Robust Hallucination Detection in LLMs via Adaptive Token Selection
This addresses safety concerns for broader LLM deployment by improving detection of hallucinations in free-form text, though it is an incremental advance over prior token-based methods.
The paper tackles the problem of hallucination detection in large language models (LLMs) by proposing HaMI, a method that adaptively selects critical tokens for robust detection, and it significantly outperforms existing state-of-the-art approaches on four benchmarks.
Hallucinations in large language models (LLMs) pose significant safety concerns that impede their broader deployment. Recent research in hallucination detection has demonstrated that LLMs' internal representations contain truthfulness hints, which can be harnessed for detector training. However, the performance of these detectors is heavily dependent on the internal representations of predetermined tokens, fluctuating considerably when working on free-form generations with varying lengths and sparse distributions of hallucinated entities. To address this, we propose HaMI, a novel approach that enables robust detection of hallucinations through adaptive selection and learning of critical tokens that are most indicative of hallucinations. We achieve this robustness by an innovative formulation of the Hallucination detection task as Multiple Instance (HaMI) learning over token-level representations within a sequence, thereby facilitating a joint optimisation of token selection and hallucination detection on generation sequences of diverse forms. Comprehensive experimental results on four hallucination benchmarks show that HaMI significantly outperforms existing state-of-the-art approaches.