BiasConnect: Dual Bias-Control Frameworks

• BiasConnect is a term representing two distinct bias-control frameworks: a supervisor-based mitigation agent for retrieval and a counterfactual diagnostic tool for text-to-image models.
• The retrieval framework utilizes a multi-agent workflow in LangGraph that scores and selects minimally biased sources, while the text-to-image system applies counterfactual prompt interventions to evaluate bias interactions.
• Evaluations show significant bias reduction and actionable insights in both applications, despite trade-offs in latency, retry rates, and complex causal dynamics.

BiasConnect is a name used in 2025 arXiv literature for two distinct bias-oriented frameworks. In one line of work, it denotes a supervisor-based Bias Mitigation Agent for agentic retrieval and answer generation, designed to select sources that are both highly relevant and minimally biased before synthesis. In another, it denotes a counterfactual causal diagnostics tool for intersectional bias in text-to-image (TTI) models, designed to estimate how intervening on one bias axis affects others. The two systems share a common emphasis on moving bias analysis upstream—into source selection in retrieval or into causal interaction analysis in generative auditing—rather than treating bias solely as a post hoc property of final outputs (Singh et al., 26 Aug 2025, Shukla et al., 12 Mar 2025).

1. Terminological scope and disambiguation

The term BiasConnect does not refer to a single unified architecture. It names at least two technically unrelated frameworks:

BiasConnect variant	Domain	Core function
Bias Mitigation Agent	Agentic retrieval and answer generation	Selects sources using relevance and explicit bias scoring before writing
Causal diagnostics tool	Text-to-image generative modeling	Quantifies pairwise bias interactions via counterfactual prompt interventions and causal graphs

The retrieval-oriented BiasConnect is explicitly a multi-agent workflow implemented in LangGraph, with fairness enforced at the retrieval and source-selection stage. The TTI-oriented BiasConnect is explicitly a causal auditing layer rather than a mitigation algorithm itself; in the thesis formulation, it sits between TIBET and InterMit in the dependency chain TIBET → BiasConnect/BiasGraph → InterMit (Shukla, 28 Aug 2025).

A recurrent misconception is to treat these usages as variants of one method. The literature instead presents them as separate systems addressing different objects: retrieved documents in one case, and generated-image attribute distributions in the other. Their shared name reflects a shared research concern with bias interactions and control, not a shared implementation or formalism (Shukla, 28 Aug 2025).

2. Retrieval-stage bias mitigation in agentic AI

In "Bias Mitigation Agent: Optimizing Source Selection for Fair and Balanced Knowledge Retrieval" (Singh et al., 26 Aug 2025), BiasConnect is a supervisor-based Bias Mitigation Agent for agentic retrieval and answer generation. Its stated purpose is to improve the fairness of knowledge retrieval by not simply selecting the most relevant source, but by explicitly measuring bias in candidate documents and selecting sources that are both relevant and minimally biased. The paper’s central architectural claim is that, in agentic AI, fairness should be enforced at the retrieval/source-selection stage, not only at the final text-generation stage.

The system is organized as a multi-agent workflow in LangGraph. A Manager Agent coordinates specialized Worker Agents: a Knowledge Agent retrieves documents, a Bias Detection Agent scores them for bias, a Source Selection Agent chooses among candidates according to fairness-aware criteria, and a Writer Agent synthesizes the final answer using only the selected source. This last restriction is operationally important: answer generation is constrained to a vetted document rather than allowed to aggregate from a broader, potentially skewed candidate set (Singh et al., 26 Aug 2025).

The architecture maintains an explicit shared state

$$\mathcal{S} = (\mathcal{C}, \alpha, \kappa, \mu, \varrho),$$

where $\mathcal{C}=\{c_1,\dots,c_n\}$ is the retrieved candidate set, $\alpha$ is the document selected for answer generation, $\kappa$ is the current retry attempt, $\mu$ is the maximum number of retries, and $\varrho$ stores the reason for rejection if no source is accepted. Each document is represented as

$d_i = (\chi, \rho, \beta, \gamma),$

with $\chi$ the text content, $\rho \in [0,1]$ the relevance score, $\beta \in [0,1]$ the bias confidence score, and $\mathcal{C}=\{c_1,\dots,c_n\}$0 a binary label where 1 means biased and 0 means unbiased. This explicit state representation is used to coordinate retries, rejection reasons, and fairness checks in a controlled way (Singh et al., 26 Aug 2025).

3. Selection policies, retries, and constrained generation

The retrieval workflow begins with a user query $\mathcal{C}=\{c_1,\dots,c_n\}$1. The Manager Agent initializes state and invokes the Knowledge Agent, which uses ChromaDB as the retriever and fetches top-$\mathcal{C}=\{c_1,\dots,c_n\}$2 documents by vector similarity. From that point, BiasConnect supports three operational modes: No Source Selection, Zero-Shot, and Few-Shot (Singh et al., 26 Aug 2025).

In No Source Selection, the system behaves like a standard retrieval pipeline: it retrieves the most relevant source and uses it directly. In Zero-Shot, source selection uses the model’s own reasoning together with explicit bias and relevance criteria, without in-context examples. In Few-Shot, prompt demonstrations encode how to trade off bias, confidence, and relevance (Singh et al., 26 Aug 2025).

The Bias Detection Agent applies the pre-trained Dbias classifier to each candidate, producing a bias confidence score $\mathcal{C}=\{c_1,\dots,c_n\}$3 and a binary bias label $\mathcal{C}=\{c_1,\dots,c_n\}$4. These outputs are then consumed by the Source Selection Agent. In zero-shot mode, the first-attempt rule is

$\mathcal{C}=\{c_1,\dots,c_n\}$5

Thus, the selected source must be labeled unbiased and assigned sufficiently high confidence in that judgment; among such candidates, the highest-relevance document is chosen. If no document satisfies the rule, the system retries. On the final attempt, the criteria are relaxed so that an answer can still be produced (Singh et al., 26 Aug 2025).

In few-shot mode, the selector is described as using prompt demonstrations to implicitly learn a scoring function over $\mathcal{C}=\{c_1,\dots,c_n\}$6:

$\mathcal{C}=\{c_1,\dots,c_n\}$7

The paper presents this as a more stable and nuanced ranking procedure than pure thresholding, especially when content is subjective or ambiguous. If no acceptable source exists, the Manager updates $\mathcal{C}=\{c_1,\dots,c_n\}$8 and re-runs retrieval, often after query expansion from $\mathcal{C}=\{c_1,\dots,c_n\}$9 to $\alpha$0 to improve the candidate pool. The Writer Agent then receives the original query $\alpha$1 and the selected source $\alpha$2, and is instructed to answer strictly from that source (Singh et al., 26 Aug 2025).

4. Experimental profile of the retrieval agent

The retrieval-oriented BiasConnect was evaluated on 112 curated queries and annotated news articles from the MBIC and BABE datasets, both of which provide media-bias supervision. The system was tested with three OpenAI reasoning models: GPT-4o-mini, GPT-4.1, and GPT-4.1-mini. The reported evaluation dimensions were relevance, bias confidence, latency, and retry rate across the three operational modes (Singh et al., 26 Aug 2025).

The baseline was No Source Selection, which the paper characterizes as fast but highly biased. For GPT-4o-mini, the baseline bias rate was 49.11%; for GPT-4.1 and GPT-4.1-mini, the corresponding values were 56.25% and 52.68%. The principal result is that zero-shot source selection with GPT-4o-mini reduced bias to 8.93%, reported as an 81.82% bias reduction relative to the baseline. This is identified as the strongest fairness result in the paper. At the same time, GPT-4o-mini in zero-shot mode achieved an average relevance score of 0.366 ± 0.367, which was higher than the baseline on the reported scale (Singh et al., 26 Aug 2025).

The paper also reports a systematic efficiency trade-off. Fairness-aware selection increases latency because the system may retrieve multiple candidates, classify them, select among them, and retry with query expansion. For GPT-4o-mini, zero-shot averaged 41.38 s, compared with 17.88 s for the no-selection baseline. Retry behavior followed the same pattern: GPT-4o-mini zero-shot retried on 70.54% of queries, whereas few-shot reduced the retry rate to 29.46%. Few-shot also reduced bias substantially, with a 14.29% bias rate for GPT-4o-mini, while being somewhat faster than zero-shot (Singh et al., 26 Aug 2025).

These results suggest a clear engineering position within agentic retrieval: optimizing only for relevance can preserve high bias, whereas introducing classifier-based bias detection and fairness-aware source selection can substantially reduce bias, albeit with higher orchestration cost and longer end-to-end latency (Singh et al., 26 Aug 2025).

5. Counterfactual diagnostics for intersectional bias in text-to-image models

In the TTI literature, BiasConnect refers to a different system: a causal diagnostics tool for studying intersectional bias in black-box text-to-image generative models. Its motivating premise is that bias axes such as gender, age, ethnicity, clothing, emotion, body type, disability, and environment are often not independent. Intervening on one axis can improve another, leave it unchanged, or make it worse. This is the problem the original BiasConnect paper formalizes as bias interaction, and the thesis later situates it as the next step after TIBET, which identifies prompt-relevant axes and measures them independently (Shukla et al., 12 Mar 2025, Shukla, 28 Aug 2025).

The method proceeds by counterfactual prompt intervention. Given a prompt $\alpha$3 and bias axes $\alpha$4, it changes one axis at a time, synthesizes images with a black-box TTI model, and extracts attributes with MiniGPT-v2 using VQA-style questions. The causal intuition is intervention-based: if changing axis $\alpha$5 alters the empirical distribution of another axis $\alpha$6, then there is an intersectional dependency from $\alpha$7 to $\alpha$8 (Shukla, 28 Aug 2025).

The motivation is illustrated with prompt-level examples in which apparently local edits have distributed effects. The thesis notes that prompts such as “an old man at a church” versus “an Asian old man at a church” can alter not only ethnicity, but also emotion, posture, clothing, and overall scene mood. The operational concern is therefore not merely whether a model is biased along a single protected dimension, but whether mitigating one axis inadvertently amplifies or distorts another. The TTI-oriented BiasConnect is introduced precisely to diagnose these cross-axis trade-offs before mitigation is applied (Shukla, 28 Aug 2025).

6. Intersectional Sensitivity, BiasGraph, and empirical findings

The core quantitative object in the TTI formulation is Intersectional Sensitivity. For a target axis $\alpha$9, the empirical distribution from the initial prompt is compared to a desired distribution $\kappa$0 using Wasserstein-1 distance:

$\kappa$1

with

$\kappa$2

After intervening on another axis $\kappa$3, the resulting distribution over $\kappa$4 is compared again:

$\kappa$5

The intersectional effect is then defined as

$\kappa$6

A positive value means mitigating $\kappa$7 improves $\kappa$8; a negative value means it worsens $\kappa$9; zero means no effect. The thesis notes that $\mu$0 after normalization, and that a matrix $\mu$1 of these values forms an intersectionality matrix (Shukla, 28 Aug 2025).

BiasConnect complements this treatment-effect estimate with pairwise causal discovery. It builds contingency tables over extracted attributes and applies a chi-square conditional independence test. In the occupation experiments, the paper uses a p-value threshold of 0.0001; in the TIBET setting it uses 0.05 because those concepts are more diverse and prompt-specific. Surviving directed relationships are visualized as a pairwise causal graph or BiasGraph, in which nodes are bias axes and directed edges indicate measurable downstream impact (Shukla et al., 12 Mar 2025).

The empirical program spans two datasets. The first is a custom occupation dataset with 26 occupational prompts and eight bias dimensions: gender, age, ethnicity, environment, disability, emotion, body type, and clothing. The second is TIBET, with 100 creative prompts, LLM-generated bias axes and counterfactuals, and 48 images per prompt/counterfactual. Evaluated models include Stable Diffusion 1.4, Stable Diffusion 3.5, Flux-dev, Playground v2.5, and Kandinsky 2.2; the VQA backbone is MiniGPT-v2 (Shukla et al., 12 Mar 2025).

The reported findings emphasize that cross-axis effects are neither rare nor uniformly beneficial. In a motivating study on 26 occupational prompts with Stable Diffusion 1.4 and ITI-GEN, independently mitigating eight bias dimensions negatively affected other axes in over 29.4% of cases. Specific examples include a positive clothing-to-emotion effect for librarians in Stable Diffusion 3.5, reported as $\mu$2, and a negative ethnicity-to-gender effect for athletes, reported as $\mu$3. The thesis also highlights a prompt such as “a photo of a musician”, where increasing gender diversity can skew age distribution so that older musicians become less represented (Shukla et al., 12 Mar 2025, Shukla, 28 Aug 2025).

Validation is framed in predictive terms. The original paper reports an average correlation of $\mu$4 between BiasConnect’s estimated treatment effects and actual post-mitigation observations using ITI-GEN on Stable Diffusion 1.4; the abstract rounds this to $\mu$5. The later InterMit paper reports $\mu$6 for ITI-GEN, and the thesis reports $\mu$7 for hardprompt-style mitigation. These results are presented as evidence that the estimates are meaningfully predictive of downstream mitigation effects rather than merely descriptive of pre-mitigation prompt behavior (Shukla et al., 12 Mar 2025, Shukla et al., 22 May 2025).

7. Downstream role, strengths, and limitations

The TTI-oriented BiasConnect is closely tied to later mitigation work but remains conceptually distinct from it. In the thesis, BiasConnect is the causal auditing layer that informs InterMit, which uses the BiasConnect sensitivity matrix together with user-defined priorities and target distributions to decide which axis to mitigate next. The 2025 InterMit paper reports that this BiasConnect-guided mitigation achieves lower bias—0.33 vs. 0.52—with fewer mitigation steps—2.38 vs. 3.15 average steps—than a traditional baseline, while also yielding superior image quality (Shukla et al., 22 May 2025).

This relationship clarifies another common misunderstanding. In retrieval, BiasConnect is itself a mitigation system. In TTI auditing, BiasConnect is not a mitigation algorithm itself; it is a prompt-sensitive, model-agnostic diagnostic mechanism for anticipating trade-offs, selecting mitigation axes, and comparing architectures by the dependency structures they learn (Shukla, 28 Aug 2025).

The papers are also explicit about limitations. In the retrieval setting, fairness-aware source selection increases latency and retry rate because the system may need multiple retrieval-classification-selection cycles with query expansion (Singh et al., 26 Aug 2025). In the TTI setting, BiasConnect depends on the accuracy of VQA-based attribute extraction, on the faithfulness of counterfactual prompts, and on the normative choice of the ideal distribution $\mu$8. It captures pairwise dependencies rather than full higher-order causal structure, does not model indirect causal effects, and does not by itself produce an optimal multi-axis mitigation policy (Shukla, 28 Aug 2025, Shukla et al., 12 Mar 2025).

Taken together, these systems show how the name BiasConnect has been used to denote two different upstream bias-control paradigms. One embeds fairness into retrieval-time source vetting for agentic language systems; the other exposes interaction structure among bias axes in generative vision models through counterfactual interventions and causal graphs. Their common significance lies less in a shared implementation than in a shared methodological stance: bias should be analyzed where decisions are being structured, whether that structure is a source-selection workflow or an intervention graph over intersecting attributes (Singh et al., 26 Aug 2025, Shukla et al., 12 Mar 2025).