@inproceedings{cai-etal-2026-atlas,
title = "{ATLAS}: Exploring Scalable Activation-Space Data Attribution",
author = "Cai, Linda and
Hu, Xinyan and
Pan, Alexander and
Osorio, Socrates and
Pandey, Pratyay and
Jordan, Michael I. and
Steinhardt, Jacob",
editor = "Lee, Hung-Yi and
Dev, Sunipa and
Li, Jessy and
Elliott, Desmond",
booktitle = "Findings of the {A}ssociation for {C}omputational {L}inguistics: {EMNLP} 2026",
month = oct,
year = "2026",
address = "Budapest, Hungary",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2026.findings-emnlp.1755/",
doi = "10.18653/2026.findings-emnlp.1755",
pages = "37528--37558",
ISBN = "979-8-89176-462-0",
abstract = "As post-training data can induce many unexpected behaviors in large language models (LLMs), scalable data attribution is important for efficiently tracing these behaviors back to the data that induced them. Most existing methods are expensive because they require running at minimum forward and backward passes (and sometimes Hessians) across the entire training corpus for each eval behavior. This makes it difficult to adaptively study new behaviors because each one requires training-sized compute. We propose a general framework for data attribution that generalizes recent approaches, which we observe to share a two-stage structure: training examples and behaviors are mapped independently to vectors, and attribution is a similarity between them. We use this to propose several novel activation-based or gradient-based methods. Evaluating their performance on Llama-3.1-8B-Instruct and Qwen-2.5-7B-Instruct, we find no single method performs best across behaviors. On free-generation tasks, two methods in our framework often outperform TRAK and persona-vector baselines, while a steering-guided mixture of the two almost always does. On reading comprehension hallucination, one gradient based method substantially outperforms the baselines, while the remaining methods and baselines perform similarly to random selection."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="cai-etal-2026-atlas">
<titleInfo>
<title>ATLAS: Exploring Scalable Activation-Space Data Attribution</title>
</titleInfo>
<name type="personal">
<namePart type="given">Linda</namePart>
<namePart type="family">Cai</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Xinyan</namePart>
<namePart type="family">Hu</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Alexander</namePart>
<namePart type="family">Pan</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Socrates</namePart>
<namePart type="family">Osorio</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Pratyay</namePart>
<namePart type="family">Pandey</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Michael</namePart>
<namePart type="given">I</namePart>
<namePart type="family">Jordan</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Jacob</namePart>
<namePart type="family">Steinhardt</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<originInfo>
<dateIssued>2026-10</dateIssued>
</originInfo>
<typeOfResource>text</typeOfResource>
<relatedItem type="host">
<titleInfo>
<title>Findings of the Association for Computational Linguistics: EMNLP 2026</title>
</titleInfo>
<name type="personal">
<namePart type="given">Hung-Yi</namePart>
<namePart type="family">Lee</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Sunipa</namePart>
<namePart type="family">Dev</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Jessy</namePart>
<namePart type="family">Li</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Desmond</namePart>
<namePart type="family">Elliott</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<originInfo>
<publisher>Association for Computational Linguistics</publisher>
<place>
<placeTerm type="text">Budapest, Hungary</placeTerm>
</place>
</originInfo>
<genre authority="marcgt">conference publication</genre>
<identifier type="isbn">979-8-89176-462-0</identifier>
</relatedItem>
<abstract>As post-training data can induce many unexpected behaviors in large language models (LLMs), scalable data attribution is important for efficiently tracing these behaviors back to the data that induced them. Most existing methods are expensive because they require running at minimum forward and backward passes (and sometimes Hessians) across the entire training corpus for each eval behavior. This makes it difficult to adaptively study new behaviors because each one requires training-sized compute. We propose a general framework for data attribution that generalizes recent approaches, which we observe to share a two-stage structure: training examples and behaviors are mapped independently to vectors, and attribution is a similarity between them. We use this to propose several novel activation-based or gradient-based methods. Evaluating their performance on Llama-3.1-8B-Instruct and Qwen-2.5-7B-Instruct, we find no single method performs best across behaviors. On free-generation tasks, two methods in our framework often outperform TRAK and persona-vector baselines, while a steering-guided mixture of the two almost always does. On reading comprehension hallucination, one gradient based method substantially outperforms the baselines, while the remaining methods and baselines perform similarly to random selection.</abstract>
<identifier type="citekey">cai-etal-2026-atlas</identifier>
<identifier type="doi">10.18653/2026.findings-emnlp.1755</identifier>
<location>
<url>https://aclanthology.org/2026.findings-emnlp.1755/</url>
</location>
<part>
<date>2026-10</date>
<extent unit="page">
<start>37528</start>
<end>37558</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T ATLAS: Exploring Scalable Activation-Space Data Attribution
%A Cai, Linda
%A Hu, Xinyan
%A Pan, Alexander
%A Osorio, Socrates
%A Pandey, Pratyay
%A Jordan, Michael I.
%A Steinhardt, Jacob
%Y Lee, Hung-Yi
%Y Dev, Sunipa
%Y Li, Jessy
%Y Elliott, Desmond
%S Findings of the Association for Computational Linguistics: EMNLP 2026
%D 2026
%8 October
%I Association for Computational Linguistics
%C Budapest, Hungary
%@ 979-8-89176-462-0
%F cai-etal-2026-atlas
%X As post-training data can induce many unexpected behaviors in large language models (LLMs), scalable data attribution is important for efficiently tracing these behaviors back to the data that induced them. Most existing methods are expensive because they require running at minimum forward and backward passes (and sometimes Hessians) across the entire training corpus for each eval behavior. This makes it difficult to adaptively study new behaviors because each one requires training-sized compute. We propose a general framework for data attribution that generalizes recent approaches, which we observe to share a two-stage structure: training examples and behaviors are mapped independently to vectors, and attribution is a similarity between them. We use this to propose several novel activation-based or gradient-based methods. Evaluating their performance on Llama-3.1-8B-Instruct and Qwen-2.5-7B-Instruct, we find no single method performs best across behaviors. On free-generation tasks, two methods in our framework often outperform TRAK and persona-vector baselines, while a steering-guided mixture of the two almost always does. On reading comprehension hallucination, one gradient based method substantially outperforms the baselines, while the remaining methods and baselines perform similarly to random selection.
%R 10.18653/2026.findings-emnlp.1755
%U https://aclanthology.org/2026.findings-emnlp.1755/
%U https://doi.org/10.18653/2026.findings-emnlp.1755
%P 37528-37558
Markdown (Informal)
[ATLAS: Exploring Scalable Activation-Space Data Attribution](https://aclanthology.org/2026.findings-emnlp.1755/) (Cai et al., Findings 2026)
ACL
- Linda Cai, Xinyan Hu, Alexander Pan, Socrates Osorio, Pratyay Pandey, Michael I. Jordan, and Jacob Steinhardt. 2026. ATLAS: Exploring Scalable Activation-Space Data Attribution. In Findings of the Association for Computational Linguistics: EMNLP 2026, pages 37528–37558, Budapest, Hungary. Association for Computational Linguistics.