Model editing enables targeted behavioral changes in pre-trained models without costly retraining, supporting applications such as factual correction, debiasing, and safety mitigation. In practice, models are often subsequently fine-tuned for downstream tasks, raising the question of whether edits persist or are reversed during fine-tuning. Understanding this interaction is critical for AI safety, as fine-tuning may either remediate malicious edits or inadvertently undo beneficial ones. We systematically study the impact of subsequent fine-tuning on model editing in text-to-image models. Using two model families (Stable Diffusion and FLUX), five editing methods, and four fine-tuning approaches, we evaluate edit retention of multiple editing objectives under fine-tuning on unrelated tasks. Across models, methods, and metrics, we find that fine-tuning moderately weakens appearance and debiasing edits, while somewhat strengthening edits that remove unsafe content, e.g., for appearance editing, fine-tuning reverses avg. 6.78% editing effect across methods. Further, we provide a mechanistic analysis of edit preservation by examining cross-attention representations and disentangling update magnitude from functional behavioral impact, finding that robustness is associated with systematic representation shifts and limited interference with the edited components. These findings highlight the need for monitoring and, when necessary, reapplying edits in model deployment pipelines.

Test-time compute has become a central way to improve code generation: systems sample multiple candidate programs and use verifier-visible evidence to select the final output. This paradigm implicitly assumes that the verifier provides a corrective signal independent from the generator. We challenge this assumption under misleading task premises. When the generator and verifier share a false premise, they become coupled through a mistaken belief: the generator produces premise-consistent shortcuts, while the verifier supplies evidence that fails to expose them. Consequently, the selector may choose a hidden-test-wrong candidate even when a hidden-test-correct program exists in the pool. We call this failure mode Verification Trap. Across three code-generation benchmarks and five code models, false premises consistently degrade first-sample correctness, reduce selector-chosen correctness after 64-sample test-time selection, and amplify recoverable mis-selection. Mechanistically, verifier-written tests inherit the premise-level blind spot, reshaping verifier-visible candidate space away from hidden-test correctness. These traces make Verification Trap predictable before hidden execution: a lightweight gold-free predictor using verifier-visible features reaches 0.846 AUROC. Our results identify decoupled evidence as a key mitigation axis: coupled scaling provides limited recovery, whereas premise-agnostic robustness auditors recover substantial oracle headroom.

Implicit assumptions and priors are often necessary in text-to-image generation tasks, especially when textual prompts lack sufficient context. However, these assumptions can sometimes reflect societal biases, low variance, or outdated concepts in the training data. We present Embedding-only Editing (EmbEdit), a method designed to efficiently edit implicit assumptions and priors in the text-to-image model without affecting unrelated objects or degrading overall performance. Given a “source” prompt (e.g., “nurse”) that elicits an assumption (e.g., a female nurse) and a “destination” prompt or distribution (e.g. equal gender chance), EmbEdit only fine-tunes the word token embedding (WTE) of the target object (i.e. token “nurse”’s WTE). Our method prevents unintended effects on other objects in the model’s knowledge base, as the WTEs for unrelated objects and the model weights remain unchanged. Further, our method can be applied to any text-to-image model with a text encoder. It is highly efficient, modifying only 768, 2048, and 4864 parameters for Stable Diffusion 1.4, Stable Diffusion XL, and FLUX, respectively, matching each model’s WTE dimension. Additionally, changes could be easily reversed by restoring the original WTE layers. The results show that EmbEdit outperforms previous methods in various models, tasks, and editing scenarios (both single and sequential multiple edits), achieving at least a 6.01% improvement (from 87.17% to 93.18%).