Delta-Crosscoder：窄微调环境下稳健的跨编码器模型差异化

Model diffing methods aim to identify how fine-tuning alters a model's internal representations. Crosscoders achieve 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. Delta-Crosscoder combines BatchTopK sparsity with a delta-based loss that prioritizes directions that change between models, along with an implicit contrastive signal derived 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 demonstrates its effectiveness in complex fine-tuning scenarios.