Large Language Models

LLMs self-discover optimal reasoning structures for complex problems, boosting performance by up to 32% compared to existing methods.

Certified selective generation controls language model hallucinations by leveraging textual entailment and a novel semi-supervised algorithm, guaranteeing a controlled false discovery rate.

Enhance Transformer models via Selective Self-Attention (SSA), a principled context control method that boosts accuracy and efficiency.

SelectIT leverages LLMs’ intrinsic uncertainty to efficiently select high-quality instruction tuning data, enhancing model performance without extra resources.

This paper presents novel statistical methods to reliably watermark and segment LLMs-generated text, ensuring source traceability even after user modifications.

Training-free architecture search finds optimal subnets in LLMs, boosting inference speed and slashing memory needs without retraining.

SDP4Bit achieves up to 4.08x speedup in LLM training by quantizing weight differences and gradients to ~4 bits, maintaining accuracy.

Massive language models improve with bigger datastores at inference time. A 1.4 trillion-token datastore, MASSIVEDS, shows that retrieval-based LMs outperform larger, solely-trained models on knowled…

Boosting LLM performance: This research shows how larger language models need bigger vocabularies for optimal efficiency and performance.

Direct Alignment Algorithms (DAAs) for LLM alignment suffer from over-optimization, even without explicit reward models; this paper empirically demonstrates this and proposes scaling laws to understan…

Revolutionizing LLM training: Constant learning rate with cooldown replaces cosine schedule, enabling cost-effective scaling experiments!

SaulLM-54B & SaulLM-141B achieve state-of-the-art performance on legal tasks by scaling up model size, employing a specialized instruction-following protocol, and aligning model outputs with human pre…

SAFEWORLD: a new benchmark reveals and fixes LLMs’ struggle with diverse safety standards.

S2FT: Structured Sparse Fine-Tuning achieves state-of-the-art LLM fine-tuning performance, training efficiency, and inference scalability by selecting sparsely and computing densely.

S-STE achieves efficient 2:4 sparse pre-training by introducing a novel continuous pruning function, overcoming the limitations of previous methods and leading to improved accuracy and speed.

LLMs struggle with out-of-distribution (OOD) generalization. This research introduces ‘rule extrapolation’ using formal languages to rigorously evaluate OOD behavior in various LLM architectures, rev…

RouterDC: A query-based router trained via dual contrastive learning assembles multiple LLMs, significantly outperforming individual LLMs and existing routing methods on both in- and out-of-distributi…

Robust Prompt Optimization (RPO) creates robust LLM defenses against jailbreaking attacks by optimizing a transferable suffix, achieving state-of-the-art robustness.

Leveraging model-generated synthetic data for LLM finetuning significantly improves efficiency when using both positive and strategically constructed negative examples, resulting in an eight-fold incr…

Risk-Averse RLHF fine-tunes LLMs to minimize toxic outputs while maintaining performance.