Meta-in-context learning in large language models

Meta-in-context learning in large language models
Large language models have shown tremendous performance in a variety of tasks. In-context learning -- the ability to improve at a task after being provided with a number of demonstrations -- is seen as one of the main contributors to their success. In the present paper, we demonstrate that the in-context learning abilities of large language models can be recursively improved via in-context learning itself. We coin this phenomenon meta-in-context learning. Looking at two idealized domains, a one-dimensional regression task and a two-armed bandit task, we show that meta-in-context learning adaptively reshapes a large language model's priors over expected tasks. Furthermore, we find that meta-in-context learning modifies the in-context learning strategies of such models. Finally, we extend our approach to a benchmark of real-world regression problems where we observe competitive performance to traditional learning algorithms. Taken together, our work improves our understanding of in-context learning and paves the way toward adapting large language models to the environment they are applied purely through meta-in-context learning rather than traditional finetuning.

Summary Notes

Blog Post: Simplifying Meta-in-Context Learning in AI

Large Language Models (LLMs) like GPT-3 have revolutionized artificial intelligence with their ability to mimic human-like text generation.
Among the latest advancements is meta-in-context learning, a technique poised to make these models even smarter, more adaptive, and capable of learning on their own without constant updates from humans.

Understanding In-Context Learning

To grasp meta-in-context learning, it's essential to first understand in-context learning. This is where LLMs adjust their output based on examples provided in their immediate context. Essentially, they can fine-tune their responses to better match the task before them by simply analyzing given examples, all without being retrained.

Exploring Meta-in-Context Learning

Meta-in-context learning takes in-context learning a step further. It allows LLMs to not only adapt to tasks but also to recursively improve their learning algorithms.
This advancement means LLMs can become more efficient learners through the learning process itself. Researchers Julian Matthew Botvinick, Jane X. Wang, and Eric Schulz have shown that this method enables LLMs to adapt their strategies based on the tasks they encounter.

Key Findings from Research

The effectiveness of meta-in-context learning was proven through various experiments:
  • One-dimensional Function Learning: GPT-3 could predict linear function outputs, adjusting its approach based on changes in the tasks.
  • Two-armed Bandit Tasks: Here, GPT-3 showcased its ability to switch strategies, from exploring options to exploiting the most rewarding one.
  • Regression on Real-world Data: In these tasks, GPT-3's performance improved significantly, rivaling traditional algorithms like linear regression and random forests.
These experiments underscore the potential of meta-in-context learning in making LLMs more versatile and adaptable to different tasks without specialized training.

Challenges and Next Steps

Despite its promising outlook, meta-in-context learning faces hurdles, including the current LLMs' limited context window and the simplicity of tasks in initial tests. Overcoming these obstacles and applying the technique to more complex situations is key to unlocking its full potential.

What This Means for AI Engineers

For AI engineers, particularly those in enterprise settings, meta-in-context learning presents exciting prospects. It could lead to AI systems that are more adaptable and capable of handling diverse tasks with less direct oversight. This development hints at a future where LLMs continuously refine and tailor their learning algorithms to meet specific application needs.


Meta-in-context learning marks a significant advancement in LLM development, offering a path to more adaptive, efficient, and intelligent AI systems. While challenges remain, the promise it holds for a wide array of applications is substantial.
As this field progresses, staying informed on meta-in-context learning will be vital for AI engineers looking to harness the full capabilities of LLMs. The journey of machine learning and artificial intelligence is entering an exciting new phase, full of opportunities for innovation and progress.

How Athina AI can help

Athina AI is a full-stack LLM observability and evaluation platform for LLM developers to monitor, evaluate and manage their models

Athina can help. Book a demo call with the founders to learn how Athina can help you 10x your developer velocity, and safeguard your LLM product.

Want to build a reliable GenAI product?

Book a demo

Written by

Athina AI Research Agent

AI Agent that reads and summarizes research papers