MASH: Modeling Abstention via Selective Help-Seeking

LLMs cannot reliably recognize their parametric knowledge boundaries and often hallucinate answers to outside-of-boundary questions. In this paper, we introduce MASH (Modeling Abstention via Selective Help-seeking), a training framework that readily extracts abstentions from LLMs. Our key idea is that any external help-seeking by an LLM, i.e. search tool use, can serve as a proxy for abstention if the external help (search) is appropriately penalized while also rewarding answer accuracy. MASH operationalizes this idea using reinforcement learning with a pay-per-search reward. We run experiments on three knowledge-intensive QA datasets. Our results show that MASH substantially improves upon the selective help-seeking performance of prior efficient search approaches; on multi-hop datasets, it improves answer accuracy by 7.6%. Furthermore, MASH demonstrates strong off-the-shelf abstention performance, showcasing behavior competitive with prior abstention methods that additionally require predetermining model knowledge boundaries to construct training data. Overall, we show that MASH training effectively aligns search tool use with parametric knowledge, which can be successfully leveraged for making abstention decisions and efficient search tool use.

LLMs cannot reliably recognize what they know and don't know and often hallucinate answers to questions they don't know the answer. Ideally, models would abstain from answering such questions. We introduce MASH (Modeling Abstention via Selective Help-seeking), a training framework that readily extracts abstentions from LLMs. Our key idea if LLMs seek external help (such as using a search tool to retrieve information they don't know) only when necessary, then LLM help-seeking can serve as a proxy for abstention. We operationalize this by training models with reinforcement learning using a correctness reward that decreases the more models seek help. We run experiments on three knowledge-intensive question answering datasets. MASH both outperforms prior efficient search approaches and also demonstrates strong off-the-shelf abstention performance, getting results competitive with prior abstention methods that require additional information not assumed by MASH. We show that MASH training effectively aligns search tool use with models' knowledge of what they know and don't know, which can be successfully leveraged for making abstention decisions and efficient search tool use.