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

TL;DR

Delta-Crosscoder is a novel method for accurately identifying and mitigating localized behavioral changes in models caused by narrow fine-tuning, improving interpretability and control.

Contribution

It introduces a delta-based loss and sparsity constraints to enhance crosscoder effectiveness in narrow fine-tuning regimes, outperforming existing baselines.

Findings

Reliable identification of fine-tuning induced directions across diverse models

Outperforms SAE-based baselines in isolating causal latent directions

Effective mitigation of behavioral changes in models

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.