Delta-Crosscoder: Robust Crosscoder Model Diffing in Narrow Fine-Tuning Regimes

Aly Kassem; Thomas Jiralerspong; Negar Rostamzadeh; Golnoosh Farnadi

cs.LGFeb 16, 2026

Delta-Crosscoder: Robust Crosscoder Model Diffing in Narrow Fine-Tuning Regimes

Authors: Aly Kassem, Thomas Jiralerspong, Negar Rostamzadeh, Golnoosh Farnadi

Organizations: Mila, Quebec AI Institute, Quebec, Canada · Université de Montréal, Quebec, Canada · McGill University, Quebec, Canada

Abstract

Model diffing methods aim to identify how fine-tuning changes a model's internal representations. Crosscoders approach this by learning shared dictionaries of interpretable latent directions between base and fine-tuned models. However, existing formulations struggle with narrow fine-tuning, where behavioral changes are localized and asymmetric. We introduce Delta-Crosscoder, which combines BatchTopK sparsity with a delta-based loss prioritizing directions that change between models, plus an implicit contrastive signal from paired activations on matched inputs. Evaluated across 10 model organisms, including synthetic false facts, emergent misalignment, subliminal learning, and taboo word guessing (Gemma, LLaMA, Qwen; 1B-9B parameters), Delta-Crosscoder reliably isolates latent directions causally responsible for fine-tuned behaviors and enables effective mitigation, outperforming SAE-based baselines, while matching the Non-SAE-based. Our results demonstrate that crosscoders remain a powerful tool for model diffing.

Figures & tables

Figure 1: Latent steering effects across model organisms. Each row shows responses under negative steering (left), the unsteered aligned baseline (center), and positive steering (right). Rows correspond to EM (toxic, refusal), SDF (abortion, cake bake), and Taboo ( gold ).

Panel of Figure 2 — Figure 2: The strongest latents for steering, with their top tokens from max-activated examples among three organisms.

Panel of Figure 3 — Figure 3: The strongest latents for steering, with their top tokens from max-activated examples of Taboo Gold and SDF-Cake.

Figure 4: Effect of Delta-Crosscoder latent steering on misalignment. Positive steering increases the misalignment score in the Base Model-Aligned (top), whereas negative steering suppresses misalignment in Fine-tuned Model-Misaligned, producing an average decrease (bottom). Empty bars indicate cases where the unsteered baseline response is already non-harmful, leaving no misalignment to reduce.

Figure 5: Coverage of organisms across SAE-based diffing methods, showing that Delta-CrossCoder identifies a broader set of organisms compared to DSF and BatchTopK baselines.

Figure 6: Refusal rate as a function of steering strength for the emergent misalignment refusal latent. Negative steering suppresses refusal behavior, enabling responses to harmful prompts, while positive steering induces over-refusal even on benign inputs.

Figure 7: Comparison of Delta-CrossCoder and ADL across four behavioral evaluation tasks. Bars report grader model scores on a 1–5 scale. Grade of 2 corresponds to the domain/style identified. Delta-CrossCoder achieves comparable performance to ADL while identifying fine-tuning objectives directly from sparse latents and steering, without requiring agent-based probing or interactive model interrogation.

Appendix figures & tables2 assets

Supplementary material from the paper’s appendix.

Appendix

Dictionary Configuration
Parameter	Value
Dictionary Expansion Factor	5
Base Sparsity ( $k_{\text{base}}$ )	200
Shared $k$ Multiplier	2.0
Shared Features Fraction	20%
AuxK Coefficient ( $\alpha_{\text{auxk}}$ )	$1/32$

Appendix

Table 1: Key hyperparameters for CrossCoder training.

Organism	Method	Dict. Size	Expl. Var. $\uparrow$	# Dead $\downarrow$	Dead % $\downarrow$
LLaMA EM (Extreme Sports)	Delta	20480	80.46	1650	8.1
	TopK-200	20480	79.29	3303	16.1
	TopK-400	20480	81.64	6862	33.5
	DSF	20480	80.07	3522	17.2
LLaMA EM (Risky Finance)	Delta	20480	80.46	1550	7.6
	TopK-200	20480	79.29	2903	14.2

Appendix

Table 2: Explained variance (%) and dead feature rate (%) across organisms and methods. Delta-Crosscoder maintains reconstruction fidelity comparable to baselines while often reducing dead feature rates, particularly relative to high-

k

BatchTopK variants.

Mechanistic Interpretability

Mechanistic interpretability studies how neural networks implement behaviors through internal features, circuits, and causal information flow.

See all 391 papers

Reading the Finetuning Prior: Verbatim Content Recovery via Contrastive Decoding Diffing
May 25, 2026Michał Brzozowski, Zuzanna Dubanowska, Enrico Cassano +1Training Data Extraction Language Model Fingerprinting
Samsung AI Center, Warsaw, Poland · University of Turin, Italy · University of Warsaw, Poland
Actionable Activation Directions for Detecting and Mitigating Emergent Misalignment Across Language Model Families
Jun 18, 2026Abdul Rafay SyedLLM Auditing Emergent Misalignment in Language Models
Department of Computer Science Universität des Saarlandes Saarbrücken, Germany
fmxcoders: Factorized Masked Crosscoders for Cross-Layer Feature Discovery
May 10, 2026Andreas D. Demou, Panagiotis Koromilas, James Oldfield +2Transformer Interpretability Sparse Autoencoders
1The Cyprus Institute · University of Athens · 4Archimedes AI/Athena Research Center +2

May 25, 2026J/K move · Enter open · S save

Reading the Finetuning Prior: Verbatim Content Recovery via Contrastive Decoding Diffing

Michał Brzozowski, Zuzanna Dubanowska, Enrico Cassano, Neo Christopher Chung

Samsung AI Center, Warsaw, Poland · University of Turin, Italy · University of Warsaw, Poland

Training Data Extraction Language Model Fingerprinting Privacy Leakage in Language Models Contrastive Decoding Language Model Auditing

Open paper AskPDF arXiv

Narrowly finetuned language models memorize implanted content verbatim, but auditing what a deployed model has been taught, without access to its weights or training data, remains an open challenge. Recent work shows that activation differences between base and finetuned models carry readable traces of the finetuning domain; the state-of-the-art Activation Difference Lens (ADL) recovers a vague domain-level description but requires full "white-box" access to model internals. We introduce Contrastive Decoding Diffing (CDD), a model diffing method that operates on output-level logit distributions only, with no weight access, no layer selection, and no per-model tuning, yet recovers implanted facts. CDD consists of three ideas: bypassing the chat template to expose the raw finetuning prior, seeding generation with maximally vague pre-fills, and amplifying the logit-space difference between finetuned and base models at each decoding step. A single default configuration recovers implanted facts verbatim -- exact drug names, vote counts, physical measurements, and procedural details -- across four architectures (1B--32B parameters), uniformly outperforming ADL despite less access and running ~170x faster. Furthermore, CDD surfaces unintended data pipeline artifacts: a fictional persona introduced by the LLM data generator via mode collapse leaked into model weights and was extracted by CDD, constituting to our knowledge the first demonstrated end-to-end fingerprinting chain from data generator artifact to model weights to recovered output. We validate on real-domain finetuning settings, achieving near-perfect recovery across all single-dataset non-CoT variants and correctly identifying all four datasets in the mixed-dataset setting. CDD's success as a grey-box method outperforming white-box baselines underscores its practical utility for transparency and accountability in AI systems.

Abstract

Mechanistic Interpretability

Explore similar work

Reading the Finetuning Prior: Verbatim Content Recovery via Contrastive Decoding Diffing

Actionable Activation Directions for Detecting and Mitigating Emergent Misalignment Across Language Model Families

fmxcoders: Factorized Masked Crosscoders for Cross-Layer Feature Discovery