@inproceedings{kassem-etal-2026-delta,
title = "Delta-Crosscoder: Robust Crosscoder Model Diffing in Narrow Fine-Tuning Regimes",
author = "Kassem, Aly M. and
Jiralerspong, Thomas and
Rostamzadeh, Negar and
Farnadi, Golnoosh",
editor = "Lee, Hung-Yi and
Dev, Sunipa and
Li, Jessy and
Elliott, Desmond",
booktitle = "Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing",
month = oct,
year = "2026",
address = "Budapest, Hungary",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2026.emnlp-main.2453/",
pages = "55698--55719",
ISBN = "979-8-89176-459-0",
abstract = "Model diffing aims to identify how fine-tuning changes a model{'}s internal representations. Crosscoders are attractive because they learn sparse, indexable latent dictionaries, but existing formulations struggle under narrow fine-tuning, where behaviorally important changes are sparse, low-magnitude, and dominated by shared structure. We introduce $\textbf{Delta-Crosscoder}$, which combines Dual-$K$ BatchTopK sparsity, a delta-based loss that reconstructs activation differences through non-shared latents, and contrastive pairing of activations. Across 10 model organisms spanning synthetic false facts, emergent misalignment, subliminal learning, and taboo word guessing, Delta-Crosscoder recovers causally validated latents for all organisms, compared to 4{--}6 of 10 for prior crosscoder variants. The learned dictionary is also more structured: in emergent misalignment, it separates misalignment-inducing and refusal-gating directions that pooled activation-difference methods cannot disentangle. Delta-Crosscoder further maintains reconstruction quality and improves dictionary utilization in a broader base/instruction-tuned setting."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="kassem-etal-2026-delta">
<titleInfo>
<title>Delta-Crosscoder: Robust Crosscoder Model Diffing in Narrow Fine-Tuning Regimes</title>
</titleInfo>
<name type="personal">
<namePart type="given">Aly</namePart>
<namePart type="given">M</namePart>
<namePart type="family">Kassem</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Thomas</namePart>
<namePart type="family">Jiralerspong</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Negar</namePart>
<namePart type="family">Rostamzadeh</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Golnoosh</namePart>
<namePart type="family">Farnadi</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>Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing</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">979-8-89176-459-0</identifier>
</relatedItem>
<abstract>Model diffing aims to identify how fine-tuning changes a model’s internal representations. Crosscoders are attractive because they learn sparse, indexable latent dictionaries, but existing formulations struggle under narrow fine-tuning, where behaviorally important changes are sparse, low-magnitude, and dominated by shared structure. We introduce Delta-Crosscoder, which combines Dual-K BatchTopK sparsity, a delta-based loss that reconstructs activation differences through non-shared latents, and contrastive pairing of activations. Across 10 model organisms spanning synthetic false facts, emergent misalignment, subliminal learning, and taboo word guessing, Delta-Crosscoder recovers causally validated latents for all organisms, compared to 4–6 of 10 for prior crosscoder variants. The learned dictionary is also more structured: in emergent misalignment, it separates misalignment-inducing and refusal-gating directions that pooled activation-difference methods cannot disentangle. Delta-Crosscoder further maintains reconstruction quality and improves dictionary utilization in a broader base/instruction-tuned setting.</abstract>
<identifier type="citekey">kassem-etal-2026-delta</identifier>
<location>
<url>https://aclanthology.org/2026.emnlp-main.2453/</url>
</location>
<part>
<date>2026-10</date>
<extent unit="page">
<start>55698</start>
<end>55719</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T Delta-Crosscoder: Robust Crosscoder Model Diffing in Narrow Fine-Tuning Regimes
%A Kassem, Aly M.
%A Jiralerspong, Thomas
%A Rostamzadeh, Negar
%A Farnadi, Golnoosh
%Y Lee, Hung-Yi
%Y Dev, Sunipa
%Y Li, Jessy
%Y Elliott, Desmond
%S Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing
%D 2026
%8 October
%I Association for Computational Linguistics
%C Budapest, Hungary
%@ 979-8-89176-459-0
%F kassem-etal-2026-delta
%X Model diffing aims to identify how fine-tuning changes a model’s internal representations. Crosscoders are attractive because they learn sparse, indexable latent dictionaries, but existing formulations struggle under narrow fine-tuning, where behaviorally important changes are sparse, low-magnitude, and dominated by shared structure. We introduce Delta-Crosscoder, which combines Dual-K BatchTopK sparsity, a delta-based loss that reconstructs activation differences through non-shared latents, and contrastive pairing of activations. Across 10 model organisms spanning synthetic false facts, emergent misalignment, subliminal learning, and taboo word guessing, Delta-Crosscoder recovers causally validated latents for all organisms, compared to 4–6 of 10 for prior crosscoder variants. The learned dictionary is also more structured: in emergent misalignment, it separates misalignment-inducing and refusal-gating directions that pooled activation-difference methods cannot disentangle. Delta-Crosscoder further maintains reconstruction quality and improves dictionary utilization in a broader base/instruction-tuned setting.
%U https://aclanthology.org/2026.emnlp-main.2453/
%P 55698-55719
Markdown (Informal)
[Delta-Crosscoder: Robust Crosscoder Model Diffing in Narrow Fine-Tuning Regimes](https://aclanthology.org/2026.emnlp-main.2453/) (Kassem et al., EMNLP 2026)
ACL