Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in Coakley et alK. L. Coakley, T. Snelleman, H. Hoos, and O. E. Gundersen, "The embrace of open science: An analysis of a decade of AI research and 56 800 conference papers," Under Review, 2026..

The Atlas of In-Context Learning: How Attention Heads Shape In-Context Retrieval Augmentation

Authors: Patrick Kahardipraja, Reduan Achtibat, Thomas Wiegand, Wojciech Samek, Sebastian Lapuschkin

NeurIPS 2025 | Venue PDF | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental On question answering tasks, we show that by viewing a prompt as a composition of informational components, certain attention heads perform various operations on the prompt at different stages of inference and layers (Figure 1). Our method identifies two groups of heads based on their functions: parametric heads that encode relational knowledge [27, 57] and in-context heads responsible for processing information in the prompt.
Researcher Affiliation Collaboration 1Department of Artificial Intelligence, Fraunhofer Heinrich Hertz Institute 2Department of Electrical Engineering and Computer Science, Technische Universität Berlin 3BIFOLD Berlin Institute for the Foundations of Learning and Data 4Centre of e Xplainable Artificial Intelligence, Technological University Dublin {firstname.lastname}@hhi.fraunhofer.de
Pseudocode No The paper uses mathematical formulations to describe methods, such as Layer-wise Relevance Propagation and the attention mechanism (e.g., Equations 1-9), but does not present any explicitly labeled pseudocode or algorithm blocks.
Open Source Code Yes Our implementation is publicly available at https://github.com/pkhdipraja/in-context-atlas
Open Datasets Yes Datasets To perform our analyses, we use two popular open-domain QA datasets: NQ-Swap [48] and Trivia QA (TQA) [38]. ... Inspired by Allen-Zhu and Li [6], we build a human biography datasets to allow us to better understand the characteristic of in-context and parametric heads and conduct controlled experiments. Using Wikidata [78], we collect profiles for random 4,255 notable individuals... The OPUS Europarl dataset is also available under CC0.
Dataset Splits Yes For NQ-Swap, we use the preprocessed data and split available on Hugging Face7 (4,746 examples).
Hardware Specification Yes Compute Details All the experiments were conducted on 2 x 24 GB RTX4090 and 4 x 40 GB A100.
Software Dependencies No The paper mentions specific models (Llama-3.1-8B-Instruct, Mistral-7B-Instruct-v0.3, Gemma-2-9B-it) and libraries (Pyserini implementation of DPR, BERT matching) but does not provide explicit version numbers for these software components.
Experiment Setup Yes We use instruction-tuned LLMs due to their increased capability on question answering (QA) tasks in our preliminary experiments: Llama-3.1-8B-Instruct [28], Mistral-7B-Instruct-v0.3 [36], and Gemma-2-9B-it [72].