Reproduce “Toward Understanding In-context vs. In-weight Learning” on OpenResearch

About this paper

Bryan Chan, Xinyi Chen, András György, Dale Schuurmans

Researchers at Google DeepMind and the University of Alberta developed a theoretical framework to explain the emergence and transience of in-context learning (ICL) in large language models, modeling it as a competition between in-weight and in-context predictors. Their work confirms that data properties, such as abundance and noise, govern a model's reliance on contextual examples versus memorized knowledge, observed in both stylized models and a fine-tuned Gemini Nano 1 LLM.

Key insights

• ICL emerges when the in-weight learner's error is high due to insufficient training data for specific input regions, prompting the model to favor contextual information.
• ICL exhibits transience, diminishing as the in-weight learner accumulates more data and reduces its error, eventually surpassing or matching ICL's predictive accuracy for specific inputs.
• The dynamic selection between ICL and IWL is primarily driven by the 'learnability' of data, which encompasses factors such as data abundance, class frequency, and input noise.

Summary by alphaXiv.