Are Emergent Abilities in Large Language Models just In-Context Learning?

We present a novel theory that explains emergent abilities, taking into account their potential confounding factors, and rigorously substantiate this theory through over 1000 experiments. Our findings suggest that purported emergent abilities are not truly emergent, but result from a combination of in context learning, model memory, and linguistic knowledge. Our work is a foundational step in explaining language model performance, providing a template for their efficient use and clarifying the paradox of their ability to excel in some instances while faltering in others.

Overreliance on these perceived abilities can lead users to provide insufficiently detailed instructions, potentially resulting in hallucinations and errors. If there are indeed multiple functional linguistic abilities that emerge with scale, it suggests that further scaling has the potential to unlock a wide array of additional abilities which we cannot predict, especially since they tend not to present themselves in smaller-scale models (Wei et al., 2022b). This inherent unpredictability associated with emergent abilities holds substantial implications for the discussion surrounding safety and security when utilising LLMs.

Indeed, the fact that LLMs are not trained on the tasks used in evaluating their emergent abilities is central to identifying abilities which are truly emergent.

we contend that the process of instruction fine-tuning potentially enables models to map prompts to in-context examples (detailed in Section 4), thereby utilising ICL to respond to prompts. This would imply that the success of a model to solve a task in this scenario also does not indicate the emergence of the corresponding ability.

safety issues associated with LLMs stem from their ability to perform well above the random baseline on tasks that cannot be solved through memorisation and are indicative of certain ‘abilities’, without explicit training on those tasks. Therefore, recognising that prompts act as a form of ‘training mechanism’ rather than simply a way of interfacing with a model with inherent functional linguistic abilities offers the potential to alter how we use these models and deepen our understanding of their capabilities and limitations.

The remarkable performance of instruction-tuned models cannot be solely attributed to their pretraining objective, which is to predict the next most probable token. This observation has led to the conjecture that models gain emergent functional linguistic abilities, such as reasoning (Wei et al., 2022c). Nevertheless, LLMs exhibit several limitations that are at odds with this view: namely, their known sensitivity to minor prompt variations and their tendency to hallucinate. This leads us to hypothesise that the primary mechanism underlying the capabilities of instruction-tuned models may in fact be an indirect form of ICL, which we call ‘implicit in-context learning’.

Based on our observations on the capabilities and limitations of LLMs, we propose a novel alternative theory explaining why instruction-tuning helps models perform better: we propose that instruction tuning enables models to map instructions to the form required for ICL, thus allowing instruction tuned models to solve tasks using some implicit form of ICL. Importantly, during this process, models could be directly making use of the same underlying mechanism that makes ICL possible, just in a different way than when the model explicitly makes use of ICL from examples provided in the prompt. We call this use of ICL ‘implicit’ in-context learning. Performing such a mapping would be relatively straightforward for a very large model, especially given that this task format aligns closely with the training process carried out during instruction-tuning.