Large Language Models

SVD-based weight pruning surprisingly boosts in-context learning in large language models, especially when applied to deeper layers, offering a novel approach to model compression and efficiency.

OLLM: An end-to-end LLM method builds ontologies from scratch, outperforming subtask approaches and improving semantic accuracy with novel evaluation metrics.

EAGLE: Guiding LLMs using latent embeddings for controlled text generation.

Novel trajectory volatility score (TV Score) significantly improves out-of-distribution detection in mathematical reasoning by leveraging dynamic embedding trajectories, outperforming existing GLM met…

Elo rating’s reliability for LLM evaluation is challenged, revealing inconsistencies and suggesting new, more robust methods are needed for accurate model ranking.

EFFI-LEARNER: A novel self-optimization framework dramatically improves the efficiency of LLM-generated code by iteratively refining code based on execution profiles.

Novel sketching algorithms enable scalable gradient and Hessian analysis for large language models, revealing insights into their intrinsic dimensionality and challenging existing assumptions.

TRIPLE: Efficient prompt optimization using fixed-budget best-arm identification.

LoRA-Inlaid: a novel multi-task LLM serving system boosts throughput by 1.58x, latency by 1.76x, and job completion time by 2x, while improving SLO attainment by 10x, all while maintaining model quali…

PromptEval efficiently estimates LLM performance across many prompts, providing robust performance metrics and enabling reliable LLM comparisons.

Learning to rank request outputs improves LLM scheduling, resulting in 2.8x lower chatbot latency and 6.5x higher synthetic data generation throughput.

Adaptive Dense-to-sparse Constrained Optimization (ADC) efficiently jailbreaks LLMs by transforming discrete token optimization into a continuous process, achieving higher success rates than existing …

LLaVolta significantly boosts multi-modal LLMs by using visual context compression, achieving substantial training cost reduction and enhanced inference efficiency without performance loss.

TREACLE: a reinforcement learning policy efficiently selects LLMs and prompts, achieving up to 85% cost savings while maintaining high accuracy in answering reasoning questions.

Boosting LLM robustness against attacks efficiently: Continuous adversarial training in embedding space outperforms discrete methods, achieving improved robustness with less computation.

This paper presents a novel watermarking scheme for language models that is both undetectable and robust to a constant fraction of adversarial edits (insertions, deletions, substitutions).

AI alignment beyond human supervision is achieved via easy-to-hard generalization: training reward models on easy tasks to effectively evaluate and improve generators on harder tasks, achieving superh…

LLMs’ emotional decision-making is assessed using a novel framework, EAI, showing that emotions significantly alter ethical and strategic choices in games. This reveals crucial biases, necessitati…

DuQuant: Dual transformations distribute outliers for stronger quantized LLMs.

DropBP: Accelerate LLM fine-tuning by 44% while preserving accuracy!