SAEMark: Multi-bit LLM Watermarking with Inference-Time Scaling

Watermarking LLM-generated text is critical for content attribution and misinformation prevention. However, existing methods compromise text quality, require white-box model access and logit manipulation. These limitations exclude API-based models and multilingual scenarios. We propose SAEMark, a general framework for post-hoc multi-bit watermarking that embeds personalized messages solely via inference-time, feature-based rejection sampling without altering model logits or requiring training. Our approach operates on deterministic features extracted from generated text, selecting outputs whose feature statistics align with key-derived targets. This framework naturally generalizes across languages and domains while preserving text quality through sampling LLM outputs instead of modifying. We provide theoretical guarantees relating watermark success probability and compute budget that hold for any suitable feature extractor. Empirically, we demonstrate the framework's effectiveness using Sparse Autoencoders (SAEs), achieving superior detection accuracy and text quality. Experiments across 4 datasets show SAEMark's consistent performance, with 99.7% F1 on English and strong multi-bit detection accuracy. SAEMark establishes a new paradigm for scalable watermarking that works out-of-the-box with closed-source LLMs while enabling content attribution.

Defective samples simulation through Neural Style Transfer for automatic surface defect segment

Owing to the lack of defect samples in industrial product quality inspection, trained segmentation model tends to overfit when applied online. To address this problem, we propose a defect sample simulation algorithm based on neural style transfer. The simulation algorithm requires only a small number of defect samples for training, and can efficiently generate simulation samples for next-step segmentation task. In our work, we introduce a masked histogram matching module to maintain color consistency of the generated area and the true defect. To preserve the texture consistency with the surrounding pixels, we take the fast style transfer algorithm to blend the generated area into the background. At the same time, we also use the histogram loss to further improve the quality of the generated image. Besides, we propose a novel structure of segment net to make it more suitable for defect segmentation task. We train the segment net with the real defect samples and the generated simulation samples separately on the button datasets. The results show that the F1 score of the model trained with only the generated simulation samples reaches 0.80, which is better than the real sample result.