Natural Language Processing

LoQT enables efficient large language model training on consumer hardware via quantized weights and low-rank weight updates, overcoming memory limitations.

New dataset MUSIC-AVQA-R and a multi-faceted cycle collaborative debiasing strategy significantly improve audio-visual question answering robustness.

LLMs often generate factually inaccurate long-form text. This work introduces LongFact, a new benchmark dataset of 2280 fact-seeking prompts, and SAFE, a novel automated evaluation method that outperf…

Loki: Low-rank Keys for Efficient Sparse Attention accelerates attention mechanisms in LLMs by exploiting the low-dimensionality of key vectors. It dynamically selects key tokens based on approximate…

LOFIT: Localized fine-tuning boosts LLMs’ performance by selectively training only a small subset of attention heads, achieving comparable accuracy to other methods while using significantly fewer par…

Transformers’ learning dynamics depend heavily on initialization and Markovian data properties, leading to either global or local minima; this paper proves this, offers initialization guidelines, and …

TEA-GLM leverages LLMs for zero-shot graph learning by aligning GNN representations with LLM token embeddings, achieving state-of-the-art performance on unseen datasets and tasks.

LLMDFA: A novel LLM-powered framework performs compilation-free and customizable dataflow analysis, achieving high accuracy in bug detection by decomposing the task into sub-problems and mitigating L…

LLM-Check efficiently detects LLM hallucinations in a single response, using internal model analysis, improving real-time applications.

LLM Processes leverage LLMs to create probabilistic regression models guided by natural language, enabling seamless integration of expert knowledge and improving prediction accuracy.

LLM dataset inference reliably detects if a dataset was used in training, overcoming limitations of existing membership inference attacks.

LLM circuit analyses remain consistent across model scales and extensive training, enabling more efficient interpretability research.

LLaNA: A novel Multimodal Large Language Model directly processes NeRF weights to enable NeRF captioning and Q&A, outperforming traditional 2D/3D-based methods.

LLaMo, a novel Large Language Model-based Molecular Graph Assistant, uses multi-level graph projection and instruction tuning to achieve superior performance on diverse molecular tasks.

LIVE, a novel learnable in-context vector, significantly improves visual question answering by reducing computational costs and enhancing accuracy compared to traditional ICL methods.

Lisa: a novel lazy safety alignment method safeguards LLMs against harmful fine-tuning attacks by introducing a proximal term to constrain model drift, significantly improving alignment performance.

LISA, a layerwise importance sampling method, dramatically improves memory-efficient large language model fine-tuning, outperforming existing methods while using less GPU memory.

Transformers’ in-context learning mirrors human episodic memory, with specific attention heads acting like the brain’s contextual maintenance and retrieval system.

Scaling causal language models reveals a connection between neural collapse properties, model size, and improved generalization, highlighting NC’s broader relevance to LLMs.

Large Language Models struggle with compositional tasks, requiring exponentially more data than expected for learning compared to learning sub-tasks individually. This paper reveals surprising sample …