Semantic Consistency as an Unsupervised Hallucination Signal in Large Language Models

Charan Sai Ponnada

venue IEEE International Conference on Computing, Communication and Intelligent Systems (InCODE) 2026

date April 2026

status in-review

abstract

This work proposes paraphrase variance across K=5 semantic-preserving rewrites as a label-free, unsupervised proxy for hallucination in large language model outputs, requiring no ground-truth answers at inference time. The signal is evaluated using BERTScore, NLI contradiction rate and AUC-ROC on the TriviaQA and Natural Questions benchmarks. Experiments run on Llama-3-8B-Instruct (BF16, FlashAttention-2) on 2x NVIDIA L40S, and extend to Llama-3-70B for a cross-scale validation claim.

at a glance

Paraphrases — K = 5 semantic-preserving rewrites
Benchmarks — TriviaQA, Natural Questions
Models — Llama-3-8B-Instruct, Llama-3-70B
Metrics — BERTScore, NLI contradiction rate, AUC-ROC

keywords

Hallucination Detection, Large Language Models, Semantic Consistency, Unsupervised Evaluation, NLI, BERTScore, AI Safety

bibtex

@inproceedings{ponnada2026semantic, title={Semantic Consistency as an Unsupervised Hallucination Signal in Large Language Models}, author={Ponnada, Charan Sai}, booktitle={Proceedings of IEEE InCODE 2026}, year={2026} }