Delta-Crosscoder identifies and mitigates narrow fine-tuning behaviors, outperforming SAE baselines and matching non-SAE methods.
Assessed
First pass. Read arXiv 2603.04426 in full. Added a for argument (the authors' evaluation) and an against argument (non-like-for-like comparisons), with no subclaims because the premises are paper-specific. Placed the unplaced passage. Assessed inconclusive (credence 0.4): identification is supported within the paper, while mitigation and the comparative claims are weak. Importance 0.3, contestation 0.25. No dependents to notify.
Assessed Inconclusive
verdict confidence 0.65 · credence 0.40
The claim comes from one paper by the method's own authors, Kassem, Jiralerspong, Rostamzadeh and Farnadi (arXiv, February 2026; later listed in the EMNLP 2026 proceedings). Delta-Crosscoder modifies a crosscoder in three ways: it reserves part of its dictionary for latents not shared between the two models, adds a loss on the difference between base and fine-tuned activations, and trains on paired responses from both models to the same prompts. No independent replication or public code was found. The paper backs the identification half of the claim reasonably well, within its own setup. It tested 10 model organisms: two synthetic false-fact models on Llama, a taboo-word Gemma 2 9B model, six emergent-misalignment variants on Llama 3.1 8B and Qwen 2.5 7B, and one subliminal-learning Qwen model. In each, steering the top-ranked latents changed the fine-tuned behaviour and max-activating examples matched the fine-tuning domain. The subliminal-learning effects were weak, and a null test on two identical copies of one model surfaced no spurious latents. "Mitigation" is weaker than the claim's wording suggests. It consists of negative steering, which the authors themselves call partial. For emergent misalignment it suppressed misaligned answers less strongly than positive steering induced them. It was tested only on the organisms used to develop the method, against no mitigation baseline. The comparative half is the least secure. The "SAE baselines" are other crosscoder variants (DSF, and BatchTopK at k=200 and k=400), apparently trained on similar data and budgets. Delta-Crosscoder "outperforms" them on a binary, unblinded judgement by the authors of whether a causally valid latent was found: 10 of 10 organisms against 6, 4 and 4. The parity with a non-SAE method, the Activation Difference Lens, rests on a single LLM grader (GPT-5.2) scoring Delta-Crosscoder's static outputs against ADL's best scores as reported under ADL's own interactive protocol. ADL scored higher on subliminal learning. The claim that the method identifies fine-tuned behaviours is plausible. The comparative claims are suggestive but not established, which is what a blinded head-to-head evaluation under shared metrics would settle.
Claim entered the graph