TL;DR
OmniDebias shifts the fairness paradigm for omni-modal large language models (OLLMs) from "attribute-centric" to "bias-centric." By disentangling social stereotypes from legitimate knowledge and consolidating them into a single, shared bias representation, it enables unified debiasing across multiple modalities and generalizes to unseen sensitive attributes without requiring separate, conflicting objectives.
What Problem Does This Paper Address?
As AI moves toward Omni-modal Large Language Models (OLLMs)—models that handle text, images, and more in a single joint embedding space—fairness becomes a scaling nightmare. Traditional "attribute-centric" debiasing requires a separate fairness objective for every combination of attribute (gender, race, age) and modality (text-to-image, image-to-text).
This approach has three major flaws:
- Scalability: It’s impossible to enumerate every possible demographic concept and interaction.
- Mutual Interference: Optimizing for one attribute (e.g., gender) can inadvertently degrade fairness for another (e.g., race) because their representations are entangled.
- Generalization: These methods fail on "out-of-distribution" (OOD) attributes that weren't explicitly targeted during training.
Methods
The authors introduce OmniDebias, which treats bias as a structured, shared phenomenon rather than a collection of independent problems.
1. Bias Disentanglement
The core idea is to exploit activation inconsistencies. Ideally, switching an attribute (e.g., male $\leftrightarrow$ female) should produce the same "shift" in the model's internal activations regardless of the context. However, in biased models, the shift for "CEO" looks different than the shift for "nurse" because stereotypes are entangled with the occupation's representation. By measuring these deviations across many contexts, OmniDebias isolates the "bias" component from "legitimate knowledge."
2. Bias Unification
Instead of treating gender bias and racial bias as separate vectors, the method hierarchically aggregates these extracted biases into a unified bias representation ($\Phi^{(l)}$). This captures the underlying "cross-attribute structure" of how the model encodes stereotypes.
3. Mitigation
This unified representation can be used in two ways:
- OmniDebias-PE (Post-hoc Editing): A training-free method that projects model weights onto the orthogonal complement of the bias subspace at inference time.
- OmniDebias-FT (Fine-tuning): An adaptive approach that dynamically extracts and penalizes bias-aligned components during training without needing ground-truth labels.
Datasets
The researchers evaluated their method on two representative OLLMs (Janus-Pro-7B and BAGEL-7B-MoT) using:
- Extended WinoBias: 114 professions paired with gender, race, and age attributes for Text-to-Text (T2T) and Text-to-Image (T2I) tasks.
- CMSC Dataset: 18 social concepts (personality, education, etc.) for Image-to-Text (I2T) description.
- Socioeconomic Status (SES): Used as an unseen/OOD attribute to test if the model could generalize fairness to a concept it wasn't specifically "debiased" for.
Key Insights and Ideas
- Bias has a "Shared Signature": The most striking insight is that different social biases (gender, race, age) share a common underlying structure in the model's latent space. Capturing this "unified bias" is more effective than chasing individual attributes.
- Knowledge Preservation: By focusing on the inconsistency of shifts across contexts, the method avoids "throwing the baby out with the bathwater." It removes the stereotypical associations while keeping the actual meaning of the attributes (e.g., the model still knows what "female" means, just not that "female" implies "homemaker").
- Zero-Shot Fairness: Because the bias representation is unified, the model shows improved fairness on attributes it never saw during the debiasing process (like Socioeconomic Status).
Key Results
OmniDebias consistently outperformed attribute-centric baselines across 11 different attribute-modality settings.
- Quantitative: On Janus-Pro-7B, OmniDebias-FT improved fairness metrics across the board while maintaining or even improving generation quality (measured by Perplexity and MUSIQ scores). For example, in T2I tasks, it significantly reduced demographic skew compared to the vanilla model.
- Qualitative: As seen in Figure 4, while the vanilla model predominantly generates white, middle-aged men when prompted for a "CEO," OmniDebias produces a much more balanced and diverse set of images.
Limitations and Future Work
- Coarse Categories: The current study uses relatively coarse demographic categories (e.g., binary gender). Real-world identities are much more granular (non-binary, gender-fluid), which may present more complex bias structures.
- Modality Expansion: While it covers text and images, future OLLMs will include audio and video. Extending OmniDebias to these high-dimensional temporal modalities is a key next step.
- Architectural Heterogeneity: The method works best on models with shared autoregressive representations. Models with highly specialized, modality-specific decoders (like certain flow-based image generators) might require further adaptation.
Summary updated