- The paper proposes CCHD, a constrained optimization approach that enforces prediction consistency between original and paraphrased claims for enhanced hallucination detection.
- It utilizes both paraphrase-consistency and label-preservation constraints with a gradient descent-ascent method, achieving up to +1.26 F1 improvement on benchmark datasets.
- CCHD is scalable and inference-efficient, relying on training-time back-translation for paraphrasing and supporting multiple paraphrase strategies without extra runtime cost.
Constrained Paraphrase Consistency for LLM Hallucination Detection: A Technical Analysis
Motivation and Background
Hallucination detection in LLM-generated text, notably in tasks such as abstractive summarization, is a persistent challenge, with prior work documenting up to 30% factual inconsistencies in generated content. Existing solutions frequently rely on either computationally expensive LLM-based evaluators or lightweight discriminative detectors trained on extensive annotated or synthetically augmented datasets. However, these approaches suffer from inefficiency, potential bias, and underutilization of implicit consistency present in semantically equivalent paraphrases.
The paper introduces Consistency-Constrained Hallucination Detector (CCHD), reframing hallucination detection as a constrained optimization problem. The core premise is that a claim and its semantically equivalent paraphrases should exhibit prediction-level consistency regarding factual support from a given document. Current discriminative detectors do not explicitly enforce such consistency and treat paraphrased and original samples identically, thus failing to exploit structure in the training data.
Methodology
Constrained Optimization Framework
CCHD structures training as a constrained optimization problem. The primary objective is the cross-entropy loss on original document-claim pairs. Two families of constraints are imposed:
- Paraphrase-consistency constraints bound the divergence (using Jeffreys divergence) between prediction distributions on original and paraphrased views.
- Label-preservation constraints ensure paraphrased views maintain alignment with the ground-truth label, with controllable slack for paraphrase noise.
The optimization problem is solved via gradient descent-ascent (GDA) over model parameters and per-view Lagrange multipliers. This primal-dual approach allows adaptive penalization, directly enforcing consistency and supervision balance, using only a few additional scalar dual variables. The architecture of base detectors (e.g., DeBERTa, Flan-T5) remains unchanged, and the method induces no inference-time overhead as paraphrase generation is confined to training.
Paraphrase Generation and Constraint Instantiation
To instantiate paraphrase views, CCHD adopts back-translation, leveraging MT systems to create paraphrases that widely vary stylistically while preserving semantics. This strategy avoids costly LLM-based paraphrasers and additional annotation. The framework is agnostic to backbone choice and supports extensions such as multiple pivot languages and alternative paraphrasing strategies, simply increasing the number of dual multipliers.
Both prediction-level and embedding-level constraints are supported. Prediction-level constraints regularize output distribution divergence, while embedding-level constraints penalize distance in learned representation space.
Empirical Evaluation
CCHD is evaluated on the LLM-AggreFact benchmark, encompassing 11 factuality and grounded consistency datasets. The method is instantiated with both DeBERTa and Flan-T5 backbones. Macro-averaged F1 scores demonstrate that CCHD-DBT and CCHD-FT5 outperform all baseline detectors, including FactCG, MiniCheck, AlignScore, and FactCC, with gains of up to +1.26 F1 points over the strongest non-CCHD baseline. Notably, CCHD achieves the highest F1 on 7 of 11 tasks, highlighting robustness and cross-task generalization.
Ablation Studies and Sensitivity Analysis
Ablation studies validate the necessity of both constraint families and adaptive GDA optimization. CCHD achieves the highest macro-F1, outperforming variants using only paraphrase-consistency, only label-preservation, and fixed dual multipliers. Sensitivity analysis reveals that paraphrase pivot language and constraint type affect performance. French-to-English back-translation with prediction-level constraints yields the strongest average results, likely due to syntactic similarity and robust MT systems. Embedding-level constraints provide complementary gains on structure-sensitive tasks but generally underperform compared to prediction-level counterparts.
Implications and Future Directions
Practical Implications
CCHD offers a scalable and efficient solution for hallucination detection in LLMs, requiring no inference-time paraphrasing and no additional annotation. Its constrained training strategy enhances label efficiency and model reliability, making it suitable for deployment in high-throughput factuality detection pipelines. The method is extensible, allowing for adaptation to additional pivots, paraphrasers, and constraint types.
Theoretical Implications
Formulating hallucination detection as a constrained optimization problem enables principled enforcement of prediction-level agreement across paraphrastic views. This challenges prevailing practices in data augmentation and regularization, suggesting that soft-consistency constraints may be a generalizable technique for other discriminative tasks where semantic invariance is expected.
Future Directions
Potential future enhancements include integration of multiple pivot languages, adaptive paraphrase selection, and evaluation on hallucination detection in multimodal LLMs, as factual consistency across modalities becomes increasingly relevant. The theoretical machinery of primal-dual optimization could be further explored for complex constraint hierarchies and multimodal consistency regularization.
Conclusion
CCHD establishes a new discriminative framework for LLM hallucination detection, leveraging constrained optimization to enforce cross-paraphrase consistency and label preservation. Empirical results show substantial improvements in factuality detection performance, with no runtime cost at inference. CCHD's methodological innovations open pathways for improved robustness and generalization in factuality assessment, with broad applicability across generative AI evaluation domains.
(2606.08158)