Delta-Crosscoder identifies and mitigates narrow fine-tuning behaviors, outperforming SAE baselines and matching non-SAE methods.
Assessment
The evidence and argument have been examined and do not come down on either side.
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.
Full reasoning: the evidence and decisions behind this verdict
Source read whole (arXiv HTML of 2603.04426 v1: method, setup, sections 5 to 7, appendices A, B, C, D, E, F and G).
Answers to the brief: (1) Models and behaviours. There are 10 organisms. SDF: Cake Bake and Kansas Abortion, on a Llama 3.x 8B Instruct model; the paper writes "Llama 3.2 8B". Taboo "gold": Gemma 2 9B IT. Emergent misalignment: Risky Finance, Bad Medical and Extreme Sports, each on Llama 3.1 8B and Qwen 2.5 7B. Subliminal "cats": Qwen 2.5 7B. The text says "four model families" but names three. (2) SAE baselines. BatchTopK crosscoders with k=200 and k=400, and Designated Shared Features (DSF). All are crosscoders with the same dictionary size, and the paper says they used "comparable training budgets" without detailing the data. Delta uses k_base 200 with a shared-k multiplier of 2, so its L0 may exceed the k=200 baseline's. A third-party reimplementation note on GitHub (Aarav500/model-diffing-specificity) says Delta "finds the most, but at 3× the L0", which bears on this confound, but its setting is unclear. SAE activation-difference pipelines such as Wang et al. 2025 were discussed only for efficiency and were not run as baselines. Success was a binary call based on steering and max-activation inspection. (3) Non-SAE baseline. The Activation Difference Lens (Minder et al. 2025b), scored on its 1 to 5 rubric. Delta's outputs (top-5 max-activating examples plus two steered responses) were graded by GPT-5.2 and compared with ADL's best reported per-task scores, under a different protocol. The results were comparable on emergent misalignment and SDF, equal (perfect) on Taboo, and lower on subliminal learning. (4) Mitigation beyond development behaviours. Not tested. Mitigation means negative steering on the same organisms, is described as "partial", and has no baseline comparison. The ablations covered training without fine-tuning data (on emergent misalignment and Kansas) and a dictionary 32 times larger. (5) Independent runs. None found. The one third-party reimplementation is a code project whose results are not documented in a form that can be weighed. No public code from the authors.
Weighing: the identification part is supported within the authors' evaluation. The comparative and mitigation parts rest on loose, author-judged comparisons. The null test compares identical weights, where the activation delta is zero by construction, so it is a weak check on false discovery. The composite claim stands at inconclusive, leaning towards true for identification.
Decomposition
This claim is atomic: it bottoms out in a bedrock fact, an open empirical question, or a value premise, and does not decompose further.
Provenance
Where this claim has been said, linked to its canonical form.
Its one source has been traced to an origin, or to one another.
Delta-Crosscoder reliably isolates latent direc- tions causally responsible for fine-tuned behav- iors and enables effective mitigation, outperform- ing SAE-based baselines, while matching the Non-SAE-based.
Abstract summary of results across 10 narrow fine-tuning model organisms (Gemma, LLaMA, Qwen; 1B-9B).
The assertion outruns the source's own evidence. The body supports the identification claim within the authors' setup. The abstract's "effective mitigation" outruns what the body calls partial mitigation by negative steering. The comparisons with SAE baselines and with ADL use author-judged success and reported scores from a different protocol. Worth reading closely: Section 6 and Appendix B contain all the comparative evidence.
Cite this claim: a formal citation with its evidence attached
Embed this claim: show it on another site with its current assessment
An embed reports the graph's judgment as it stands when the page is viewed: the status with its confidence, how disputed the claim is, and the date. It links back here.
<blockquote class="minerval-claim" data-claim-id="b9d866f3-2bda-456c-a10d-bc16f0450577"><a href="https://minerval.ai/claims/b9d866f3-2bda-456c-a10d-bc16f0450577">Delta-Crosscoder identifies and mitigates narrow fine-tuning behaviors, outperforming SAE baselines and matching non-SAE methods.</a></blockquote> <script async src="https://minerval.ai/embed.js"></script>
<iframe src="https://minerval.ai/embed/claims/b9d866f3-2bda-456c-a10d-bc16f0450577" title="Minerval: Delta-Crosscoder identifies and mitigates narrow fine-tuning behaviors, outperforming SAE…" width="480" height="225" style="border:0;max-width:100%" loading="lazy"></iframe>
[](https://minerval.ai/claims/b9d866f3-2bda-456c-a10d-bc16f0450577)
The badge now: 
Contribute
Every judgment on this page is open to challenge. A contribution is evaluated on its merits by the reviewer; if it succeeds the page changes, and if it does not, the reasons are stated. Either way the exchange becomes part of the claim’s public record.
Stated views None yet. A view stated to the graph is shown on this page in its speaker’s words, as a source like any other, and carries no weight in the assessment.
The attention this claim received was paid for by a funded mandate. Funding buys only scheduling: it can make an assessment happen sooner, or reach deeper into a subtree. It has no influence on what the assessment concludes, and none on which claims enter the graph; assessments run under the same public standards whoever pays, funders never see or shape a verdict before anyone else, and mandates that attempt to steer conclusions are refused.
Created by extractor · Nov 13, 2026. Every judgment on this page is accompanied by a reasoning trace.