You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Large Language Models (LLMs) trained on extensive textual corpora haveemerged as leading solutions for a broad array of Natural Language Processing(NLP) tasks. Despite their notable performance, these models are prone tocertain limitations such as misunderstanding human instructions, generatingpotentially biased content, or factually incorrect (hallucinated) information.Hence, aligning LLMs with human expectations has become an active area ofinterest within the research community. This survey presents a comprehensiveoverview of these alignment technologies, including the following aspects. (1)Data collection: the methods for effectively collecting high-qualityinstructions for LLM alignment, including the use of NLP benchmarks, humanannotations, and leveraging strong LLMs. (2) Training methodologies: a detailedreview of the prevailing training methods employed for LLM alignment. Ourexploration encompasses Supervised Fine-tuning, both Online and Offline humanpreference training, along with parameter-efficient training mechanisms. (3)Model Evaluation: the methods for evaluating the effectiveness of thesehuman-aligned LLMs, presenting a multifaceted approach towards theirassessment. In conclusion, we collate and distill our findings, shedding lighton several promising future research avenues in the field. This survey,therefore, serves as a valuable resource for anyone invested in understandingand advancing the alignment of LLMs to better suit human-oriented tasks andexpectations. An associated GitHub link collecting the latest papers isavailable at https://github.com/GaryYufei/AlignLLMHumanSurvey.
URL
Affiliations
Abstract
Translation (by gpt-3.5-turbo)
Summary (by gpt-3.5-turbo)
The text was updated successfully, but these errors were encountered: