---
title: 'Bias Transfer Hypothesis (BTH): Implications in AI'
url: https://www.emergentmind.com/topics/bias-transfer-hypothesis-bth
type: topic
---

# Bias Transfer Hypothesis (BTH): Implications in AI

Searching arXiv for recent 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 [2509.08146]. Earlier BTH work often reported weak or absent transfer for **masked language models (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 [2412.03537]. 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 [2207.02842].

## 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 language models, **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 [2412.03537]. 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 [2509.08146].

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 [2509.08146]. 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 [2207.02842]. In multilingual representation learning, it describes the carryover of embedding-space gender bias into cross-lingual downstream systems [2005.00699]. 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 [2310.13746]. 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 [2412.03537]. 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 Language Models” 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 **\(\rho \ge 0.94\)** [2412.03537]. “Bias after Prompting: Persistent Discrimination in Large Language Models” 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 [2509.08146].

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 [2509.08146].

## 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 [2509.08146]. 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 [2509.08146].
- **Question answering / broader demographic bias** using **BBQ-lite**, covering **age**, **nationality**, **physical appearance**, **religion**, **socio-economic status**, and **sexual orientation** [2509.08146].

The prompting regimes comprise **Intrinsic**, **Zero-shot prompting**, **Few-shot prompting**, and **Chain-of-Thought (CoT) prompting** [2509.08146]. A major methodological contribution is systematic variation of few-shot composition: **Number of shots** \(20, 40, 60, 80, 100\); **Stereotypical makeup** \(neutral, anti-stereotypical, pro-stereotypical\); **Occupational distribution** \(in-distribution WinoBias occupations or out-of-distribution occupations from Winogender\); and **Representational balance** \(equally sampled occupations or reweighted by the model’s own intrinsic O-SB\) [2509.08146].

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 [2509.08146]. **Aggregate Selection Bias (A-SB)** averages absolute O-SB across occupations, with **0% is ideal** [2509.08146]. **Referential Prediction Accuracy (RPA)** measures referent prediction performance, especially on Type 2 sentences [2509.08146]. The central BTH statistic is:

$$
\rho = \text{PearsonCorr}(\text{O-SB}_{intrinsic}, \text{O-SB}_{prompted})
$$

with **strong correlation** defined as \(\rho \ge 0.7\) and **moderate correlation** as \(0.7 > \rho \ge 0.40\) [2509.08146].

## 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 **\(\rho \ge 0.94\)** [2509.08146]. The 2024 study likewise reports **\(\rho \ge 0.94\)** overall for zero-shot and few-shot prompting, with negligible \(p\)-values, arguing that the same occupations that are biased in the base model remain biased after prompting [2412.03537]. The preservation is not confined to standard prompting: **CoT prompting** also shows strong transfer, with reported values **\(\rho \ge 0.97\)** [2509.08146].

Several nuances recur across studies. **Type 1 ambiguous sentences** are generally **more biased** than **Type 2 unambiguous** ones [2509.08146]. 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 [2509.08146]. 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 [2509.08146].

The same pattern holds when isolating prompting from instruction tuning. For **Mistral 7B v0.3**, reported intrinsic-to-prompted correlations are **\(\rho = 0.98\)** for zero-shot and **\(\rho = 0.96\)** for few-shot; for **Falcon 40B (non-IFT)** they are **\(\rho = 0.96\)** and **\(\rho = 0.97\)**, respectively [2509.08146]. Between base and instruction-tuned counterparts, bias magnitude may be somewhat reduced, but intrinsic bias patterns remain highly correlated, with **Mistral: \(\rho = 0.99\)** and **Falcon: \(\rho = 0.98\)** [2509.08146]. 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: \(\rho \ge 0.98\)**, **Physical appearance: \(\rho \ge 0.79\)**, **Socio-economic status: \(\rho \ge 0.99\)**, **Religion: \(\rho \ge 0.69\)**, **Sexual orientation: \(\rho \ge 0.47\)**, and **Nationality: \(\rho \approx 0.42\) to \(0.66\)** depending on setup [2509.08146]. 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 **\(\rho \ge 0.90\)** [2509.08146]. The 2024 WinoBias study reports a similar robustness result for few-shot length and stereotypical composition, with **\(\rho \ge 0.97\)** [2412.03537].

## 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” [2509.08146]. Some methods reduce **A-SB**, some improve or preserve **RPA**, but **none consistently reduce bias transfer across models, tasks or demographics** [2509.08146].

The results are therefore mixed rather than uniformly negative. Baseline intrinsic–prompted correlations are often around **\(\rho \approx 0.96\) to \(0.98\)** [2509.08146]. The causality-based prompt can lower transfer into the **moderate** range, for example on **Llama 3 8B** to **\(\rho = 0.69\)** [2509.08146]. The anti-stereotyping prompts can even flip correlation negative in specific cases, including **Llama 3 8B: \(\rho = -0.62\) or \(-0.47\)**, **Llama 3 70B: \(\rho = -0.80\) or \(-0.74\)**, and **Mistral 7B: \(\rho = -0.56\) or \(-0.62\)** [2509.08146]. 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 [2509.08146].

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 [2412.03537]. 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 [2509.08146] |
| Transfer learning in vision | Source sensitivity to triggers or spurious features | Residual sensitivity after transfer, measured by ASR [2207.02842] |
| Multilingual embeddings | Gender bias in aligned embedding spaces | Gender accuracy gaps in cross-lingual occupation prediction [2005.00699] |
| Fair multi-task learning | Fairness-gradient conflict across tasks | Higher fairness violation than STL, i.e. \(DG(t) > 0\) [2310.13746] |
| CLIP downstream adaptation | Pre-training social bias in CLIP backbones | Weak and inconsistent downstream correlation after adaptation [2508.17750] |
| Hypothesis transfer in bandits | Source model parameter estimate | Biased regularization or weighted target estimate [2211.07387] |

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 [2207.02842]. 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 [2207.02842]. 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 [2005.00699]. 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 [2508.17750]. The strongest reported correlation is only **Spearman’s \(\rho = 0.61\)** with **\(p = 0.06\)**, 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 [2508.17750]. 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 [2508.17750].

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 \(t\) than the corresponding single-task learner, formalized as **\(DG(t) > 0\)** [2310.13746]. 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 [2211.07387]. 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 [2509.08146]. These results directly challenge the earlier view, derived from MLM fine-tuning, that fairness in pretraining has limited downstream relevance [2412.03537]. In this setting, the practical conclusion is clear: **fairness in pre-trained models matters because prompt adaptation tends to preserve, not erase, their biases** [2509.08146].

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 [2207.02842]. Second, transfer depends on the **measurement regime**: occupation-level bias correlation, ASR, \(|\mathrm{Diff}|\), demographic disparity, spurious correlation, or local-cluster analysis can lead to different conclusions [2005.00699]. Third, the notion of “bias” itself varies by field, ranging from demographic unfairness to spurious feature reliance, model-induced artifacts, or fairness-gradient interference [2310.13746].

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 [2509.08146]; [2508.17750]. 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 [2509.08146]; [2207.02842]; [2005.00699]. Under that reading, BTH has become a central organizing concept for understanding why downstream fairness cannot be studied independently of pretraining.

Source: https://www.emergentmind.com/topics/bias-transfer-hypothesis-bth