Large language models (LLMs) have shown significant improvements in many natural language processing (NLP) tasks, accelerating their rapid adoption across many industries. These models are resource-intensive, requiring extensive computational resources both during training and inference, leading to increased energy consumption and negative environmental impact. As their adoption accelerates, the sustainability of LLMs has become a critical issue, necessitating strategies to optimize their runtime efficiency without compromising performance. Hence, it is imperative to identify the parameters that significantly influence the performance and energy efficiency of LLMs. To that end, in this work, we investigate the effect of important parameters on the performance and energy efficiency of LLMs during inference and examine their trade-offs. First, we analyze how different types of models with varying numbers of parameters and architectures perform on tasks like text generation, question answering, and summarization by benchmarking LLMs such as Falcon-7B, Mistral-7B-v0.1, T5-3B, GPT-2, GPT-J-6B, and GPT-Neo-2.7B. Second, we study input and output sequence characteristics such as sequence length concerning energy consumption, performance, and throughput. Finally, we explore the impact of hardware-based power-saving techniques, i.e., Dynamic Voltage Frequency Scaling (DVFS), on the models' latency and energy efficiency. Our extensive benchmarking and statistical analysis reveal many interesting findings, uncovering how specific optimizations can reduce energy consumption while maintaining throughput and accuracy. This study provides actionable insights for researchers and practitioners to design energy-efficient LLM inference systems.

There is increasing interest in distilling task-specific knowledge from large language models (LLM) to smaller student models. Nonetheless, LLM distillation presents a dual challenge: 1) there is a high cost associated with querying the teacher LLM, such as GPT-4, for gathering an ample number of demonstrations; 2) the teacher LLM might provide imperfect outputs with a negative impact on the student's learning process. To enhance sample efficiency within resource-constrained, imperfect teacher scenarios, we propose a three-component framework leveraging three signal types. The first signal is the student's self-consistency (consistency of student multiple outputs), which is a proxy of the student's confidence. Specifically, we introduce a ``teaching assistant'' (TA) model to assess the uncertainty of both the student's and the teacher's outputs via confidence scoring, which serves as another two signals for student training. Furthermore, we propose a two-stage training schema to first warm up the student with a small proportion of data to better utilize student's signal. Experiments have shown the superiority of our proposed framework for four complex reasoning tasks. On average, our proposed two-stage framework brings a relative improvement of up to 20.79% compared to fine-tuning without any signals across datasets.

Large language models show impressive abilities in memorizing world knowledge, which leads to concerns regarding memorization of private information, toxic or sensitive knowledge, and copyrighted content. We introduce the problem of Large Scale Knowledge Washing, focusing on unlearning an extensive amount of factual knowledge. Previous unlearning methods usually define the reverse loss and update the model via backpropagation, which may affect the model's fluency and reasoning ability or even destroy the model due to extensive training with the reverse loss. Existing works introduce additional data from downstream tasks to prevent the model from losing capabilities, which requires downstream task awareness. Controlling the tradeoff of unlearning and maintaining existing capabilities is also challenging. To this end, we propose LAW (Large Scale Washing) to update the MLP layers in decoder-only large language models to perform knowledge washing, as inspired by model editing methods and based on the hypothesis that knowledge and reasoning are disentanglable. We derive a new objective with the knowledge to be unlearned to update the weights of certain MLP layers. Experimental results demonstrate the effectiveness of LAW in forgetting target knowledge while maintaining reasoning ability. The code will be open-sourced at https://github.com/wangyu-ustc/LargeScaleWashing.

Dense Retrieval (DR) is now considered as a promising tool to enhance the memorization capacity of Large Language Models (LLM) such as GPT3 and GPT-4 by incorporating external memories. However, due to the paradigm discrepancy between text generation of LLM and DR, it is still an open challenge to integrate the retrieval and generation tasks in a shared LLM. In this paper, we propose an efficient LLM-Oriented Retrieval Tuner, namely LMORT, which decouples DR capacity from base LLM and non-invasively coordinates the optimally aligned and uniform layers of the LLM towards a unified DR space, achieving an efficient and effective DR without tuning the LLM itself. The extensive experiments on six BEIR datasets show that our approach could achieve competitive zero-shot retrieval performance compared to a range of strong DR models while maintaining the generation ability of LLM.

We conjecture that hidden state vectors corresponding to individual input tokens encode information sufficient to accurately predict several tokens ahead. More concretely, in this paper we ask: Given a hidden (internal) representation of a single token at position $t$ in an input, can we reliably anticipate the tokens that will appear at positions $\geq t + 2$? To test this, we measure linear approximation and causal intervention methods in GPT-J-6B to evaluate the degree to which individual hidden states in the network contain signal rich enough to predict future hidden states and, ultimately, token outputs. We find that, at some layers, we can approximate a model's output with more than 48% accuracy with respect to its prediction of subsequent tokens through a single hidden state. Finally we present a "Future Lens" visualization that uses these methods to create a new view of transformer states.

Existing event-centric NLP models often only apply to the pre-defined ontology, which significantly restricts their generalization capabilities. This paper presents CEO, a novel Corpus-based Event Ontology induction model to relax the restriction imposed by pre-defined event ontologies. Without direct supervision, CEO leverages distant supervision from available summary datasets to detect corpus-wise salient events and exploits external event knowledge to force events within a short distance to have close embeddings. Experiments on three popular event datasets show that the schema induced by CEO has better coverage and higher accuracy than previous methods. Moreover, CEO is the first event ontology induction model that can induce a hierarchical event ontology with meaningful names on eleven open-domain corpora, making the induced schema more trustworthy and easier to be further curated.