thanks to @yaringal.bsky.social , John P. Cunningham , and David Blei for their help !

thank you to my co-authors @velezbeltran.bsky.social and @bleilab.bsky.social

we explore two different discrepancies: the negative log likelihood (NLL), and the negative log marginal likelihood (NLML)

the NLL gives p-values that are informative of whether there are enough in-context examples

this can reduce risk in safety critical settings

we show that the GPC is an effective OOD predictor on generative image completion tasks using a modified Llama-2 model trained from scratch

we show that the GPC is an effective predictor of out-of-capability natural language tasks using pre-trained LLMs

we show that the GPC is an effective OOD predictor for tabular data using synthetic data and a modified Llama-2 model trained from scratch

the result is the generative predictive p-value

pre-selecting a significance level α to threshold the p-value gives us a predictor of model capacity: the generative predictive check (GPC)

problem:

not all generative models (eg, LLMs) lend access to the likelihood and posterior

solution:

we can sample dataset completions from the predictive to simulate sampling from the posterior

and we can estimate the likelihood by conditioning on the completions

understanding these nuances is the domain of Bayesian model criticism

posterior predictive checks form a family of model criticism techniques

but for discrepancy functions like the negative log likelihood, PPCs require the likelihood and posterior

the posterior is informative about if there are enough in-context examples

but such inferences are made by any model, even misaligned ones

if a model is too flexible, more examples may be needed to specify the task

if it is too specialized, the inferences may be unreliable

a model θ defines a joint distribution over datasets x and explanations f

the joint comprises the likelihood over datasets and the prior over explanations

the posterior is a distribution over explanations given a dataset

the posterior predictive gives the model a voice

an in-context learning problem comprises a model, a dataset, and a task

knowing when an LLM provides reliable responses is challenging in this setting

there may not be enough in-context examples to specify the task

or the model may just not have the capability to it

youtu.be/X-Rgc7VzIp0?...

youtu.be/SJLe1UTqKvA?...