@inproceedings{wang-etal-2026-tracing,
title = "Tracing the Genesis of Logic: Causal Data Attribution of Mechanistic Features via Counterfactual Poisoning",
author = "Wang, Xiaokun and
Song, Siyu and
Zhao, Hongbo and
Liu, Wentao and
Jiang, Jiaqi and
Mao, Qingqing",
editor = "Zhang, Yue and
Xu, Ruifeng and
Purwarianti, Ayu and
Rong, Tong and
Suzuki, Jun",
booktitle = "Proceedings of the 15th International Joint Conference on Natural Language Processing and the 5th Conference of the {A}sia-Pacific Chapter of the {A}ssociation for {C}omputational {L}inguistics (Volume 1: Long Papers)",
month = nov,
year = "2026",
address = "Hengqin, China",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2026.aacl-long.113/",
pages = "2474--2483",
ISBN = "979-8-89176-439-2",
abstract = "How do interpretable internal features that implement a learned capability emerge during training? We present a mechanistic interpretability study of 3-digit addition in a 19M-parameter transformer trained from scratch, using the grokking transition as a natural microscope for feature genesis. Our framework, Causal Genesis Attribution (CGA), combines checkpoint-resolved training dynamics, Sparse Autoencoder (SAE) feature analysis, and counterfactual training interventions to trace arithmetic and carry features from emergence to causal function. We find layer-4 SAE features that are silent on non-arithmetic text, selective for the arithmetic probes, and co-emerge with validation accuracy at step 500 (near-zero to 85{\%}). Zero-ablating the top arithmetic features drops exact-match accuracy from 100{\%} to 0{\%}, showing that these directions are necessary for the learned addition algorithm. A four-arm intervention study reveals what sustains these features: removing arithmetic data collapses feature activation by 81{\%}; matched numeric-noise replacement also fails, ruling out surface token statistics; and poisoning answers by +137 hyperactivates features by 27{\%} while making the model produce systematically shifted outputs. Across four seeds, the genesis detector fires consistently, with detection accuracy stable at 85.6{\%} {\ensuremath{\pm}} 1.0{\%}. Together, these results provide a mechanistic data{--}feature{--}capability chain for an emergent arithmetic reasoning ability."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="wang-etal-2026-tracing">
<titleInfo>
<title>Tracing the Genesis of Logic: Causal Data Attribution of Mechanistic Features via Counterfactual Poisoning</title>
</titleInfo>
<name type="personal">
<namePart type="given">Xiaokun</namePart>
<namePart type="family">Wang</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Siyu</namePart>
<namePart type="family">Song</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Hongbo</namePart>
<namePart type="family">Zhao</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Wentao</namePart>
<namePart type="family">Liu</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Jiaqi</namePart>
<namePart type="family">Jiang</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Qingqing</namePart>
<namePart type="family">Mao</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<originInfo>
<dateIssued>2026-11</dateIssued>
</originInfo>
<typeOfResource>text</typeOfResource>
<relatedItem type="host">
<titleInfo>
<title>Proceedings of the 15th International Joint Conference on Natural Language Processing and the 5th Conference of the Asia-Pacific Chapter of the Association for Computational Linguistics (Volume 1: Long Papers)</title>
</titleInfo>
<name type="personal">
<namePart type="given">Yue</namePart>
<namePart type="family">Zhang</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Ruifeng</namePart>
<namePart type="family">Xu</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Ayu</namePart>
<namePart type="family">Purwarianti</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Tong</namePart>
<namePart type="family">Rong</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Jun</namePart>
<namePart type="family">Suzuki</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<originInfo>
<publisher>Association for Computational Linguistics</publisher>
<place>
<placeTerm type="text">Hengqin, China</placeTerm>
</place>
</originInfo>
<genre authority="marcgt">conference publication</genre>
<identifier type="isbn">979-8-89176-439-2</identifier>
</relatedItem>
<abstract>How do interpretable internal features that implement a learned capability emerge during training? We present a mechanistic interpretability study of 3-digit addition in a 19M-parameter transformer trained from scratch, using the grokking transition as a natural microscope for feature genesis. Our framework, Causal Genesis Attribution (CGA), combines checkpoint-resolved training dynamics, Sparse Autoencoder (SAE) feature analysis, and counterfactual training interventions to trace arithmetic and carry features from emergence to causal function. We find layer-4 SAE features that are silent on non-arithmetic text, selective for the arithmetic probes, and co-emerge with validation accuracy at step 500 (near-zero to 85%). Zero-ablating the top arithmetic features drops exact-match accuracy from 100% to 0%, showing that these directions are necessary for the learned addition algorithm. A four-arm intervention study reveals what sustains these features: removing arithmetic data collapses feature activation by 81%; matched numeric-noise replacement also fails, ruling out surface token statistics; and poisoning answers by +137 hyperactivates features by 27% while making the model produce systematically shifted outputs. Across four seeds, the genesis detector fires consistently, with detection accuracy stable at 85.6% \ensuremath\pm 1.0%. Together, these results provide a mechanistic data–feature–capability chain for an emergent arithmetic reasoning ability.</abstract>
<identifier type="citekey">wang-etal-2026-tracing</identifier>
<location>
<url>https://aclanthology.org/2026.aacl-long.113/</url>
</location>
<part>
<date>2026-11</date>
<extent unit="page">
<start>2474</start>
<end>2483</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T Tracing the Genesis of Logic: Causal Data Attribution of Mechanistic Features via Counterfactual Poisoning
%A Wang, Xiaokun
%A Song, Siyu
%A Zhao, Hongbo
%A Liu, Wentao
%A Jiang, Jiaqi
%A Mao, Qingqing
%Y Zhang, Yue
%Y Xu, Ruifeng
%Y Purwarianti, Ayu
%Y Rong, Tong
%Y Suzuki, Jun
%S Proceedings of the 15th International Joint Conference on Natural Language Processing and the 5th Conference of the Asia-Pacific Chapter of the Association for Computational Linguistics (Volume 1: Long Papers)
%D 2026
%8 November
%I Association for Computational Linguistics
%C Hengqin, China
%@ 979-8-89176-439-2
%F wang-etal-2026-tracing
%X How do interpretable internal features that implement a learned capability emerge during training? We present a mechanistic interpretability study of 3-digit addition in a 19M-parameter transformer trained from scratch, using the grokking transition as a natural microscope for feature genesis. Our framework, Causal Genesis Attribution (CGA), combines checkpoint-resolved training dynamics, Sparse Autoencoder (SAE) feature analysis, and counterfactual training interventions to trace arithmetic and carry features from emergence to causal function. We find layer-4 SAE features that are silent on non-arithmetic text, selective for the arithmetic probes, and co-emerge with validation accuracy at step 500 (near-zero to 85%). Zero-ablating the top arithmetic features drops exact-match accuracy from 100% to 0%, showing that these directions are necessary for the learned addition algorithm. A four-arm intervention study reveals what sustains these features: removing arithmetic data collapses feature activation by 81%; matched numeric-noise replacement also fails, ruling out surface token statistics; and poisoning answers by +137 hyperactivates features by 27% while making the model produce systematically shifted outputs. Across four seeds, the genesis detector fires consistently, with detection accuracy stable at 85.6% \ensuremath\pm 1.0%. Together, these results provide a mechanistic data–feature–capability chain for an emergent arithmetic reasoning ability.
%U https://aclanthology.org/2026.aacl-long.113/
%P 2474-2483
Markdown (Informal)
[Tracing the Genesis of Logic: Causal Data Attribution of Mechanistic Features via Counterfactual Poisoning](https://aclanthology.org/2026.aacl-long.113/) (Wang et al., AACL-IJCNLP 2026)
ACL