Bias Transfer Hypothesis (BTH): Implications in AI
- Bias Transfer Hypothesis (BTH) is a theory that intrinsic biases in pretrained models transfer into task-specific behavior, as shown by strong correlations in prompt-adapted LLMs.
- Recent studies reveal that prompt-based adaptation, unlike fine-tuning for MLMs, preserves bias patterns with Pearson’s correlations often exceeding 0.94 on tasks like WinoBias.
- Mitigation attempts using debiasing prompts can reduce bias magnitude but typically fail to alter the intrinsic bias structure preserved during adaptation.
Searching arXiv for papers on the Bias Transfer Hypothesis and closely related bias-transfer work. The Bias Transfer Hypothesis (BTH) is the theory that social biases internalized during pre-training transfer into task-specific behavior after adaptation. In the recent LLM literature, BTH is formulated primarily as an association claim rather than a causal one: “transfer” is operationalized as correlation between intrinsic bias in a pre-trained model and prompt-adapted bias in downstream behavior, typically through occupation-level bias scores or related task-level disparity measures (Sivakumar et al., 9 Sep 2025). Earlier BTH work often reported weak or absent transfer for masked LLMs (MLMs) under fine-tuning, but more recent work argues that this conclusion does not generalize to causal LLMs under prompting, where strong alignment between pre-trained and adapted bias patterns is repeatedly observed (Mackraz et al., 2024). Across adjacent literatures, the same core idea appears in transfer learning, multilingual representation alignment, multimodal CLIP adaptation, fair multi-task learning, contextual bandits, and even behavioral or financial modeling: a bias embedded in an upstream representation, hypothesis, data-generating process, or shared training signal can persist, reappear, or be reinterpreted downstream, although the strength and stability of that transfer depend strongly on the adaptation mechanism and measurement regime (Salman et al., 2022).
1. Definition and conceptual scope
In its standard machine-learning formulation, BTH asks whether the bias of a pretrained model transfers to its adapted task-specific behavior. For causal LLMs, intrinsic bias denotes bias in the model’s original next-token generation behavior, while prompt-adapted bias denotes bias after zero-shot, few-shot, or related prompt-based adaptation on a downstream task (Mackraz et al., 2024). High alignment between these two bias types is treated as evidence that the same occupations, groups, or attributes remain biased in the same direction before and after adaptation (Sivakumar et al., 9 Sep 2025).
A central interpretive distinction in the recent LLM work is that BTH is treated as correlation, not causation. The 2025 study on prompt-adapted causal LLMs explicitly states that, despite the name, it makes no causal claim; instead, it measures whether intrinsic and prompted biases are aligned through the Pearson correlation between occupation-level intrinsic and prompted scores (Sivakumar et al., 9 Sep 2025). This framing makes BTH an empirical regularity about preservation of bias structure rather than a claim about mechanistic necessity.
Beyond LLMs, the hypothesis generalizes in several directions. In transfer learning for vision, BTH refers to the persistence of a source model’s sensitivity to spurious features or triggers after adaptation to a downstream task (Salman et al., 2022). In multilingual representation learning, it describes the carryover of embedding-space gender bias into cross-lingual downstream systems (Zhao et al., 2020). In fair multi-task learning, the term is used analogically: fairness improvements on one task can worsen another task’s fairness when shared parameters receive conflicting fairness gradients, so unfairness is effectively transferred across tasks (Roy et al., 2023). A plausible implication is that BTH is best understood as a family of transfer phenomena whose common structure is the persistence or re-expression of upstream bias under downstream reuse.
2. From fine-tuned MLMs to prompt-adapted causal LLMs
Recent work distinguishes sharply between two empirical regimes. Prior BTH literature mostly examined MLMs adapted by fine-tuning, often finding that intrinsic bias in pretrained MLMs does not correlate well with downstream bias after adaptation (Mackraz et al., 2024). That literature motivated the pessimistic conclusion that fairness in pretraining has limited effect on downstream fairness.
The more recent prompt-based literature argues that this conclusion is not universal. “Evaluating Gender Bias Transfer between Pre-trained and Prompt-Adapted LLMs” extends BTH to instruction-tuned causal models and reports that intrinsic biases in pretrained Mistral, Falcon, and Llama models are strongly correlated with biases after zero-shot and few-shot prompting on WinoBias, with overall values (Mackraz et al., 2024). “Bias after Prompting: Persistent Discrimination in LLMs” broadens the same argument across models, tasks, demographics, and prompting regimes, again concluding that prompt adaptation often preserves the sign and distribution of intrinsic bias rather than washing it out (Sivakumar et al., 9 Sep 2025).
This shift matters because prompting changes only the input, not the weights. The prompt-based studies therefore isolate a deployment regime in which bias transfer cannot be attributed to weight-update dynamics alone. This suggests that the weakness of BTH in earlier MLM fine-tuning studies was at least partly regime-specific rather than a general property of pretraining and adaptation (Sivakumar et al., 9 Sep 2025).
3. Operationalization in causal LLM studies
The prompt-adaptation literature uses a relatively unified evaluation setup. The core models include Llama 3 8B, Llama 3 70B, Falcon 40B, and Mistral 7B / 7B v0.3, with both instruction-tuned and, in some cases, non-instruction-tuned variants to isolate the effect of prompting (Sivakumar et al., 9 Sep 2025). The principal downstream settings are:
- Coreference resolution / gender bias using WinoBias, with 3,160 sentences, balanced by gender pronoun and split into Type 1 (ambiguous) and Type 2 (unambiguous) items (Sivakumar et al., 9 Sep 2025).
- Question answering / broader demographic bias using BBQ-lite, covering age, nationality, physical appearance, religion, socio-economic status, and sexual orientation (Sivakumar et al., 9 Sep 2025).
The prompting regimes comprise Intrinsic, Zero-shot prompting, Few-shot prompting, and Chain-of-Thought (CoT) prompting (Sivakumar et al., 9 Sep 2025). A major methodological contribution is systematic variation of few-shot composition: Number of shots $20, 40, 60, 80, 100$; Stereotypical makeup ; Occupational distribution ; and Representational balance (Sivakumar et al., 9 Sep 2025).
The key metrics are concise and task-specific. Occupation Selection Bias (O-SB) is the difference in generation rates for an occupation under male versus female pronouns; negative values indicate female-leaning bias and positive values male-leaning bias (Sivakumar et al., 9 Sep 2025). Aggregate Selection Bias (A-SB) averages absolute O-SB across occupations, with 0% is ideal (Sivakumar et al., 9 Sep 2025). Referential Prediction Accuracy (RPA) measures referent prediction performance, especially on Type 2 sentences (Sivakumar et al., 9 Sep 2025). The central BTH statistic is:
with strong correlation defined as and moderate correlation as (Sivakumar et al., 9 Sep 2025).
4. Empirical support in prompt-adapted LLMs
The most direct support for BTH in modern LLMs comes from the WinoBias results. Across models, prompt-adapted gender biases remain strongly correlated with intrinsic biases, with reported values typically (Sivakumar et al., 9 Sep 2025). The 2024 study likewise reports overall for zero-shot and few-shot prompting, with negligible $20, 40, 60, 80, 100$0-values, arguing that the same occupations that are biased in the base model remain biased after prompting (Mackraz et al., 2024). The preservation is not confined to standard prompting: CoT prompting also shows strong transfer, with reported values $20, 40, 60, 80, 100$1 (Sivakumar et al., 9 Sep 2025).
Several nuances recur across studies. Type 1 ambiguous sentences are generally more biased than Type 2 unambiguous ones (Sivakumar et al., 9 Sep 2025). Prompting often improves RPA, especially on pro-stereotypical items and male-pronoun cases, which indicates that higher apparent task performance can coexist with preserved unfairness (Sivakumar et al., 9 Sep 2025). Intrinsic biases are often higher in magnitude than prompted biases, but the orientation is preserved, meaning prompting may reduce amplitude without changing the ranking or sign pattern of the occupational bias landscape (Sivakumar et al., 9 Sep 2025).
The same pattern holds when isolating prompting from instruction tuning. For Mistral 7B v0.3, reported intrinsic-to-prompted correlations are $20, 40, 60, 80, 100$2 for zero-shot and $20, 40, 60, 80, 100$3 for few-shot; for Falcon 40B (non-IFT) they are $20, 40, 60, 80, 100$4 and $20, 40, 60, 80, 100$5, respectively (Sivakumar et al., 9 Sep 2025). Between base and instruction-tuned counterparts, bias magnitude may be somewhat reduced, but intrinsic bias patterns remain highly correlated, with Mistral: $20, 40, 60, 80, 100$6 and Falcon: $20, 40, 60, 80, 100$7 (Sivakumar et al., 9 Sep 2025). This indicates that instruction tuning does not eliminate bias transfer.
The question-answering experiments on BBQ-lite extend the result beyond gender. Reported correlations include Age: $20, 40, 60, 80, 100$8, Physical appearance: $20, 40, 60, 80, 100$9, Socio-economic status: 0, Religion: 1, Sexual orientation: 2, and Nationality: 3 to 4 depending on setup (Sivakumar et al., 9 Sep 2025). The interpretation given is that transfer is strong for some demographics, moderate for others, and broad enough that it cannot be treated as a purely gender-specific phenomenon.
Few-shot variation does not dissolve the pattern. Even when few-shot contexts are altered by sample size, stereotype composition, occupational distribution, or representational balance, bias magnitude can decrease but intrinsic–prompted correlations remain typically 5 (Sivakumar et al., 9 Sep 2025). The 2024 WinoBias study reports a similar robustness result for few-shot length and stereotypical composition, with 6 (Mackraz et al., 2024).
5. Mitigation attempts and qualified interpretations
Prompt-based mitigation methods have not consistently broken BTH in the causal-LLM setting. The 2025 prompting study evaluates In-line debiasing prompts inspired by Constitutional AI, Self-debiasing via explanation, Self-debiasing via reprompting, Thinking Fair and Slow with instruction-based and role-based pre-prompts, Prompting Fairness as a causality-based debiasing method, and two anti-stereotyping prompts labeled “anti-stereotyping all” and “anti-stereotyping most” (Sivakumar et al., 9 Sep 2025). Some methods reduce A-SB, some improve or preserve RPA, but none consistently reduce bias transfer across models, tasks or demographics (Sivakumar et al., 9 Sep 2025).
The results are therefore mixed rather than uniformly negative. Baseline intrinsic–prompted correlations are often around 7 to 8 (Sivakumar et al., 9 Sep 2025). The causality-based prompt can lower transfer into the moderate range, for example on Llama 3 8B to 9 (Sivakumar et al., 9 Sep 2025). The anti-stereotyping prompts can even flip correlation negative in specific cases, including Llama 3 8B: 0 or 1, Llama 3 70B: 2 or 3, and Mistral 7B: 4 or 5 (Sivakumar et al., 9 Sep 2025). However, these reversals are explicitly not presented as a general solution, but as a toy demonstration that explicit counter-stereotyping can reverse correlation in some settings (Sivakumar et al., 9 Sep 2025).
A recurring misconception is that successful prompt engineering is sufficient to neutralize upstream bias. The prompt-based BTH literature rejects that claim. The consistent finding is that prompting can change bias magnitude, but typically not the underlying pattern of which occupations or groups are favored or disfavored (Mackraz et al., 2024). This suggests that prompt-level mitigation alone may be structurally limited when the base model already encodes a stable bias geometry.
6. Related formulations across machine learning and adjacent fields
BTH has been instantiated in several neighboring research areas, often with different bias objects and different adaptation operators.
Selected BTH formulations across domains
| Domain | Upstream bias object | Downstream manifestation |
|---|---|---|
| Prompt-adapted causal LLMs | Intrinsic occupation-level or demographic bias | Prompted bias in WinoBias and BBQ-lite (Sivakumar et al., 9 Sep 2025) |
| Transfer learning in vision | Source sensitivity to triggers or spurious features | Residual sensitivity after transfer, measured by ASR (Salman et al., 2022) |
| Multilingual embeddings | Gender bias in aligned embedding spaces | Gender accuracy gaps in cross-lingual occupation prediction (Zhao et al., 2020) |
| Fair multi-task learning | Fairness-gradient conflict across tasks | Higher fairness violation than STL, i.e. 6 (Roy et al., 2023) |
| CLIP downstream adaptation | Pre-training social bias in CLIP backbones | Weak and inconsistent downstream correlation after adaptation (Ramos et al., 25 Aug 2025) |
| Hypothesis transfer in bandits | Source model parameter estimate | Biased regularization or weighted target estimate (Bilaj et al., 2022) |
The vision transfer-learning study “When does Bias Transfer in Transfer Learning?” provides a particularly clear mechanistic account. In a linearized setting, SGD updates during fine-tuning lie in the span of target data, so any source bias component orthogonal to that span cannot be changed by fine-tuning (Salman et al., 2022). Empirically, the paper shows that a source bias can survive transfer even when the target dataset is explicitly de-biased, especially in fixed-feature transfer, using attack success rate (ASR) to quantify residual trigger sensitivity (Salman et al., 2022). This is conceptually close to the causal-LLM prompting case, where adaptation also leaves the base representation largely intact.
The multilingual embeddings literature supports a refined BTH rather than a simple persistence law. Gender bias in multilingual fastText spaces transfers into downstream occupation prediction, but its magnitude depends on alignment direction and target space; aligning to gender-rich languages can reduce both intrinsic and extrinsic bias, whereas aligning to English often increases it (Zhao et al., 2020). This suggests that transfer can reshape bias rather than merely copy it.
By contrast, the CLIP literature offers a counterexample to any universal version of BTH. “From Global to Local: Social Bias Transfer in CLIP” studies 29 CLIP models and finds difficulty in discovering consistent trends in bias transfer to downstream VQA and captioning systems (Ramos et al., 25 Aug 2025). The strongest reported correlation is only Spearman’s 7 with 8, and a proposed explanation is that different CLIP backbones converge toward nearly identical representation spaces after downstream adaptation, reducing the explanatory role of pre-training bias (Ramos et al., 25 Aug 2025). The paper also shows that a model less biased globally can be more biased in a local subset such as the “guitars” group, indicating that bias-transfer conclusions depend strongly on whether measurement is global or local (Ramos et al., 25 Aug 2025).
Other fields use “bias transfer” more analogically but still preserve the core logic. In FairBranch, bias transfer in fair multi-task learning means a multi-task model can be more unfair on task 9 than the corresponding single-task learner, formalized as 0 (Roy et al., 2023). In contextual bandits, hypothesis transfer learning treats a source parameter vector as a useful or harmful bias that should influence the target learner only when aligned with the new task, via convex weighting or biased regularization (Bilaj et al., 2022). A plausible implication is that BTH can describe both descriptive persistence of unwanted bias and deliberate reuse of prior hypotheses, depending on whether the transferred structure is normatively harmful or instrumentally useful.
7. Interpretation, limitations, and broader significance
Across the literature, BTH is neither uniformly validated nor uniformly falsified. The strongest support currently comes from prompt-adapted causal LLMs, where intrinsic and prompted biases are repeatedly reported as strongly correlated across models, tasks, demographics, and prompting regimes (Sivakumar et al., 9 Sep 2025). These results directly challenge the earlier view, derived from MLM fine-tuning, that fairness in pretraining has limited downstream relevance (Mackraz et al., 2024). In this setting, the practical conclusion is clear: fairness in pre-trained models matters because prompt adaptation tends to preserve, not erase, their biases (Sivakumar et al., 9 Sep 2025).
At the same time, the broader literature imposes several qualifications. First, transfer depends on the adaptation operator: prompting, fixed-feature transfer, full-network fine-tuning, alignment, branch-specific gradient correction, and downstream CLIP adaptation do not behave identically (Salman et al., 2022). Second, transfer depends on the measurement regime: occupation-level bias correlation, ASR, 1, demographic disparity, spurious correlation, or local-cluster analysis can lead to different conclusions (Zhao et al., 2020). Third, the notion of “bias” itself varies by field, ranging from demographic unfairness to spurious feature reliance, model-induced artifacts, or fairness-gradient interference (Roy et al., 2023).
The main controversy is therefore not whether upstream structure can survive downstream reuse, but how broadly one can generalize from one adaptation regime to another. The recent LLM results invalidate the assumption that bias does not transfer under prompt adaptation, but the CLIP results show that pre-training bias is not always a dependable predictor of downstream bias (Sivakumar et al., 9 Sep 2025); (Ramos et al., 25 Aug 2025). This suggests that BTH should be treated as a conditional empirical hypothesis: strong in some architectures and adaptation pathways, weak or unstable in others.
A plausible synthesis is that BTH is most robust when downstream adaptation leaves the upstream representational structure largely intact, and weakest when adaptation induces strong convergence or overwriting. The literature on prompt-adapted causal LLMs, transfer learning with frozen or partially frozen backbones, and aligned multilingual embeddings is consistent with that interpretation (Sivakumar et al., 9 Sep 2025); (Salman et al., 2022); (Zhao et al., 2020). Under that reading, BTH has become a central organizing concept for understanding why downstream fairness cannot be studied independently of pretraining.