Geometry-Aware Hallucination Detection in Large Language Models

Researchers introduce GA-ICL, a geometry-aware framework that improves hallucination detection in large language models by selecting better in-context learning demonstrations. Rather than relying on surface-level text similarity, the method uses latent representations and prototype geometry to choose demonstrations, achieving stronger performance across factual verification and hallucination detection benchmarks while maintaining robustness across model scales.

Hallucination—the generation of plausible-sounding but factually incorrect information—represents one of the most significant limitations of contemporary large language models. This research addresses a critical gap in the current mitigation landscape by focusing on demonstration selection within in-context learning, a relatively underexplored lever for improving factual reliability. While prior approaches have tackled hallucinations through decoding modifications, retrieval augmentation, or model fine-tuning, those methods either require computational overhead or structural changes to deployed systems.

The geometry-aware approach represents a meaningful advancement in understanding how LLMs process information at the representational level. By moving beyond lexical similarity metrics to prototype-based selection grounded in latent geometry, GA-ICL captures meaningful semantic and structural relationships that surface-level heuristics miss. This aligns with broader trends in representation learning that emphasize the importance of manifold structure in high-dimensional spaces.

For developers and AI practitioners, GA-ICL offers a practical, training-light alternative for improving model reliability without parameter modification—a significant advantage for deployed systems where fine-tuning proves impractical. The framework's demonstrated robustness across temperature variations and different model architectures (Phi-14B, Qwen3-32B) suggests genuine generalizability rather than task-specific optimization.

The scaling behavior remains particularly noteworthy: while smaller models sometimes benefit from simpler lexical retrieval on QA tasks, larger models consistently favor geometry-aware selection. This pattern suggests that representational complexity increases with model scale, making sophisticated selection mechanisms increasingly valuable. Future work should examine whether similar geometric principles apply to other ICL-dependent tasks and whether the approach transfers to emerging model architectures.

• →GA-ICL improves hallucination detection by selecting in-context demonstrations based on learned prototype geometry rather than surface-level similarity.
• →The method demonstrates strong performance gains on dialogue and summarization tasks while maintaining robustness across temperature variations and model architectures.
• →Geometry-aware selection scales effectively to larger models, outperforming baseline approaches including on tasks where smaller models show limitations.
• →The approach requires no LLM parameter modification, making it practically deployable in existing systems without retraining.
• →Latent representation structure becomes increasingly important for demonstration selection as model scale increases.