Hybrid Attribution Priors for Explainable and Robust Model Training

1 INTRODUCTION

• Self-derived priors can inherit model bias, whereas human-annotated priors offer transparency and generalizability but are expensive to construct and difficult to scale.Source: §1 · pp. 1–2
• The introduction asks whether existing methods have systematic limitations, how to extract and combine high-quality LLM priors, and whether refined priors support interpretability and robust supervision across settings.Source: §1 · p. 2
• LIME and SHAP highlight label-relevant tokens but often emphasize shared keywords rather than cues that distinguish semantically similar classes, potentially amplifying their errors.The introduction illustrates this with confusion between “set alarm” and “query alarm.”Source: §1 · p. 2
• Across full-data, few-shot, and adversarial settings, the refined priors are reported to improve classification accuracy, interpretability, and robustness to adversarial perturbations.The method is evaluated across three diverse datasets.Source: §1 · p. 2
• CAP enriches text with task instructions and label space to produce salient, discriminative priors, while CAP Hybrid combines CAP with existing attribution methods.The model’s self-attributions are aligned with the enriched priors during training.Source: §1 · pp. 2–3

2.1 MODEL ROBUSTNESS AND INTERPRETABILITY

• Integrated gradients estimates feature contributions along a baseline-to-input path, while LIME uses masked perturbations and a locally weighted surrogate; SHAP assigns feature importance within a unified framework.Source: §2.1 · p. 3

2.2 ATTRIBUTION-DRIVEN LEARNING

• Prior attribution-guided learning includes gradient regularization, explanation distillation, human rationale supervision, and multi-view attribution distillation, but some approaches depend on costly annotations or parameter updates.The paper proposes automatically extracting comprehensive priors from LLMs without human effort or model fine-tuning.Source: §2.2 · p. 3

3 ANALYSIS OF ATTRIBUTION PRIORS

• The analysis examines attribution methods used to generate priors for small language models in classification tasks.Source: §3 · p. 3

3.0.1 INTEGRATED GRADIENT

• IG priors are extracted from a trained model and can align with its decisions, but inherit its biases and may miss semantic label information, especially in few-shot or fine-grained settings.IG captures local gradient-based information.Source: §3.0.1 · pp. 3–4

3.0.2 LIME AND SHAP

• LIME and SHAP highlight label-affine words and produce highly correlated scores (Pearson correlation 0.8); LIME also reveals substantial keyword overlap among some classes.The authors call the overlapping classes “confused classes.”Source: §3.0.2 · p. 4
• Under standard training, most misclassifications occur in regions with high LIME keyword overlap, motivating priors that capture both label affinity and discriminative features.Figure 2 describes the distribution of wrong predictions across overlap levels.Source: §3.0.2 · p. 4, Figure 2 · p. 4

4 THE PROPOSED METHOD

• The framework extracts class-aware LLM priors, fuses them with existing attribution signals, and aligns target-model attributions with the fused priors.The paper describes this process as supporting more faithful and robust training.Source: §4 · pp. 4–5

4.1 CLASS-AWARE ATTRIBUTION PRIORS

• CAP prompts an LLM with task instructions and label space, samples word masks, scores masked variants by true-label probability, and fits per-word attribution scores.Masking samples multiple perturbation patterns rather than evaluating all 2^d possibilities.Source: §4.1 · p. 5

4.2 HYBRID ATTRIBUTION PRIORS

• The hybrid prior normalizes and combines LIME, CAP, and IG signals using word-wise mean or max aggregation.LIME captures target-label affinity, CAP emphasizes class-boundary terms, and IG captures model sensitivity.Source: §4.2 · p. 6

4.3 EXPLANATION-GUIDED LEARNING

• Training aligns normalized external priors with target-model IG self-attributions using MSE, while jointly optimizing attribution guidance and cross-entropy classification loss.The hyperparameter β balances attribution guidance against the standard classification objective.Source: §4.3 · p. 6

5.1 DATASETS AND BASELINES

• The evaluation covers HWU64, Banking77, and Clinc150, comparing Integrated Gradients and LIME as explanation-guided learning baselines.The experiments report accuracy, comprehensiveness, and sufficiency across full-data, few-shot, and adversarial settings.Source: §5.1 · p. 7, §5.3 · p. 7

5.2 ADVERSARIAL DATASET CONSTRUCTION

• Adversarial test examples are generated by adding or replacing keywords from another class while retaining the original label.Target classes are selected using CAP-derived class keyword lists and keyword-distribution overlap; the sets are generated with GPT-4o.Source: §5.2 · p. 7

5.3 IMPLEMENTATION DETAILS

• RoBERTa-Base is used for training and IG extraction, while Llama-3-8B-Instruct supplies LIME and CAP attribution scores.Comprehensiveness measures confidence drop when the rationale is removed; sufficiency measures how well the rationale alone supports the prediction, with smaller sufficiency scores indicating a minimal confidence drop.Source: §5.3 · p. 7

5.4.1 FEW-SHOT RESULTS

• 79.09% accuracy on HWU64 was achieved by CAPHybrid in few-shot training, alongside comprehensiveness of 69.12% and sufficiency of 30.66%.CAPHybrid also improved adversarial accuracy by 8% on HWU64 and 4% on Banking77.Source: Table 1 · p. 8, §5.4.1 · p. 8

5.4 RESULT ANALYSIS

• The 5-shot results report accuracy, comprehensiveness, and sufficiency across three datasets, with hybrid size indicating how many attribution sources are combined.The reported single-prior methods outperform the RoBERTa-Base baseline across all evaluation metrics, while IG, LIME, and CAP show different strengths.Source: Table 1 · p. 8, §5.4.1 · p. 8

5.4.2 FULL TRAINING RESULTS

• 93.51% test accuracy and 84.97% adversarial accuracy were achieved by CAPHybrid on full-data Banking77, versus 92.63% and 70.29%, respectively.Its comprehensiveness was 68.21% versus 56.73%; sufficiency was 23.98% versus 23.98%.Source: §5.4.2 · p. 8, Table 2 · p. 9

5.5 QUALITATIVE ANALYSIS OF DATASET PROPERTIES AND METHOD EFFECTIVENESS

• At difficulty levels 3 and 4, attribution-guided training was particularly effective on challenging examples with high token-level similarity between confusing classes.The analysis treats lexical overlap between ground-truth labels and original incorrect predictions as a proxy for instance difficulty.Source: §5.5 · p. 9
• Banking77 and HWU64 contain clusters of highly similar classes, whereas Clinc150 has more uniformly distributed, better-separated class representations.The paper reports marginal gains on Clinc150, where simple LIME priors suffice to provide meaningful guidance.Source: §5.5 · p. 9

• Surface-level priors can fail to distinguish fine-grained classes when categories have high semantic overlap.Source: §6 · p. 10
• CAP uses task-specific, label-informed semantic constraints, while CAPHybrid fuses complementary attribution sources into richer, more balanced priors.Source: §6 · p. 10
• Across datasets, the paper reports that its method enhances performance, interpretability, and robustness.Source: §6 · p. 10

• HWU64 contains 64 intents from 18 everyday domains, with utterances balanced across intents.Source: §A.2 · p. 12
• Banking77 contains 13,083 labeled customer utterances spanning 77 fine-grained banking intents.Source: §A.2 · p. 12
• Clinc150 contains 22,500 utterances evenly distributed across 150 intents in 10 general domains.Source: §A.2 · p. 12
• Integrated Gradients and LIME are selected as explanation-guided learning baselines, using model-derived and LLM-derived attribution scores, respectively.IG scores are computed from a trained model and used to supervise a new model; LIME scores from a large language model guide the smaller model.Source: §A.2 · p. 12

• The appendix derives Equation (3) by differentiating the stated objective and transforming its left-hand side.Source: §A.3 · pp. 12–13
• The attribution parameter vector α is estimated from mask vectors m_i with associated scores z_i, using regularization coefficient λ > 0.Source: §A.3 · p. 13
• For the transformation, the proof sets t = α^T m_i, a scalar, and uses t m_i = m_i t and α^T m_i = m_i^T α.Source: §A.3 · p. 13

A.4 PROOF OF POSITIVE DEFINITE MATRIX A

• A = M^T M + λI is symmetric and positive definite when λ > 0.The section states that I is the identity matrix.Source: §A.4 · p. 13
• For nonzero real column vector x, the M-term is nonnegative and λ||x||2 is positive, establishing positive definiteness.The proof concludes that A is positive definite.Source: §A.4 · p. 13
• The proof evaluates the matrix property using an arbitrary nonzero real column vector x.Source: §A.4 · p. 13

A.5 TRAINING PARAMETER SETTING

• CAP priors use λ = 0.1 and n = 100; RoBERTa-Base uses learning rate 5e-5, batch size 16, patience 10, gradient clipping 1.0, and attribution loss weight β = 1.For hybrid priors, the reported test results use the mean or max aggregation combination with the best validation performance.Source: §A.5 · p. 14

A.6 THE COMPARSION OF DIFFERENT AGGREGATION OPERATION

• Mean aggregation achieves the best predictive performance in most cases among the compared accuracy aggregation strategies.Source: §A.6 · p. 14

A.7 ADVERSARIAL DATASET GENERATION PROMPT

• The appendix presents adversarial dataset-generation prompts, including a Keyword Replacement prompt template in Table 5.Source: §A.7 · p. 14, Table 5 · p. 15