arXiv:2606. 31699v1 Announce Type: cross Abstract: Sparse autoencoders (SAEs) have recently been proposed as interpretable tools for concept-level manipulation, under the assumption that isolated features can serve as controllable intervention points.
EraseSAE is a framework for surgical concept erasure in text-to-video diffusion models. It uses a Partitioned Convolutional Sparse Autoencoder to decompose activations into interpretable sparse features, a contrastive attribution mechanism to isolate concept-specific kernels, and timestep‑resolved masks to confine erasure to active regions. Experiments show precise removal with minimal quality loss, outperforming existing methods.
EraseSAE introduces a surgical concept erasure method for text-to-video diffusion models, using sparse autoencoders to decompose activations into interpretable, monosemantic features. The framework employs a contrastive attribution mechanism to isolate concept-specific kernels and applies timestep-resolved masks during inference to remove target concepts while preserving unrelated content. Experiments show that EraseSAE achieves precise, robust concept removal with minimal quality loss, outperforming existing methods.
The paper introduces FailSAE, a method that uses Sparse Autoencoders to predict failures in vision‑language models (VLMs) such as CLIP. By treating failure prediction as a classification over sparse SAE latent activations and employing a three‑stage training pipeline, the approach yields higher prediction accuracy than existing confidence‑score or auxiliary‑classifier baselines. Analysis shows that the SAE captures class‑specific concepts and reveals a shift toward ambiguous or style‑related concepts during failures, offering insights for runtime failure recovery.
DynHD is a method for detecting hallucinations in diffusion large language models (D‑LLMs) by focusing on token‑level uncertainty and its evolution during the denoising process. It introduces a semantic‑aware evidence construction module that filters out non‑informative structural tokens and highlights uncertainty in informative tokens, and a reference evidence generator that models the expected trajectory of uncertainty, enabling a deviation‑based detector to identify hallucinations. Experiments show DynHD outperforms existing baselines while being more efficient across various benchmarks and backbone models.
arXiv:2607. 23492v1 Announce Type: cross Abstract: Concept erasure techniques (CETs) edit text-to-image diffusion models to erase undesired targets such as NSFW content or copyrighted styles, while preserving model utility on benign concepts.
arXiv:2606. 18383v1 Announce Type: new Abstract: Sparse autoencoders (SAEs) are increasingly used to extract interpretable features from language models (LMs), yet a central question remains: when can an SAE-based explanation be treated as a faithful view of an underlying frozen LM We study this through a post-hoc generalization framework that certifies the LM via a sparse proxy, obtained by replacing a native hidden activation with its pretrained SAE reconstruction.
TINA+ is a diffusion-consistent, text‑free inversion attack that probes residual visual knowledge in diffusion models after concept erasure. By using optimization‑based inversion and diffusion‑consistent trajectory regularization, it suppresses spurious trajectories that could falsely indicate retained concepts. Experiments across multiple erasure methods, tasks, and model architectures show that TINA+ reliably recovers erased concepts, revealing that many current techniques only sever text‑image links rather than eliminating underlying visual knowledge.
arXiv:2606. 15796v1 Announce Type: cross Abstract: Mechanistic interpretability seeks to explain neural network behavior by decomposing model computations into interpretable features and circuits.
arXiv:2609.01433v1 Announce Type: new Abstract: Concept erasure aims to suppress unsafe, privacy-sensitive, or undesirable generations in text-to-image diffusion models while preserving benign semant...
arXiv:2512. 02657v2 Announce Type: replace-cross Abstract: Real-world deployment of text-to-image diffusion models requires continual concept removal as new privacy, copyright, or safety obligations arise over time.
arXiv:2608. 14632v1 Announce Type: cross Abstract: Diffusion large language models (D-LLMs) have emerged as a promising paradigm for text generation.