Link to ICLR

TL;DR: We present a method to evaluate the alignment between the decision-making logic of Large Language Models (LLMs) and human cognition in a case study on legal LLMs.

Link to arxiv Link to AAAI

TL;DR: We propose TransMamba, a novel sequence-level hybrid framework that unifies Transformer and Mamba by sharing parameter matrices and memory converter. Extensive experiments demonstrates that TransMamba achieves superior training efficiency and performance compared to single and hybrid baselines.

Link to arxiv

TL;DR: DeepMath-Creative introduces a benchmark and evaluation criteria for mathematical creativity, concluding that current LLMs rely more on learned-pattern recombination than genuine creative insight.

Link to NAACL

TL;DR: We propose a simple yet effective framework for continuous speech tokenizer and a text-tospeech model based on continuous speech tokens, which shows greater performance and better information preservation rate.

Link to arxiv

TL;DR: We propose a novel LLMs knowledge distillation framework named direct preference knowledge distillation and demonstrate the effectiveness of preference and implicit reward in the KD process.

Link to arxiv Link to ICML

TL;DR: This paper examines floating-point quantization in LLM training, proposes a unified scaling law, and offers key recommendations for improving model performance and cost-efficiency.

Link to arxiv Link to CVPR

TL;DR: We build a verb-taxonomy action semantic space to align existing datasets into a unified ‘Pangea’ label system and propose a physical-to-semantic model that leverages it, delivering consistent improvements, strongest in transfer learning.