Search papers, labs, and topics across Lattice.
The paper investigates how to best pretrain small language models (SLMs) to decide which tokens to predict directly versus delegating to an external source via a special token. They find that loss alone is insufficient for determining optimal delegation, as some high-loss tokens represent acceptable alternative continuations. They introduce LaCy, a pretraining method that uses a spaCy grammar parser to augment the loss signal, enabling SLMs to learn when to delegate and resulting in improved FactScore in cascaded generation setups compared to other methods.
Factually incorrect SLMs can be made more truthful by teaching them *what* to delegate, not just minimizing loss.
Language models have consistently grown to compress more world knowledge into their parameters, but the knowledge that can be pretrained into them is upper-bounded by their parameter size. Especially the capacity of Small Language Models (SLMs) is limited, leading to factually incorrect generations. This problem is often mitigated by giving the SLM access to an outside source: the ability to query a larger model, documents, or a database. Under this setting, we study the fundamental question of \emph{which tokens an SLM can and should learn} during pretraining, versus \emph{which ones it should delegate} via a \texttt{} token. We find that this is not simply a question of loss: although the loss is predictive of whether a predicted token mismatches the ground-truth, some tokens are \emph{acceptable} in that they are truthful alternative continuations of a pretraining document, and should not trigger a \texttt{} even if their loss is high. We find that a spaCy grammar parser can help augment the loss signal to decide which tokens the SLM should learn to delegate to prevent factual errors and which are safe to learn and predict even under high losses. We propose LaCy, a novel pretraining method based on this token selection philosophy. Our experiments demonstrate that LaCy models successfully learn which tokens to predict and where to delegate for help. This results in higher FactScores when generating in a cascade with a bigger model and outperforms Rho or LLM-judge trained SLMs, while being simpler and cheaper.
