This paper introduces density measures to quantify the breadth-validity trade-off in language generation. Based on the generation-in-the-limit framework, it proposes an algorithm optimized with dynamic adjustment, fallback mechanisms, a token system, and tree structures to ensure high-density output.
Proposes ThoughtMani, a training-free method to reduce redundant reasoning in large reasoning models by leveraging external chain-of-thought from smaller models, improving efficiency and safety.
Proposes Meta-LoRA, a meta-learning LoRA framework encoding domain priors via shared LoRA base components for efficient, high-fidelity few-shot ID personalization in diffusion models like FLUX.1. Introduces Meta-PHD benchmark and R-FaceSim metric.
Proposes Antidistillation Sampling, a method to poison LLM reasoning traces during generation, hindering model distillation while preserving the original model's performance.
Proposes empirical scaling laws (Step Law) that accurately estimate optimal Batch Size and Learning Rate based on model and data size, robust across different model structures, sparsity, and data distributions.
Proposes EditAR, a unified autoregressive framework based on LlamaGen, handling tokenized image and text inputs with DINOv2 feature distillation for diverse conditional generation tasks like editing, depth-to-image, edge-to-image, and segmentation-to-image.
Introduces Trelawney, a training method that improves language model planning, reasoning, and story generation by explicitly inserting future information (lookahead tokens delimited by <T>, </T>) into training sequences, enabling models to learn and utilize future goals.
Argues unsupervised object discovery is largely solved by pre-trained segmentation models (e.g., HQES, SAM). Proposes OCCAM probe framework to show OCL's focus should shift to downstream challenges like OOD generalization and compositionality using available object representations, highlighting robust foreground object selection as the new bottleneck.
Welcome to Hugo Theme Stack
