Large Language Models

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.

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 …

This paper presents a fully automated method for PDDL translation and planning using LLMs and environment interaction, achieving a 66% success rate on challenging PDDL domains.

LEDEX: A novel training framework significantly boosts LLMs’ code self-debugging by using automated data collection, supervised fine-tuning, and reinforcement learning, leading to more accurate code a…

Language models excel at generating programs for algorithmic tasks, but struggle with soft reasoning. COGEX leverages pseudo-programs and program emulation to tackle these tasks, while COTACS searches…

Large language models surprisingly solve unseen arithmetic tasks; this work reveals how they learn to compose simple skills into complex ones through in-context learning, showing a transition from mem…

RL-instructed language models excel at strategic communication in One Night Ultimate Werewolf, demonstrating the importance of discussion tactics in complex games.

Goal-conditioned contrastive learning boosts language reward model performance and enables better control of language model generation.

Learn-To-be-Efficient (LTE) trains LLMs to achieve structured sparsity, boosting inference speed by 25% at 50% sparsity without sacrificing accuracy.

LB-CL: A novel parameter-efficient continual learning method for LLMs that boosts performance and reduces forgetting by leveraging parametric knowledge transfer and maintaining orthogonal low-rank sub…

LaPael improves LLM knowledge injection by applying learned noise to early layers, enabling diverse and efficient knowledge updates without repeated external model usage.

LLM-DA dynamically adapts LLM-generated rules for accurate, interpretable temporal knowledge graph reasoning, significantly improving accuracy without fine-tuning.

Teach LLMs uncertainty for reliable high-stakes predictions: Fine-tuning with graded examples significantly improves LLM’s uncertainty calibration and generalizes well.

New conformal inference methods enhance LLM validity by providing adaptive validity guarantees and improving the quality of LLM outputs, addressing prior methods’ limitations.

ECO prompts enable efficient LLM unlearning by corrupting prompts flagged for forgetting, achieving promising results across various LLMs and tasks with minimal side effects.