SemGrad: Gradient-Based Uncertainty Quantification for LLMs
A team of researchers has introduced SemGrad, the inaugural gradient-based method for uncertainty quantification (UQ) in free-form text generation using large language models (LLMs). Unlike traditional methods that rely heavily on sampling, SemGrad offers a computationally efficient, sampling-free alternative. It functions within semantic space, assessing the stability of output distributions in response to semantically similar input variations. The approach features a Semantic Preservation Score (SPS) to pinpoint embeddings that accurately reflect semantics, serving as a basis for gradient calculations. This innovation seeks to reduce the computational burden and variability associated with current UQ methods, enhancing reliability by identifying hallucinations. The findings are available on arXiv with the identifier 2605.04638.
Key facts
- SemGrad is the first gradient-based UQ method for free-form generation.
- It is sampling-free and computationally efficient.
- It operates in semantic space, not parameter space.
- It measures stability under semantically equivalent input perturbations.
- It introduces a Semantic Preservation Score (SPS).
- It addresses high computational cost and variance of sampling-based methods.
- It aims to improve trustworthiness of LLMs by detecting hallucinations.
- Published on arXiv with identifier 2605.04638.
Entities
Institutions
- arXiv