Large Language Models

RE-Control: Aligning LLMs via dynamic representation editing using optimal control theory, achieving superior alignment with significantly fewer resources than fine-tuning.

Aligner efficiently aligns LLMs by learning to correct initial responses, achieving significant improvements in helpfulness and harmlessness across various models with resource efficiency.

ALI-Agent uses LLM-powered agents for in-depth, adaptive assessment of LLMs’ alignment with human values, overcoming limitations of existing static benchmarks.

Language model algorithms have improved drastically, halving compute needs every 8 months since 2012, surpassing Moore’s Law; however, compute scaling, not algorithms, drove most recent performance ga…

Randomly initialized transformers, with only embedding layers optimized, surprisingly excel at various algorithmic tasks, revealing inherent capabilities even before training.

AlchemistCoder enhances code LLMs by pioneering hindsight tuning on multi-source data, harmonizing conflicting styles via AlchemistPrompts, and achieving state-of-the-art performance.

AGENTPOISON: A novel backdoor attack compromises LLM agents by poisoning their memory or knowledge bases, achieving high success rates with minimal performance impact.

This paper introduces a parametric World Knowledge Model (WKM) to improve AI agent planning by integrating both global task knowledge and dynamic state knowledge, thereby overcoming current LLMs’ limi…

Adversarial Representation Engineering (ARE) offers a unified, interpretable approach for editing large language models (LLMs) by using a representation sensor as an editing oracle, enhancing model sa…

Boosting LLM efficiency, this study introduces adversarial moment-matching distillation, outperforming existing methods by matching action-value moments for superior knowledge transfer and achieving s…

TP-LLaMA boosts tool-augmented LLMs by optimizing inference trajectories using preference learning from both successful and failed attempts, achieving superior performance and efficiency.

Adaptive Layer Sparsity (ALS) revolutionizes large language model (LLM) compression by intelligently pruning less important layers, achieving significant size reduction without performance loss. It o…

Ctrl-G: A neuro-symbolic framework enables adaptable control of LLM generation by combining any LLM with a Hidden Markov Model (HMM), ensuring outputs adhere to logical constraints specified as determ…

This paper introduces segment auctions, maximizing logarithmic social welfare, for integrating ads into LLM outputs via Retrieval Augmented Generation, balancing ad revenue and output quality.

LLM compression accuracy hides crucial behavioral changes; use % flips and KL-divergence for better evaluation.

Probe sampling accelerates Greedy Coordinate Gradient (GCG) and other prompt optimization methods by up to 5.6x, achieving equal or better attack success rates, making LLM safety research faster and m…

Boost LLM inference speed by 3x! This paper refines blockwise parallel decoding (BPD) by cleverly refining draft predictions, resulting in faster text generation for large language models.

Transformers exhibit abrupt learning: training loss plateaus, then suddenly drops. This study uses matrix completion to demonstrate this phenomenon, providing insights into the model’s algorithmic sh…

LLMs improve through self-correction, but the mechanisms are unclear. This paper provides a theoretical framework and empirical evidence demonstrating that self-correction arises from in-context align…

This paper theoretically analyzes speculative decoding, revealing its optimality and providing formulas for expected rejections, paving the way for more efficient large language model inference.