Large Language Models

Boosting unit test generation efficiency, this study empirically evaluates various parameter-efficient fine-tuning methods on LLMs, demonstrating comparable performance to full fine-tuning at signific…

Tencent unveils Hunyuan-Large, a groundbreaking open-source MoE LLM boasting 389B parameters and 52B activated parameters, surpassing existing models in performance across various benchmarks.

DynaSaur: a novel LLM agent framework enabling dynamic action creation, surpassing prior methods with greater flexibility and top performance on the GAIA benchmark.

Sample-efficient LLM alignment achieved via a novel Thompson sampling algorithm (SEA), outperforming existing methods.

Swan & ArabicMTEB: New dialect-aware Arabic embedding models and benchmark achieve state-of-the-art performance, addressing limitations of existing multilingual models.

LibMoE: A new library streamlines MoE research by offering standardized training, evaluation, and a modular design, enabling efficient benchmarking of various MoE algorithms for LLMs.

Specialized Sparse Autoencoders (SSAEs) decode foundation models’ ‘dark matter’ features, efficiently extracting rare subdomain concepts for improved interpretability and safety.

LLaMo: a novel large molecular graph-language model seamlessly integrates molecular graph encoders and LLMs, achieving state-of-the-art performance in molecule description generation, property predict…

GlotCC: Open multilingual corpus & pipeline for minority languages, exceeding 1000 languages.

Constraint Back-translation enhances complex instruction following in LLMs by leveraging inherent constraints in existing datasets for efficient high-quality data creation.

BitStack: Dynamic LLM sizing for variable memory!

Steering large language and diffusion models is made easy and efficient via Activation Transport (ACT)! This novel framework uses optimal transport theory to precisely control model activations, leadi…

AAAR-1.0 benchmark rigorously evaluates LLMs’ ability to assist in four core research tasks, revealing both potential and limitations.

NeuZip dynamically compresses neural network weights, achieving memory-efficient training and inference without performance loss, significantly reducing the memory footprint of large language models.

M2RC-EVAL: A new massively multilingual benchmark for repository-level code completion, featuring fine-grained annotations and a large instruction dataset, enabling better evaluation of code LLMs acro…