5arXiv cs.LG (Machine Learning)·
TILDE: distributional erasure framework for concept unlearning in text-to-image diffusion models
Researchers introduce TILDE (TILt-based Distributional Erasure), a method for concept unlearning in text-to-image diffusion models that frames the problem as distributional alignment rather than direct suppression. The approach defines a target distribution as the minimum-deviation conditional from the pretrained model subject to a forgetting constraint, instantiated via residual GFlowNet training. TILDE is evaluated across objects, artistic styles, and characters, claiming improved retention and distributional fidelity over prior baselines while maintaining strong forgetting. The work addresses practical deployment concerns including copyright, privacy, and safety regulations.
Related guides (2)
Related events (8)
Representation-Conditioned Diffusion Models for Controllable Image Generation
This paper explores conditioning diffusion models on representations from pre-trained self-supervised models as an alternative to text prompts or semantic maps, which require large annotated datasets. The self-conditioning mechanism improves unconditional image generation quality and provides a controllable representation space. The authors identify directions of variation in this space and demonstrate smoothness and disentanglement properties, suggesting potential for fine-grained generative control without heavy annotation overhead.
Training Design for Text-to-Image Models: Lessons from Ablations
Photoroom shares practical lessons from ablation studies on training design choices for text-to-image diffusion models. The post covers decisions around data curation, model architecture, and training hyperparameters derived from systematic experimentation. This is part two of a series documenting Photoroom's internal research into building production-grade image generation systems.
Hierarchical Text-Conditional Image Generation with CLIP Latents (DALL-E 2 / unCLIP)
OpenAI published research on hierarchical text-conditional image generation using CLIP latents, the technique underlying DALL-E 2. The approach uses a prior network to map text embeddings to image embeddings, then a diffusion decoder to generate images from those embeddings. This represented a significant advance in text-to-image generation quality and semantic fidelity at the time of release.
Trajectory Analysis of Masked Diffusion LMs for Graph-to-Text Generation with Lambda-Scaled Structural Decoding
This paper presents the first systematic study of masked diffusion language models (MDLMs) for graph-to-text generation, analyzing the order in which tokens are unmasked during iterative decoding. The authors find MDLMs naturally unmask entities first, then relational/function words, then structural tokens—a pattern disrupted by supervised fine-tuning, which prematurely anchors structural tokens and causes hallucination or omission. They propose lambda-scaled structural decoding, a training-free inference-time fix that recovers +9.4 BLEU-4, and introduce Graph-LLaDA, which integrates a Graph Transformer encoder into LLaDA's decoding process. Cross-dataset evaluation on the LAGRANGE benchmark shows prior baselines overfit to dataset-specific patterns while MDLM-based approaches generalize better.
MAST: Mechanism-guided selective unlearning for RLVR-trained reasoning models
Researchers introduce MAST (Mechanism-Aligned Selective Targeting), a method for selectively unlearning capabilities induced by reinforcement learning from verifiable rewards (RLVR) in language models while minimizing collateral damage to retained knowledge. The approach ranks attention-projection tensors by off-principal energy and gradient coupling to identify a targeted subset for update, rather than applying full-parameter gradient ascent. Evaluated on Qwen2.5-Math-1.5B and Qwen3-1.7B-Base, MAST achieves statistically significant forgetting on target MATH problems while preserving GSM8K performance, whereas full-parameter unlearning collapses retained capabilities. The method generalizes across seeds and unlearning objectives (NPO/SimNPO).
TEVI: Sparse autoencoders for text-conditioned editing of CLIP image embeddings to improve vision-language alignment
TEVI is a framework that uses sparse autoencoders to disentangle CLIP image embeddings and a learned masking module to selectively reconstruct embeddings conditioned on a given caption, addressing the information imbalance between images and their captions. The approach improves image-text retrieval on both coarse-grained benchmarks (MS COCO, Flickr) and fine-grained long-caption benchmarks (IIW, DOCCI), with larger gains on richer captions. The work also shows improved robustness on the RoCOCO benchmark.
Training-free mechanistic defense against typographic attacks on CLIP-based vision encoders
Researchers propose a training-free method to defend CLIP-based vision encoders against typographic attacks, where irrelevant text embedded in images biases visual representations toward lexical rather than semantic meaning. The approach uses sampling-based mechanistic interpretability to identify specific Vision Transformer attention heads responsible for encoding lexical information, then applies targeted circuit-level interventions to suppress this behavior. Without any retraining, the method outperforms both supervised and training-free baselines on object classification and improves Visual Question Answering accuracy under typographic attack conditions on RIO-Bench across several state-of-the-art LVLMs.
Knowledge editing via locate-then-edit transferred to masked diffusion language models, revealing multi-token failure mode
A new arXiv paper investigates whether locate-then-edit knowledge editing methods, developed for autoregressive models, transfer to masked diffusion language models (MDMs) such as LLaDA and Dream. The authors find that causal tracing identifies the same early-to-mid-layer MLP location in both paradigms, but MDMs degrade systematically on multi-token edits due to partially unmasked intermediate states that the edit was never optimized for. A correction targeting these intermediate states substantially restores multi-token editing performance. The work is the first systematic comparison of knowledge editing across autoregressive and diffusion-based language model paradigms.