While the numerous parameters in Large Language Models (LLMs) contribute to their superior performance, this massive scale makes them inefficient and memory-hungry. Thus, they are hard to deploy on commodity hardware, such as one single GPU. Given the memory and power constraints of such devices, model compression methods are widely employed to reduce both the model size and inference latency, which essentially trades off model quality in return for improved efficiency. Thus, optimizing this accuracy-efficiency trade-off is crucial for the LLM deployment on commodity hardware. In this paper, we introduce a new perspective to optimize this trade-off by prompting compressed models. Specifically, we first observe that for certain questions, the generation quality of a compressed LLM can be significantly improved by adding carefully designed hard prompts, though this isn't the case for all questions. Based on this observation, we propose a soft prompt learning method where we expose the compressed model to the prompt learning process, aiming to enhance the performance of prompts. Our experimental analysis suggests our soft prompt strategy greatly improves the performance of the 8x compressed LLaMA-7B model (with a joint 4-bit quantization and 50% weight pruning compression), allowing them to match their uncompressed counterparts on popular benchmarks. Also, we demonstrate that these learned prompts can be transferred across various datasets, tasks, and compression levels. Hence with this transferability, we can stitch the soft prompt to a newly compressed model to improve the test-time accuracy in an ``in-situ'' way.

Large language models (LLMs) have been increasingly used to interact with external environments (e.g., games, compilers, APIs) as goal-driven agents. However, it remains challenging for these language agents to quickly and efficiently learn from trial-and-error as traditional reinforcement learning methods require extensive training samples and expensive model fine-tuning. We propose Reflexion, a novel framework to reinforce language agents not by updating weights, but instead through linguistic feedback. Concretely, Reflexion agents verbally reflect on task feedback signals, then maintain their own reflective text in an episodic memory buffer to induce better decision-making in subsequent trials. Reflexion is flexible enough to incorporate various types (scalar values or free-form language) and sources (external or internally simulated) of feedback signals, and obtains significant improvements over a baseline agent across diverse tasks (sequential decision-making, coding, language reasoning). For example, Reflexion achieves a 91% pass@1 accuracy on the HumanEval coding benchmark, surpassing the previous state-of-the-art GPT-4 that achieves 80%. We also conduct ablation and analysis studies using different feedback signals, feedback incorporation methods, and agent types, and provide insights into how they affect performance.

The'Godfather of AI' has warned the tech will be smarter than humans in some ways by the end of the decade - and he believes it will ultimately destroy humanity. In a doom-laden interview with 60 Minutes, Geoffrey Hinton, 75, predicted that in five years, the systems will be surpass human intelligence that would lead to the rise of'killer robots,' fake news and a boom in unemployment. Hinton is a former Google executive credited with creating the technology that became the bedrock of systems like ChatGPT and Google Bard. He recently revealed his fears that the technology could go rogue and write its own code, allowing it to modify itself. While the scientist fears many aspects of the technology, he said AI has huge benefits in healthcare, such as designing drugs and recognizing medical issues.

According to Reuters, OpenAI is exploring making its own artificial intelligence chips, even looking into an acquisition. OpenAI CEO Sam Altman previously blamed GPU shortages for users' concerns regarding the company API's speed and reliability, leading to these moves. OpenAI using its own chips could reduce its costs too. Based on analysis by Bernstein Research, each ChatGPT query costs the company around four cents. At the moment, NVIDIA controls the market for chips that power AI applications. The Microsoft supercomputer OpenAI used to develop its technology, for instance, uses 10,000 NVIDIA GPUs.

Active learning aims to construct an effective training set by iteratively curating the most informative unlabeled data for annotation, which is practical in low-resource tasks. Most active learning techniques in classification rely on the model's uncertainty or disagreement to choose unlabeled data. However, previous work indicates that existing models are poor at quantifying predictive uncertainty, which can lead to over-confidence in superficial patterns and a lack of exploration. Inspired by the cognitive processes in which humans deduce and predict through causal information, we propose a novel Explainable Active Learning framework (XAL) for low-resource text classification, which aims to encourage classifiers to justify their inferences and delve into unlabeled data for which they cannot provide reasonable explanations. Specifically, besides using a pre-trained bi-directional encoder for classification, we employ a pre-trained uni-directional decoder to generate and score the explanation. A ranking loss is proposed to enhance the decoder's capability in scoring explanations. During the selection of unlabeled data, we combine the predictive uncertainty of the encoder and the explanation score of the decoder to acquire informative data for annotation. As XAL is a general framework for text classification, we test our methods on six different classification tasks. Extensive experiments show that XAL achieves substantial improvement on all six tasks over previous AL methods. Ablation studies demonstrate the effectiveness of each component, and human evaluation shows that the model trained in XAL performs surprisingly well in explaining its prediction.

The full potential of large pretrained models remains largely untapped in control domains like robotics. This is mainly because of the scarcity of data and the computational challenges associated with training or fine-tuning these large models for such applications. Prior work mainly emphasizes effective pretraining of large models for decision-making, with little exploration into how to perform data-efficient continual adaptation of these models for new tasks. Recognizing these constraints, we introduce TAIL (Task-specific Adapters for Imitation Learning), a framework for efficient adaptation to new control tasks. Inspired by recent advancements in parameter-efficient fine-tuning in language domains, we explore efficient fine-tuning techniques -- e.g., Bottleneck Adapters, P-Tuning, and Low-Rank Adaptation (LoRA) -- in TAIL to adapt large pretrained models for new tasks with limited demonstration data. Our extensive experiments in large-scale language-conditioned manipulation tasks comparing prevalent parameter-efficient fine-tuning techniques and adaptation baselines suggest that TAIL with LoRA can achieve the best post-adaptation performance with only 1\% of the trainable parameters of full fine-tuning, while avoiding catastrophic forgetting and preserving adaptation plasticity in continual learning settings.

Can Large Language Models (LLMs) simulate human behavior in complex environments? LLMs have recently been shown to exhibit advanced reasoning skills but much of NLP evaluation still relies on static benchmarks. Answering this requires evaluation environments that probe strategic reasoning in competitive, dynamic scenarios that involve long-term planning. We conduct several controlled simulations using state-of-the-art LLMs as bidding agents. We find that through simple prompting, LLMs do indeed demonstrate many of the skills needed for effectively engaging in auctions (e.g., managing budget, adhering to long-term goals and priorities), skills that we find can be sharpened by explicitly encouraging models to be adaptive and observe strategies in past auctions. These results are significant as they show the potential of using LLM agents to model intricate social dynamics, especially in competitive settings. However, we also observe considerable variability in the capabilities of individual LLMs. Notably, even our most advanced models (GPT-4) are occasionally surpassed by heuristic baselines and human agents, highlighting the potential for further improvements in the design of LLM agents and the important role that our simulation environment can play in further testing and refining agent architectures. A long-term goal of the AI community has been the development of autonomous agents that can independently make decisions and freely interact in the environment to carry out different tasks (Steels, 1995; Franklin & Graesser, 1996). Being autonomous requires an agent to have a certain set of skills, such as the ability to do complex reasoning, and manage risk and resources, among many others. Large Language Models (LLMs) have proven to be able to solve a wide range of different reasoning problems, with the boundaries of what's possible being pushed every day (Wei et al., 2022a; Bubeck et al., 2023). Despite the increasing view of these models as autonomous agents (Wang et al., 2023a; Sumers et al., 2023; Xi et al., 2023), a crucial question remains: Can these agents effectively do sequential decision-making in dynamic environments for achieving their strategic objectives? While the potential is evident (Nakajima, 2023; Significant-Gravitas, 2023), these capabilities have yet to be rigorously evaluated. Traditional reasoning and planning benchmarks in NLP (Geva et al., 2021; Sakaguchi et al., 2021; Yuan et al., 2023) mostly assess agents in static contexts. Yet, real-world scenarios demand that autonomous agents not merely respond to input but also have the ability to create long-term goals and plans, and continuously revise their decisions. To bridge this gap, one recent line of research focuses on immersing agents in simulation environments that mimic real-world scenarios (Wang et al., 2022; Park et al., 2023; Liu et al., 2023), ones that often focus on a targeted Work done during Jiangjie's internship at Allen Institute for Artificial Intelligence.

Imitation learning (IL) enables agents to mimic expert behaviors. Most previous IL techniques focus on precisely imitating one policy through mass demonstrations. However, in many applications, what humans require is the ability to perform various tasks directly through a few demonstrations of corresponding tasks, where the agent would meet many unexpected changes when deployed. In this scenario, the agent is expected to not only imitate the demonstration but also adapt to unforeseen environmental changes. This motivates us to propose a new topic called imitator learning (ItorL), which aims to derive an imitator module that can on-the-fly reconstruct the imitation policies based on very limited expert demonstrations for different unseen tasks, without any extra adjustment. In this work, we focus on imitator learning based on only one expert demonstration. To solve ItorL, we propose Demo-Attention Actor-Critic (DAAC), which integrates IL into a reinforcement-learning paradigm that can regularize policies' behaviors in unexpected situations. Besides, for autonomous imitation policy building, we design a demonstration-based attention architecture for imitator policy that can effectively output imitated actions by adaptively tracing the suitable states in demonstrations. We develop a new navigation benchmark and a robot environment for \topic~and show that DAAC~outperforms previous imitation methods \textit{with large margins} both on seen and unseen tasks.

To tackle the high inference latency exhibited by autoregressive language models, previous studies have proposed an early-exiting framework that allocates adaptive computation paths for each token based on the complexity of generating the subsequent token. However, we observed several shortcomings, including performance degradation caused by a state copying mechanism or numerous exit paths, and sensitivity to exit confidence thresholds. Consequently, we propose a Fast and Robust Early-Exiting (FREE) framework, which incorporates a shallow-deep module and a synchronized parallel decoding. Our framework enables faster inference by synchronizing the decoding process of the current token with previously stacked early-exited tokens. Furthermore, as parallel decoding allows us to observe predictions from both shallow and deep models, we present a novel adaptive threshold estimator that exploits a Beta mixture model to determine suitable confidence thresholds. We empirically demonstrated the superiority of our proposed framework on extensive generation tasks.

The emergence of tools based on Large Language Models (LLMs), such as OpenAI's ChatGPT, Microsoft's Bing Chat, and Google's Bard, has garnered immense public attention. These incredibly useful, natural-sounding tools mark significant advances in natural language generation, yet they exhibit a propensity to generate false, erroneous, or misleading content -- commonly referred to as "hallucinations." Moreover, LLMs can be exploited for malicious applications, such as generating false but credible-sounding content and profiles at scale. This poses a significant challenge to society in terms of the potential deception of users and the increasing dissemination of inaccurate information. In light of these risks, we explore the kinds of technological innovations, regulatory reforms, and AI literacy initiatives needed from fact-checkers, news organizations, and the broader research and policy communities. By identifying the risks, the imminent threats, and some viable solutions, we seek to shed light on navigating various aspects of veracity in the era of generative AI.