Counterfactual-Enhanced Debiasing Framework

• Counterfactual-Enhanced Debiasing Framework is a family of methods that treat bias as a causal phenomenon by generating counterfactual objects to adjust predictions.
• It employs diverse construction strategies—from GANs to sensitive-attribute flipping—to simulate alternative views and mitigate spurious correlations.
• The framework integrates preprocessing, in-processing, and post-processing operators, demonstrating empirical success across multimodal, medical imaging, and NLP tasks.

“Counterfactual-Enhanced Debiasing Framework” (Editor’s term) denotes a family of methods that treat bias as a causal phenomenon and then estimate, generate, or simulate counterfactual labels, predictions, embeddings, neighborhoods, or distributions in order to remove spurious effects while preserving task-relevant signal. Across the cited literature, this pattern appears in multimodal learning under presentation bias, multilingual masked language modeling, medical imaging, graph neural networks, recommendation, zero-shot vision–language recognition, summarization, machine unlearning, conformal counterfactual inference, and post-hoc group fairness (Lin et al., 2023, Liang et al., 3 Apr 2026, Kumar et al., 2023, Wo et al., 20 Aug 2025, Guo et al., 2023, Peng et al., 30 Oct 2025, Dong et al., 2023, Chen et al., 2024, Farzaneh et al., 4 Sep 2025, Arévalo et al., 8 Apr 2026). In each case, the common objective is not merely to regularize a predictor, but to intervene on a bias-inducing pathway and then train or calibrate on the resulting counterfactual view of the problem.

1. Causal formulation and estimands

A defining property of these frameworks is an explicit causal decomposition of the prediction problem. In multimodal learning under presentation bias, the relevant variables are simple features $X_t$, multimodal features $W_t$, labels $Y_t$, model outputs $R_0$, and exposure $A_t$, with the central feedback path $R_0 \rightarrow A \rightarrow Y$ inducing presentation bias because training at $t=1$ sees $P(X,W,Y \mid A=1)$ rather than the desired joint distribution (Lin et al., 2023). In audio-based depression detection, the causal graph contains gender $G$, acoustic cues $C$, mediator $W_t$0, and prediction $W_t$1, and the bias of interest is the direct path $W_t$2, while the mediator path $W_t$3 is retained as legitimate; the debiased quantity is the Total Indirect Effect,

$W_t$4

(Hu et al., 1 Dec 2025)

Other frameworks formalize the target through potential outcomes or interventional distributions. SP-CCI defines binary treatment $W_t$5, covariates $W_t$6, and potential outcomes $W_t$7, and asks for a set-valued predictor $W_t$8 satisfying $W_t$9 under SUTVA, ignorability, and overlap (Farzaneh et al., 4 Sep 2025). Representation-level calibration for zero-shot recognition frames object-context shortcuts through variables $Y_t$0 (object), $Y_t$1 (context), $Y_t$2 (learned interaction features), and $Y_t$3 (prediction), and then estimates a Total Direct Effect by subtracting background-only activation from class scores (Peng et al., 30 Oct 2025).

A notable feature of the literature is that not all methods require a full structural causal model. Fair-ICD uses search-based counterfactual neighbors with opposite sensitive attributes and explicitly states that it avoids unrealistic SCM/do-calculus assumptions by using nearest neighbors with opposite sensitive attributes (Wo et al., 20 Aug 2025). CAFP likewise uses a purely operational counterfactual in which the sensitive attribute is flipped while features are held fixed, with no SCM or learned generative model for $Y_t$4 (Arévalo et al., 8 Apr 2026). This suggests that the framework family ranges from explicit SCM-based interventions to operational counterfactual constructions that are nonetheless used to debias a prediction rule.

2. Counterfactual objects and construction strategies

The “counterfactual” in these frameworks is not a single object. In some cases it is a missing label. Counterfactual augmentation for presentation bias generates $Y_t$5 for samples with $Y_t$6 using a multimodal counterfactual GAN whose generator consumes tabular features, text, images, and recommendation condition, with DistilBERT and ViT encoders and condition-specific discriminators $Y_t$7 and $Y_t$8 (Lin et al., 2023). In SP-CCI, the counterfactual object is a synthetic outcome label $Y_t$9 sampled from a pre-trained counterfactual model $R_0$0 and inserted into a risk-controlling conformal calibration pipeline (Farzaneh et al., 4 Sep 2025).

In LLMs, the counterfactual often takes the form of a controlled text transformation. Multiple-Debias constructs multilingual counterfactual data augmentation by symmetric swaps of sensitive terms across English, German, Spanish, Chinese, and Japanese, such as “he $R_0$1 she,” “他 $R_0$2 她,” or “Jewish $R_0$3 Christian,” while preserving native Wikipedia context (Liang et al., 3 Apr 2026). ThaiFACTUAL creates factual–counterfactual pairs by entity swaps, sentiment-controlled paraphrases, negation-preserving rewrites, and attribute/value substitutions, then uses these pairs for post-hoc calibration without fine-tuning the base LLM (Sermsri et al., 26 Sep 2025).

In vision and multimodal settings, the counterfactual may be an image, an embedding, or a neighborhood. The medical-imaging framework combines Group-DRO with Cycle-GAN counterfactual image generation, using generators $R_0$4 and $R_0$5 and a classifier-consistency loss to test whether decision changes occur at confounder regions or pathology regions (Kumar et al., 2023). Fair-ICD constructs a counterfactual augmented graph by rerouting same-sensitive edges $R_0$6 to $R_0$7, where $R_0$8 is a nearest neighbor with the opposite sensitive attribute, thereby creating heterogeneous neighborhoods before message passing (Wo et al., 20 Aug 2025). Representation-level calibration for CLIP synthesizes counterfactual embeddings by estimating object and background expectations and recombining $R_0$9 with alternative contexts:

$A_t$0

(Peng et al., 30 Oct 2025)

Generative and recommendation settings instantiate yet another pattern. CCGM edits the causal graph itself by setting entries in the adjacency matrix to zero, thereby producing samples from a counterfactual SCM rather than merely intervening on a fixed graph (Bhat et al., 2022). DB-VAE uses Pearl’s abduction–action–prediction to generate counterfactual interaction sets $A_t$1, $A_t$2, and $A_t$3, augmenting sparse recommendation data with interactions corresponding to interventions on popularity and subjective bias (Guo et al., 2023).

3. Debiasing operators and learning objectives

Once counterfactual objects are constructed, the frameworks differ primarily in how they use them. In multimodal learning under presentation bias, the core operator is distributional augmentation:

$A_t$4

Training on $A_t$5 is intended to approximate the unbiased joint $A_t$6 rather than the observed $A_t$7 (Lin et al., 2023). In post-processing fairness, the operator is score symmetrization:

$A_t$8

which averages factual and counterfactual sensitive-attribute settings without retraining the classifier (Arévalo et al., 8 Apr 2026). In depression detection, the operator is counterfactual subtraction at the fused-logit level,

$A_t$9

so that the retained prediction corresponds to the mediator-based effect rather than the direct effect of gender (Hu et al., 1 Dec 2025).

This suggests that the framework family spans preprocessing, in-processing, and post-processing. Multiple-Debias explicitly organizes the pipeline as MCDA in pre-processing, PEFT fine-tuning in processing, and multilingual Self-Debias prompts in post-processing (Liang et al., 3 Apr 2026). CoFactSum uses two inference-time counterfactual estimators—Explicit Counterfactual Masking and Implicit Counterfactual Training—and then rescales their subtraction with a Debiasing Degree Adjustment mechanism at each decoding step (Dong et al., 2023). Fair-ICD combines counterfactual neighborhood construction with an adversarial discriminator trained by the min–max objective

$R_0 \rightarrow A \rightarrow Y$0

thereby offsetting structural bias before message passing and suppressing residual sensitive leakage afterward (Wo et al., 20 Aug 2025). DoubleGen modifies generative-model training objectives directly through a doubly robust loss of the form

$R_0 \rightarrow A \rightarrow Y$1

so that confounding correction is built into diffusion, flow-matching, or autoregressive learning itself (Luedtke et al., 20 Sep 2025).

Framework	Counterfactual object	Debiasing use
CA (Lin et al., 2023)	$R_0 \rightarrow A \rightarrow Y$2 for $R_0 \rightarrow A \rightarrow Y$3	Augment $R_0 \rightarrow A \rightarrow Y$4
Multiple-Debias (Liang et al., 3 Apr 2026)	Swapped multilingual text	MLM fine-tuning and MSD
Fair-ICD (Wo et al., 20 Aug 2025)	Opposite-sensitive neighbors	Heterogeneous message passing
CAFP (Arévalo et al., 8 Apr 2026)	Sensitive-attribute-flipped scores	Prediction averaging
SP-CCI (Farzaneh et al., 4 Sep 2025)	Synthetic counterfactual labels	Debiased conformal calibration

The table makes clear that “debiasing” is implemented through distinct operators—augmentation, subtraction, adversarial invariance, averaging, or conformal calibration—even when the causal motivation is similar.

4. Theoretical properties and guarantees

Several frameworks provide explicit identification or robustness claims. Counterfactual augmentation for presentation bias states that, under consistency, SUTVA, ignorability, and positivity, the exposure-adjusted outcome distribution is identifiable from observational data, and that if $R_0 \rightarrow A \rightarrow Y$5 is close to the true counterfactual distribution, then $R_0 \rightarrow A \rightarrow Y$6 approximates $R_0 \rightarrow A \rightarrow Y$7 (Lin et al., 2023). SP-CCI proves marginal coverage under exact weights and a robustness guarantee under approximate weights; its calibrated interval has the form

$R_0 \rightarrow A \rightarrow Y$8

with $R_0 \rightarrow A \rightarrow Y$9 chosen by a debiased risk estimate based on prediction-powered inference (Farzaneh et al., 4 Sep 2025).

DoubleGen contributes a different type of guarantee: doubly robust generative training. Its objective is constructed so that confounding bias is corrected even if only one of the nuisance models—the propensity or the outcome model—is correct, and the paper further states oracle optimality and minimax rate optimality under the stated regularity conditions (Luedtke et al., 20 Sep 2025). Deconfounding flow matching strengthens this distributional perspective by proving that observational and counterfactual outcome laws have identical support and tail behavior under positivity, remain statistically close under weak confounding, and admit semiparametrically efficient estimation through an efficient influence function correction (Dance et al., 8 May 2026).

Post-hoc fairness work emphasizes pointwise symmetry results. CAFP proves that, if $t=1$0 and $t=1$1, then the direct dependence on the protected attribute is eliminated under the stated independence assumptions, and the prediction distortion is exactly half the model’s own counterfactual bias:

$t=1$2

It also gives an equalized-odds difference bound in terms of average counterfactual bias (Arévalo et al., 8 Apr 2026). By contrast, Fair-ICD states that it “ensures the fairness of GNNs under moderate/mild conditions,” but also notes that no formal theorems, proofs, or explicit conditions are provided (Wo et al., 20 Aug 2025). The literature therefore ranges from heuristic causal design to full finite-sample guarantees.

5. Representative domains and empirical record

The empirical record is broad and strongly domain-specific. In multimodal learning under presentation bias, counterfactual augmentation improved binary-classification macro F1 by $t=1$3 and minority F1 by $t=1$4 on Airbnb, and by $t=1$5 and $t=1$6 on Clothing, relative to the next-best method on unbiased evaluation; in regression, Airbnb $t=1$7 increased from $t=1$8 to $t=1$9 and Clothing $P(X,W,Y \mid A=1)$0 from $P(X,W,Y \mid A=1)$1 to $P(X,W,Y \mid A=1)$2 (Lin et al., 2023). In multilingual masked LLMs, Multiple-Debias reduced average gender bias on CrowS-Pairs from $P(X,W,Y \mid A=1)$3 to $P(X,W,Y \mid A=1)$4 for mBERT with prompt-tuning and from $P(X,W,Y \mid A=1)$5 to $P(X,W,Y \mid A=1)$6 for XLM-R with adapter-tuning, while also reporting that multilingual methods generally reduce bias more effectively than monolingual ones (Liang et al., 3 Apr 2026).

In medical and affective computing, the frameworks focus on shortcut reliance and protected-attribute effects. In medical imaging, the Spurious Correlation Latching Score fell from $P(X,W,Y \mid A=1)$7 to $P(X,W,Y \mid A=1)$8 on Dataset 1 and from $P(X,W,Y \mid A=1)$9 to $G$0 on Dataset 2 when Group-DRO supervision replaced ERM, while counterfactual images shifted from confounder regions to pathology regions (Kumar et al., 2023). In depression detection, the NetVLAD-based backbone reached $G$1, $G$2, $G$3, and $G$4 under the counterfactual debiasing procedure, outperforming sub-sampling and MixFeat on DAIC-WOZ (Hu et al., 1 Dec 2025). In multimodal respiratory sound classification, BTS-CARD improved the OOD ICBHI-style score on SPRSound from $G$5 for BTS to $G$6, while also slightly improving in-distribution performance (Koo et al., 25 Oct 2025).

Graph, vision–language, summarization, and recommendation studies show the same pattern with different observables. On Pokec-n, Fair-ICD with a GCN backbone achieved $G$7, $G$8, and $G$9, improving both fairness and accuracy relative to vanilla and strong baselines (Wo et al., 20 Aug 2025). For zero-shot CLIP calibration on Waterbirds, worst-group accuracy rose from $C$0 to $C$1 with the external-context variant on ViT-B/32, and on COCO-GB v1 the gender gap dropped from $C$2 to $C$3 with the virtual-context variant (Peng et al., 30 Oct 2025). CoFactSum reported consistent gains in factual-consistency metrics on CNN/DailyMail and XSum, including QAFactEval $C$4 on CNN/DailyMail and $C$5 on XSum, while slightly lowering ROUGE-L relative to PEGASUS (Dong et al., 2023). In recommendation, DB-VAE improved Recall@20/NDCG@100 over Mult-VAE, RecVAE, MACR, and DecRS on ML-1M, ML-20M, AliShop-7C, and Amazon-Book, and the counterfactual data module helped especially in sparse regimes (Guo et al., 2023).

Additional domain-specific variants broaden the scope rather than altering the underlying logic. ThaiFACTUAL uses counterfactual augmentation and rationale-based supervision to reduce sentiment leakage and entity favoritism in Thai political stance detection, reporting Bias-SSC $C$6, RStd $C$7, macro-F1 $C$8, and OOD $C$9 (Sermsri et al., 26 Sep 2025). MME-JD for multimodal LLMs combines modality-specific counterfactuals with adaptive expert routing and reaches $W_t$00 and $W_t$01 on MMSD2.0 with Qwen2-VL (Wu et al., 18 Sep 2025). CED for target-oriented multimodal sentiment couples detail-matched image–text counterfactuals with adaptive contrastive debiasing and reports $W_t$02 and $W_t$03 on Twitter-2015 (Liu et al., 11 Sep 2025). These cases suggest that the framework has become a general design pattern rather than a single task-specific technique.

6. Limitations, controversies, and open directions

The main limitations are tied to the validity of the counterfactual object and the stability of the debiasing operator. Presentation-bias correction requires overlap and credible generation of $W_t$04; the paper explicitly notes that positivity can be violated when exposure is deterministically suppressed and that counterfactual generation may extrapolate poorly under extreme propensities (Lin et al., 2023). SP-CCI likewise emphasizes that treatment imbalance, approximate weights, and poor synthetic labels can widen margins or leave residual bias, and identifies conditional coverage as an open direction (Farzaneh et al., 4 Sep 2025). DoubleGen and deconfounding flows add robustness, but both rely on nuisance estimation, overlap, and stable transport geometry; the latter explicitly notes that formal debiasing of data-dependent OT couplings remains open (Luedtke et al., 20 Sep 2025, Dance et al., 8 May 2026).

A second class of limitations concerns semantic validity and proxy leakage. Multiple-Debias notes translation artifacts, morphological complexity in German and Spanish, and cultural specificity of stereotypes across languages (Liang et al., 3 Apr 2026). Fair-ICD requires sensitive-attribute labels and may struggle when opposite-sensitive similar nodes are rare in highly homophilous graphs (Wo et al., 20 Aug 2025). CAFP removes direct dependence on the protected attribute, but when $W_t$05 contains proxies for $W_t$06, indirect dependence can remain, so demographic parity and equalized odds may improve without becoming perfect (Arévalo et al., 8 Apr 2026). Machine-unlearning debiasing depends on the availability of semantically consistent counterfactual examples; if those counterfactuals are poor, the forgetting procedure can contaminate the remaining dataset rather than protect it (Chen et al., 2024).

There is also a methodological divide over what should count as a “counterfactual.” Some frameworks use explicit SCMs, path-specific effects, and potential outcomes; others use lexicon swaps, nearest neighbors, sensitive-attribute flips, or representation recombination without a full structural model. This suggests an ongoing tension between causal formalism and operational practicality. A plausible implication is that future work will continue to hybridize the two: stronger theoretical guarantees for practical post-hoc methods, and more operationally efficient implementations of fully identified causal estimands. Open extensions already named in the cited literature include multi-valued and continuous treatments, multi-attribute fairness, localized or conditional coverage, broader benchmarking beyond single datasets, and more faithful counterfactual generators in high-dimensional modalities (Farzaneh et al., 4 Sep 2025, Arévalo et al., 8 Apr 2026, Dance et al., 8 May 2026, Koo et al., 25 Oct 2025).