The potential of machine learning research to create a lasting impact in academic research is clear. Recent developments in the field have seen the emergence of new techniques and approaches that are pushing the boundaries of what is possible. From auto-regressive next-token predictors to Graph Neural Networks, these advances are enabling us to solve complex tasks and display non-trivial performance on text generation and arithmetic tasks. Native Language Identification (NLI) is another area where machine learning research is making a difference. A new approach using Big Bird embeddings to classify an author's native language has been developed, outperforming traditional linguistic feature engineering models. SafetyBench is a comprehensive benchmark for evaluating the safety of Large Language Models (LLMs) that consists of 11,435 multiple choice questions across 7 categories of safety concerns. Tests of 25 popular LLMs show GPT-4 has a performance advantage, and there is room for improvement in safety. Multi-modal large language models (MLLMs) are also being developed, with a new technique for training them leading to improved truthfulness and ethical alignment in pure NLP tasks.
This paper presents a theoretical framework for studying auto-regressive next-token predictors, demonstrating that even simple models can approximate any function efficiently computed by a Turing machine. The potential for these techniques to create a lasting impact in academic research is clear, as they can be used to solve complex tasks and display non-trivial performance on text generation and arithmetic tasks.
This paper presents a new approach to Native Language Identification (NLI) that uses Big Bird embeddings to classify an author's native language. The results show that this method outperforms traditional linguistic feature engineering models, and is computationally efficient. This could have a lasting impact on NLI research, providing a promising avenue for future work.
SafetyBench is a comprehensive benchmark for evaluating the safety of Large Language Models (LLMs). It consists of 11,435 multiple choice questions across 7 categories of safety concerns, and is available in both Chinese and English. Tests of 25 popular LLMs show GPT-4 has a performance advantage, and there is room for improvement in safety. SafetyBench will enable fast and comprehensive evaluation of LLMs' safety, and help create safer LLMs.
This study presents a new technique for training multi-modal large language models (MLLMs) which can lead to improved truthfulness and ethical alignment in pure NLP tasks. Results suggest that visual instruction tuning can surpass the performance of models fine-tuned with human annotations. The potential for this technique to create a lasting impact in academic research is promising.
This paper presents a novel Graph Neural Network (GNN) based approach for predicting the expressibility of Parameterized Quantum Circuits (PQCs). The proposed method is shown to outperform existing techniques, with an RMSE of 0.03 and 0.06 on two datasets. This could have a lasting impact on academic research, as it could enable more efficient and accurate quantum machine learning and optimization algorithms.
This paper presents a novel inference method, RAIN, which allows pre-trained LLMs to align themselves with human preferences without finetuning. RAIN integrates self-evaluation and rewind mechanisms to produce responses consistent with human preferences. Results show RAIN can improve the harmlessness rate of LLaMA 30B from 82% to 97%, and reduce the attack success rate from 94% to 19%. This has the potential to create a lasting impact in academic research of the described techniques.
DNNShifter is an efficient DNN pruning system for edge computing that produces model variants with near-similar accuracy to the original dense model, while being up to 93x faster to generate and up to 5.14x smaller in size. This system has the potential to create a lasting impact in academic research by providing a fast and efficient way to generate lightweight models for edge computing.
This paper presents TopMost, a Topic Modeling System Toolkit, which provides a comprehensive suite of tools for topic modeling research and applications. It covers the entire lifecycle of topic modeling, from dataset pre-processing to model training, testing, and evaluation. This toolkit enables quick utilization, fair comparisons, and flexible extensions of different topic models, which can create a lasting impact in academic research.
This paper presents a hierarchical sequence labeling approach to extract information from resumes, which outperforms existing methods. Experiments on seven languages demonstrate the potential for this technique to create a lasting impact in academic research, with improved performance and resource efficiency.
Hydra is a multi-head low-rank adaptation method for parameter efficient fine-tuning of large-scale foundation models. It combines parallel and sequential branches to explore a broader range of optimal points and explicitly leverages pre-trained weights. Experiments demonstrate its efficiency and superior performance, suggesting its potential to create a lasting impact in academic research.