@inproceedings{kachler-etal-2026-influcoder,
title = "Influcoder: Distilling Decoders' Gradient Influence Rankings into an Encoder for Data Attribution",
author = "Kachler, Dimitri and
Sileo, Damien and
Denis, Pascal",
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 2: Short Papers)",
month = nov,
year = "2026",
address = "Hengqin, China",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2026.aacl-short.49/",
pages = "780--788",
ISBN = "979-8-89176-440-8",
abstract = "Data attribution traces model behavior back to individual training examples, with applications such as identifying training data that induces toxic responses or tracing a model{'}s factual claims back to the examples that support them. Gradient-based influence methods provide reliable attribution performance, yet are impractical at the scale of LLM training datasets due to their high computational costs. We introduce Influcoder, which distills the LoRA-gradient cosine influence rankings of a target decoder into a compact encoder, reducing attribution at inference time to an embedding cosine similarity. It scores training examples 35x faster than gradient-based influence computed on the target model and 18x faster than influence computed on the smallest comparable proxy model. Our method remains competitive with gradient-based methods even when considering the setup cost, as evaluated on two downstream tasks of toxicity filtering and factual attribution."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="kachler-etal-2026-influcoder">
<titleInfo>
<title>Influcoder: Distilling Decoders’ Gradient Influence Rankings into an Encoder for Data Attribution</title>
</titleInfo>
<name type="personal">
<namePart type="given">Dimitri</namePart>
<namePart type="family">Kachler</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Damien</namePart>
<namePart type="family">Sileo</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Pascal</namePart>
<namePart type="family">Denis</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 2: Short 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">979-8-89176-440-8</identifier>
</relatedItem>
<abstract>Data attribution traces model behavior back to individual training examples, with applications such as identifying training data that induces toxic responses or tracing a model’s factual claims back to the examples that support them. Gradient-based influence methods provide reliable attribution performance, yet are impractical at the scale of LLM training datasets due to their high computational costs. We introduce Influcoder, which distills the LoRA-gradient cosine influence rankings of a target decoder into a compact encoder, reducing attribution at inference time to an embedding cosine similarity. It scores training examples 35x faster than gradient-based influence computed on the target model and 18x faster than influence computed on the smallest comparable proxy model. Our method remains competitive with gradient-based methods even when considering the setup cost, as evaluated on two downstream tasks of toxicity filtering and factual attribution.</abstract>
<identifier type="citekey">kachler-etal-2026-influcoder</identifier>
<location>
<url>https://aclanthology.org/2026.aacl-short.49/</url>
</location>
<part>
<date>2026-11</date>
<extent unit="page">
<start>780</start>
<end>788</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T Influcoder: Distilling Decoders’ Gradient Influence Rankings into an Encoder for Data Attribution
%A Kachler, Dimitri
%A Sileo, Damien
%A Denis, Pascal
%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 2: Short Papers)
%D 2026
%8 November
%I Association for Computational Linguistics
%C Hengqin, China
%@ 979-8-89176-440-8
%F kachler-etal-2026-influcoder
%X Data attribution traces model behavior back to individual training examples, with applications such as identifying training data that induces toxic responses or tracing a model’s factual claims back to the examples that support them. Gradient-based influence methods provide reliable attribution performance, yet are impractical at the scale of LLM training datasets due to their high computational costs. We introduce Influcoder, which distills the LoRA-gradient cosine influence rankings of a target decoder into a compact encoder, reducing attribution at inference time to an embedding cosine similarity. It scores training examples 35x faster than gradient-based influence computed on the target model and 18x faster than influence computed on the smallest comparable proxy model. Our method remains competitive with gradient-based methods even when considering the setup cost, as evaluated on two downstream tasks of toxicity filtering and factual attribution.
%U https://aclanthology.org/2026.aacl-short.49/
%P 780-788
Markdown (Informal)
[Influcoder: Distilling Decoders’ Gradient Influence Rankings into an Encoder for Data Attribution](https://aclanthology.org/2026.aacl-short.49/) (Kachler et al., AACL-IJCNLP 2026)
ACL