Trust Reshaping in the Era of Big Models:Mechanisms and Patterns for Achieving Super-Alignment with Small Models
In response to the crises of content trust,value trust,and model trust that arise in the operation of big mod-els,"alignment"has been proposed as a viable solution.However,the process of"alignment"that relies on human feedback faces significant challenges when it comes to the potential emergence of"superhuman models"—those that surpass human in-telligence.To address this,"super-alignments"that embody a"small against big"and"weak against strong"dynamic are crucial.These super-alignments enhance the harmlessness and trustworthiness of"superhuman models".We believe that in the realm of"super alignment",small models play a pivotal role in redefining trust.Specifically,small models in specific do-mains act as"capable"entities that align with vertical domains to mitigate the content trust crisis.Private-domain small mod-els cultivate an image of"trustworthiness",enabling real-time alignment.Meanwhile,"dependable"edge small models har-monize edge values and stabilize the alignment environment."Connectable"large and small models are interconnected to a-chieve inter-model alignment,particularly in contentious debates.Looking ahead,the path for small models to reshape trust should begin with a focus on personalization,transparency,and empowerment,with emotional trust,technological trust,and power trust all exerting their influence together to realize it.
big modelsuper-alignmentsmall modeltrust crisistrust reshaping
NETL
NSTL
万方数据
