowledge. In contrast, OmniDebias explicitly seeks to isolate bias-related variations while preserving legitimate attribute knowledge, rather than indiscriminately removing attribute information.
Abnormal Proformance (Q6)
We thank the reviewer for identifying this typo. The correct MUSIQ score for CDA-PE on BAGEL is 74.1, rather than 23.2. We will correct this error in the revised manuscript.
If you have any further questions or concerns, please feel free to reach out.
Reference
[1] Arditi, Andy, et al. "Refusal in language models is mediated by a single direction." Advances in Neural Information Processing Systems 37 (2024): 136037-136083.
[2] Fu, Zihao, et al. "Fairimagen: Post-processing for bias mitigation in text-to-image models." Advances in Neural Information Processing Systems 38 (2026): 29488-29512.
Thanks for the reply. Based on the response, I have raised the informativeness and presentation scores. However, since the performance improvements were not significant enough, I have decided to keep my overall rating.
Response to Reviewer prUR (Part 1/2)
We sincerely appreciate your time and effort in reviewing our manuscript. Below are our responses to your valuable comments.
We would like to note that OmniDebias achieves substantial improvements on classifier-based fairness metrics, while also delivering meaningful gains on embedding-based similarity metrics (e.g., T2T and T2I SES), where improvements appear numerically modest due to differences in metric scales. More importantly, our primary contribution lies in achieving consistent fairness improvements across all 11 attribute–modality settings, which none of the competing baselines achieves.
We would like to clarify that the vanilla model does not exhibit fair age preferences; instead, it predominantly generates middle-aged individuals, reflecting stereotypical associations rather than legitimate age constraints. From a technical perspective, OmniDebias successfully mitigates such preferences, achieving improved statistical fairness under the target fairness objective. However, as shown in Fig. 11, these fairness gains do not necessarily guarantee social plausibility under context-specific constraints.
Such failure cases could be mitigated by incorporating explicit constraints, e.g., replacing younger age descriptions with “young adults over 18” during bias extraction. However, similar to prior attribute-centric methods, this requires additional domain-specific priors that are difficult to enumerate and scale across diverse social contexts. We believe that bridging this gap between statistical fairness and social plausibility represents an important direction for future research.
Unlike prior methods that indiscriminately remove attribute representations containing both stereotypical bias and factual knowledge (e.g., biological characteristics), OmniDebias explicitly seeks to disentangle bias from factual information to better preserve legitimate attribute knowledge. Specifically, we employ diverse semantic contexts and matched attribute pairs to reduce irrelevant variations, followed by semantic-content deflation in Eqs. (3)–(4) to further minimise interference with legitimate knowledge.
We would like to emphasis that the across-context mean in Eq. (2) serves as a reference for identifying context-dependent variations rather than a strictly unbiased representation, and we will make this explicit in the revised manuscript. While identifying stereotypical bias from these variations involves modelling assumptions, the observed fairness improvements and utility preservation provide empirical support for the effectiveness of this formulation. Importantly, we emphasise that requiring theoretically exact disentanglement imposes an unrealistic standard on activation-based methods (e.g., refusal vectors [1]), as such methods cannot generally guarantee fully capturing relevant information without introducing spurious components. We believe such a requirement goes beyond the scope of existing activation-based methods and should not be regarded as a limitation unique to OmniDebias.
We follow the original dataset configuration of prior work [2] to conduct our training and evaluation. Upon further examination, we identified an overlap of only 6 out of 100 professions between the bias extraction and evaluation sets. We therefore re-evaluated OmniDebias after excluding these overlapping professions from evaluation and observed negligible changes in both fairness and utility performance. Furthermore, the consistent improvements on the unseen SES attribute provide additional evidence of the generalisation capability of OmniDebias.
For Tables 1–2, we use a fixed random seed of 0 and, following the common evaluation protocol for T2T and T2I in [2], sample each test prompt 12 times to reduce sampling variability and improve evaluation reliability. Given the limited rebuttal period, we plan to conduct additional experiments across multiple random seeds in future work. For OmniDebias-PE, we apply Eq. (7) to the output projection matrices across all modality-shared Transformer layers, ensuring unified debiasing across modalities. For OmniDebias-FT, the input data in Eq. (8) are drawn from the same three-modal dataset used for bias extraction in PE, with the unified bias representation dynamically re-estimated at each training epoch.