Contrastive Attribution Scoring

• Contrastive attribution scoring is a framework that distinguishes observed outcomes by comparing them with well-chosen counterfactuals to enhance model interpretability.
• It employs methods such as latent contrastive projection, corpus similarity differences, and supervised contrastive losses to quantify feature contributions.
• Practical implementations like TRACE and WhosAI demonstrate high accuracy and robustness in source attribution and forensic analyses across modalities.

Contrastive attribution scoring is a principled methodological framework for explaining, detecting, and attributing outcomes or representations in machine learning models by explicitly contrasting alternatives. Unlike classic attribution, which explains why a model predicts what it does, contrastive attribution scoring focuses on identifying features, inputs, or patterns that distinguish an observed output from specific plausible counterfactuals or foils. It provides both theoretical clarity and empirical rigor for attribution in supervised, unsupervised, and causal modeling contexts and underpins modern approaches to source attribution in natural language processing, vision, and AI-generated content forensics.

1. Theoretical Foundations and Motivation

Contrastive attribution stems from the observation that human explanations are typically contrastive in nature: explanations answer "Why outcome $y$ rather than alternative $y'$?" rather than unconditionally "Why $y$?" In the machine learning context, this leads to methodologies that compute attribution scores not merely for features that support the observed output but for those that actively distinguish it from specified alternatives (Jacovi et al., 2021, Bertossi, 2023).

Formally, contrastive scoring operates over a choice of:

• Fact: The observed label or state (e.g., $y^*$).
• Foil: A specified alternative or counterfactual (e.g., $y'$), which could be another class, source, or generated output.
• Attribution: A quantitative score for each feature, factor, or subrepresentation reflecting its role in distinguishing the fact from the foil.

This structure enables fine-grained, cognitively aligned explanations and is compatible both with input-level interpretability and with attribution of higher-level semantic concepts.

2. Methodologies in Contrastive Attribution Scoring

Contrastive attribution scoring comprises several families of techniques, unified by the principle of measuring discriminative power with respect to specified alternatives. The dominant instantiations include:

a) Latent Contrastive Projection

A standard approach for neural classifiers is to compute the difference in logits or latent activations along the direction separating the class of interest from a foil. Given pre-final representation $h(x)$ and weight matrix $W$, project onto $u_{(y^*, y')} = W_{y^*} - W_{y'}$ to obtain the contrastive latent representation:

$$C(h_x)_{y^*,y'} = P_u h_x, \quad P_u = \frac{u u^\top}{\|u\|^2}$$

Feature-wise contrastive attribution is then quantified by comparing model outputs (or suitably normalized logits) before and after ablating or perturbing a candidate feature, both in original and contrastive-projected subspace (Jacovi et al., 2021).

b) Corpus- and Representation-Based Similarity Differences

For representation learning (e.g., in vision or unsupervised embedding models), contrastive attribution leverages the difference in average encoded similarity between the instance and:

• A reference corpus $C$ (group of interest)
• A foil set $y'$0 (contrasting group)

$y'$1

This scalar contrastive score can then be explained using post-hoc attribution methods such as gradients or SHAP, enabling contrastive corpus attribution (COCOA) (Lin et al., 2022).

c) Supervised Contrastive and Triplet Losses in Attribution

In supervised and self-supervised models, supervised contrastive objectives reshape embedding spaces to cluster facts and foils distinctly. The attribution task is resolved as nearest neighbor or centroid-based classification in the embedding space, where attribution scores are derived from local contrasts in embedding similarity:

$y'$2

At inference, the contribution of each feature or subregion to the contrastive similarity, and thus to source attribution, can be computed by evaluating or perturbing its effect on the score (Wang et al., 2024, Urueña et al., 20 Nov 2025, Cava et al., 2024).

d) Causal/Logical Counterfactuals with Responsibility and Shapley

Extending classical causal attributions, contrastive logic-based frameworks compute contrastive responsibility (Resp) or contrastive Shapley values for features by aggregating causal impacts across minimal interventions that flip the label from $y'$3 to $y'$4:

• For feature $y'$5 and minimal counterfactual $y'$6 producing $y'$7,

$y'$8

where $y'$9 is the set of features altered (Bertossi, 2023).

3. Practical Algorithms and Frameworks

Contrastive attribution scoring has seen broad adoption in source attribution, model debugging, and forensic scenarios. Prominent frameworks include:

TRACE (TRansformer-based Attribution using Contrastive Embeddings)

• Principal-sentence extraction per data source via TF–IDF ranking
• Fine-tuned SBERT encoder with a projection head, trained under supervised NT-Xent contrastive loss
• Attribution via $y$0NN (hard/soft) and centroid-based inference over normalized embeddings
• Robustness to moderate input perturbations and scalable to $y$1 distinct sources (Wang et al., 2024)

WhosAI

• BERT-based triplet (anchor, positive, negative) contrastive learning with dynamic margin
• Multi-similarity mining for informative hard positives and negatives
• Attribution by nearest-centroid classifier in embedding space, handles plug-in of new sources via new centroid computation without retraining (Cava et al., 2024)

Contrastive Corpus Attribution (COCOA)

• Post-hoc attribution of input features or regions by tracing contributions to contrastive corpus similarity $y$2
• Compatible with vanilla gradients, integrated gradients, SHAP, and occlusion-based methods
• Applied both to vision and mixed-modality (CLIP) embeddings (Lin et al., 2022)

Supervised Contrastive Open-Set Attribution

• Vision models (e.g., MambaVision-L3-256-21K backbone)
• Supervised contrastive embedding space followed by few-shot $y$3NN attribution
• Supports open-set evaluation and rapid onboarding of new generator classes (Urueña et al., 20 Nov 2025)

4. Empirical Evaluation and Performance

Contrastive attribution methods consistently demonstrate high accuracy and robustness across a range of tasks and modalities:

Framework	Domain / Task	Closed-Set Accuracy	Open-Set AUC/OSCR	Key Datasets
TRACE	LLM Source Attribution	84–97% (25 sources)	Graceful scaling	booksum, dbpedia_14, news
WhosAI	AI-Text Attribution & Detection	F1=0.999/0.990	Not reported	TuringBench (200K news)
COCOA	Representation Explanation	n/a (expl. metrics)	n/a	SimCLR, CLIP
SupCon-kNN	Vision Forensics	97.3%	96.1% / 85.1%	Custom generator splits

TRACE’s accuracy across 25–100 book sources degrades gracefully, e.g., top-1 drops from 84.4% (25 sources) to ~45–50% (100), with top-3/top-5 at ~75–80%. Text perturbations (synonym, deletion up to 15%) decrease performance by only 1–3% (Wang et al., 2024). WhosAI achieves F1>0.99 on both binary and multi-class authorship attribution (Cava et al., 2024). In vision, open-set attribution shows +14.7% and +4.3% improvements (AUC, OSCR) over prior art with minimal few-shot data (Urueña et al., 20 Nov 2025). COCOA demonstrates that attributions based on $y$4 explain model decisions under augmentations and cross-modal settings (Lin et al., 2022).

5. Interpretability, Scalability, and Robustness

Contrastive attribution is inherently interpretable due to its alignment with human-style, foil-based explanations. For example, both TRACE and COCOA provide nearest-neighbor evidence or feature attributions that directly explain why one outcome is preferred to another. Notably, most frameworks enable users to return the most relevant supporting or contrasting exemplars (e.g., sentences, images) along with similarity or attribution scores.

These techniques are scalable: centroid-based and $y$5NN approaches require only $y$6 or $y$7 operations for $y$8 sources and $y$9 memory bank entries, respectively. Hard positive/negative mining and batch computations can increase training cost, but inference is typically lightweight (Wang et al., 2024, Cava et al., 2024).

Robustness to domain shift and moderate input corruption is empirically validated. For instance, TRACE shows minimal drop in attribution accuracy under token deletion or synonym substitution, and WhosAI maintains attribution clusters under corpus- or model-wise augmentations (Wang et al., 2024, Cava et al., 2024).

6. Limitations and Future Directions

Despite their strengths, contrastive attribution methods inherit several challenges:

• Foil Choice: Quality and diagnostic value depend critically on foil selection—poorly chosen alternatives can yield uninformative or misleading attributions (Lin et al., 2022, Jacovi et al., 2021).
• Dependency on Representation Quality: If the encoder does not capture the relevant semantics, contrastive attribution cannot recover them (Lin et al., 2022).
• Computational Complexity: Rich causal or counterfactual attributions (Resp, Shap) are NP-/#P-hard in general, though tractable for restricted model classes like deterministic Boolean circuits or shallow trees (Bertossi, 2023).
• Granularity: Most contemporary methods contrast pairs of classes or sources; extensions to contrast entire sets of alternatives (foil sets) are underexplored (Jacovi et al., 2021).
• Open-World Generalization: Embedding and memory-based models approximate open-set attribution but can struggle when novel classes are very close to existing clusters (Urueña et al., 20 Nov 2025, Cava et al., 2024).

Areas of active and suggested research include automatic foil/corpus discovery, richer non-linear or structured-output contrasts, and formal calibration of similarity-based confidence scores (Lin et al., 2022, Jacovi et al., 2021, Urueña et al., 20 Nov 2025). Extensions to scalable causal/probabilistic attributions and counterfactual optimization remain open problems for advancing contrastive explainability.

7. Applications and Impact in Research and Practice

Contrastive attribution scoring undergirds a wide array of applications:

• Source Attribution in LLMs: Assigning factual supports or sources to generated text for compliance and transparency (Wang et al., 2024).
• Detection and Attribution of AI-Generated Content: Distinguishing human and synthetic texts or images, even in few-shot or open-set regimes (Cava et al., 2024, Urueña et al., 20 Nov 2025).
• Model Interpretability and Debugging: Identifying discriminative factors for model errors or biases, down to token or conceptual level (Jacovi et al., 2021, Bertossi, 2023).
• Semantic and Representation Explanation: Zero-shot object localization, augmentation robustness, and multimodal grounding in unsupervised or vision-LLMs (Lin et al., 2022).

By centering explanation on contrast and causality, these approaches set a rigorous, scalable foundation for attribution, model governance, and forensic applications across modalities and architectures.