HALLUCINATIONS IN LARGE LANGUAGE MODELS (LLM’S): CHALLENGES IN MITIGATION, TRUST, AND FUTURE DIRECTIONS

Abstract

Large Language Models (LLMs) demonstrate impressive language abilities but frequently generate hallucinations - coherent, yet false or unsubstantiated outputs. This paper interrogates why hallucination remains common in LLMs as they currently exist, providing a review of strategies for mitigation, uncertainty estimation, and trustworthiness frameworks, and asserts that hallucination is an embedded emergent phenomenon of contemporary model architecture. We assessed data-driven, model-architecture, and alignment approaches and highlighted their effectiveness, as well as the trade-offs each offers. We also explored uncertainty estimation and showed how LLMs are often very confident in their false statements. We mapped the dimensions of trustworthiness (e.g., reliability, safety, fairness, etc.) to modes of hallucination and demonstrated large gaps in operationalizable recommendations (see Table 2). In our empirical research on models such as GPT-3/InstructGPT, we highlighted that while large & aligned LMs may be effective, they are also prone to significant levels of hallucination. We put together these threads into an argument that hallucination stems from base factors - noisy web-scale data, next-token objectives, and limited reasoning capabilities. Finally, we set out paradigm-shifting research directions (e.g., neuro-symbolic, causal, meta-cognitive) that may be needed to shift beyond the hallucination-laden paradigm. This wide-ranging study makes clear that substantial and systematic interdisciplinary work is required to raise the trustworthiness of LLMs.

Topics & Concepts