Browsing: Hugging Face
Theoretical analysis of Direct Preference Optimization (DPO) reveals that log-ratio reward parameterization is optimal for learning target policy via preference…
The difficulty-aware prompting method shortens reasoning traces in a dataset, improving model performance and efficiency across various benchmarks. Existing chain-of-thought…
A novel block-wise approximate KV Cache and confidence-aware parallel decoding strategy improve the inference speed of diffusion-based large language models…
TrustVLM enhances the reliability of Vision-Language Models by estimating prediction trustworthiness without retraining, improving misclassification detection in multimodal tasks. Vision-Language…
A novel pairwise-comparison framework using CreataSet dataset trains CrEval, an LLM-based evaluator that significantly improves the assessment of textual creativity…
ChartLens enhances multimodal language models with fine-grained visual attributions, improving the accuracy of chart understanding by 26-66%. The growing capabilities…
(based on a thread on Twitter) Preferences drive modern LLM research and development: from model alignment to evaluation.But how well…
The introduction of SridBench, a benchmark for scientific figure generation, reveals that current top-tier models, such as GPT-4o-image, fall short…
CheXStruct and CXReasonBench evaluate Large Vision-Language Models in clinical diagnosis by assessing structured reasoning, visual grounding, and generalization using the…
The study explores the structural patterns of knowledge in large language models from a graph perspective, uncovering knowledge homophily and…