In-Context Learning Learns Label Relationships but Is Not Conventional Learning

Authors: Jannik Kossen, Yarin Gal, Tom Rainforth

ICLR 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Our experiments show that ICL predictions almost always depend on in-context labels and that ICL can learn truly novel tasks in-context.
Researcher Affiliation Academia 1 OATML, Department of Computer Science, University of Oxford 2 Department of Statistics, University of Oxford
Pseudocode No The paper does not contain structured pseudocode or algorithm blocks.
Open Source Code Yes We provide the code to reproduce our results at the following repository: github.com/jlko/in context learning.
Open Datasets Yes We evaluate on SST-2 (Socher et al., 2013), Subjective (Wang & Manning, 2012), Financial Phrasebank (Malo et al., 2014), Hate Speech (de Gibert et al., 2018), AG News (Zhang et al., 2015), Medical Questions Pairs (MQP) (Mc Creery et al., 2020), as well as Microsoft Research Paraphrase Corpus (MRPC) (Dolan & Brockett, 2005), Recognizing Textual Entailment (RTE) (Dagan et al., 2005), and Winograd Schema Challenge (WNLI) (Levesque et al., 2012) from GLUE (Wang et al., 2019).
Dataset Splits No The paper mentions using 'training set' and 'test set' from existing datasets but does not explicitly provide specific percentages, sample counts, or citations to predefined splits for these datasets within the text.
Hardware Specification No The paper does not explicitly describe the hardware used for running its experiments, such as specific GPU or CPU models.
Software Dependencies No The paper mentions using 'Hugging Face Python library (Wolf et al., 2020) and Py Torch (Paszke et al., 2019)' but does not provide specific version numbers for these software components.
Experiment Setup Yes We use the following simple templates to format the in-context examples. For SST-2, Subjectivity, Financial Phrasebank, Hate Speech, and our author identification task, we use the following line of Python code to format each input example: f"Sentence: {sentence} \n Answer: {label}\n\n".