In the era of Large Language Models (LLMs), Knowledge Distillation (KD) emerges as a pivotal methodology for transferring advanced capabilities from leading proprietary LLMs, such as GPT-4, to their open-source counterparts like LLaMA and Mistral. Additionally, as open-source LLMs flourish, KD plays a crucial role in both compressing these models, and facilitating their self-improvement by employing themselves as teachers. This paper presents a comprehensive survey of KD's role within the realm of LLM, highlighting its critical function in imparting advanced knowledge to smaller models and its utility in model compression and self-improvement. Our survey is meticulously structured around three foundational pillars: \textit{algorithm}, \textit{skill}, and \textit{verticalization} -- providing a comprehensive examination of KD mechanisms, the enhancement of specific cognitive abilities, and their practical implications across diverse fields. Crucially, the survey navigates the intricate interplay between data augmentation (DA) and KD, illustrating how DA emerges as a powerful paradigm within the KD framework to bolster LLMs' performance. By leveraging DA to generate context-rich, skill-specific training data, KD transcends traditional boundaries, enabling open-source models to approximate the contextual adeptness, ethical alignment, and deep semantic insights characteristic of their proprietary counterparts. This work aims to provide an insightful guide for researchers and practitioners, offering a detailed overview of current methodologies in KD and proposing future research directions. Importantly, we firmly advocate for compliance with the legal terms that regulate the use of LLMs, ensuring ethical and lawful application of KD of LLMs. An associated Github repository is available at https://github.com/Tebmer/Awesome-Knowledge-Distillation-of-LLMs.

We explore how iterative revising a chain of thoughts with the help of information retrieval significantly improves large language models' reasoning and generation ability in long-horizon generation tasks, while hugely mitigating hallucination. In particular, the proposed method -- *retrieval-augmented thoughts* (RAT) -- revises each thought step one by one with retrieved information relevant to the task query, the current and the past thought steps, after the initial zero-shot CoT is generated. Applying RAT to GPT-3.5, GPT-4, and CodeLLaMA-7b substantially improves their performances on various long-horizon generation tasks; on average of relatively increasing rating scores by 13.63% on code generation, 16.96% on mathematical reasoning, 19.2% on creative writing, and 42.78% on embodied task planning. The demo page can be found at https://craftjarvis.github.io/RAT

In the field of Learning from Demonstration (LfD), Dynamical Systems (DSs) have gained significant attention due to their ability to generate real-time motions and reach predefined targets. However, the conventional convergence-centric behavior exhibited by DSs may fall short in safety-critical tasks, specifically, those requiring precise replication of demonstrated trajectories or strict adherence to constrained regions even in the presence of perturbations or human intervention. Moreover, existing DS research often assumes demonstrations solely in Euclidean space, overlooking the crucial aspect of orientation in various applications. To alleviate these shortcomings, we present an innovative approach geared toward ensuring the safe execution of learned orientation skills within constrained regions surrounding a reference trajectory. This involves learning a stable DS on SO(3), extracting time-varying conic constraints from the variability observed in expert demonstrations, and bounding the evolution of the DS with Conic Control Barrier Function (CCBF) to fulfill the constraints. We validated our approach through extensive evaluation in simulation and showcased its effectiveness for a cutting skill in the context of assisted teleoperation.

This paper studies Reinforcement Learning (RL) techniques to enable team coordination behaviors in graph environments with support actions among teammates to reduce the costs of traversing certain risky edges in a centralized manner. While classical approaches can solve this non-standard multi-agent path planning problem by converting the original Environment Graph (EG) into a Joint State Graph (JSG) to implicitly incorporate the support actions, those methods do not scale well to large graphs and teams. To address this curse of dimensionality, we propose to use RL to enable agents to learn such graph traversal and teammate supporting behaviors in a data-driven manner. Specifically, through a new formulation of the team coordination on graphs with risky edges problem into Markov Decision Processes (MDPs) with a novel state and action space, we investigate how RL can solve it in two paradigms: First, we use RL for a team of agents to learn how to coordinate and reach the goal with minimal cost on a single EG. We show that RL efficiently solves problems with up to 20/4 or 25/3 nodes/agents, using a fraction of the time needed for JSG to solve such complex problems; Second, we learn a general RL policy for any $N$-node EGs to produce efficient supporting behaviors. We present extensive experiments and compare our RL approaches against their classical counterparts.

Large Language Models (LLMs) harness extensive data from the Internet, storing a broad spectrum of prior knowledge. While LLMs have proven beneficial as decision-making aids, their reliability is hampered by limitations in reasoning, hallucination phenomenon, and so on. On the other hand, Monte-Carlo Tree Search (MCTS) is a heuristic search algorithm that provides reliable decision-making solutions, achieved through recursive rollouts and self-play. However, the effectiveness of MCTS relies heavily on heuristic pruning and external value functions, particularly in complex decision scenarios. This work introduces an innovative approach that bolsters LLMs with MCTS self-play to efficiently resolve deterministic turn-based zero-sum games (DTZG), such as chess and go, without the need for additional training. Specifically, we utilize LLMs as both action pruners and proxies for value functions without the need for additional training. We theoretically prove that the suboptimality of the estimated value in our proposed method scales with $\tilde{\mathcal O}\Bigl(\frac{|\tilde {\mathcal A}|}{\sqrt{N}} + \epsilon_\mathrm{pruner} + \epsilon_\mathrm{critic}\Bigr)$, where \(N\) is the number of simulations, $|\tilde {\mathcal A}|$ is the cardinality of the pruned action space by LLM, and $\epsilon_\mathrm{pruner}$ and $\epsilon_\mathrm{critic}$ quantify the errors incurred by adopting LLMs as action space pruner and value function proxy, respectively. Our experiments in chess and go demonstrate the capability of our method to address challenges beyond the scope of MCTS and improve the performance of the directly application of LLMs.

In the realm of social media, understanding and predicting post reach is a significant challenge. This paper presents a Crowd Reaction AssessMent (CReAM) task designed to estimate if a given social media post will receive more reaction than another, a particularly essential task for digital marketers and content writers. We introduce the Crowd Reaction Estimation Dataset (CRED), consisting of pairs of tweets from The White House with comparative measures of retweet count. The proposed Generator-Guided Estimation Approach (GGEA) leverages generative Large Language Models (LLMs), such as ChatGPT, FLAN-UL2, and Claude, to guide classification models for making better predictions. Our results reveal that a fine-tuned FLANG-RoBERTa model, utilizing a cross-encoder architecture with tweet content and responses generated by Claude, performs optimally. We further use a T5-based paraphraser to generate paraphrases of a given post and demonstrate GGEA's ability to predict which post will elicit the most reactions. We believe this novel application of LLMs provides a significant advancement in predicting social media post reach.

A major factor in the recent success of large language models is the use of enormous and ever-growing text datasets for unsupervised pre-training. However, naively training a model on all available data may not be optimal (or feasible), as the quality of available text data can vary. Filtering out data can also decrease the carbon footprint and financial costs of training models by reducing the amount of training required. Data selection methods aim to determine which candidate data points to include in the training dataset and how to appropriately sample from the selected data points. The promise of improved data selection methods has caused the volume of research in the area to rapidly expand. However, because deep learning is mostly driven by empirical evidence and experimentation on large-scale data is expensive, few organizations have the resources for extensive data selection research. Consequently, knowledge of effective data selection practices has become concentrated within a few organizations, many of which do not openly share their findings and methodologies. To narrow this gap in knowledge, we present a comprehensive review of existing literature on data selection methods and related research areas, providing a taxonomy of existing approaches. By describing the current landscape of research, this work aims to accelerate progress in data selection by establishing an entry point for new and established researchers. Additionally, throughout this review we draw attention to noticeable holes in the literature and conclude the paper by proposing promising avenues for future research.

In the deployment of deep neural models, how to effectively and automatically find feasible deep models under diverse design objectives is fundamental. Most existing neural architecture search (NAS) methods utilize surrogates to predict the detailed performance (e.g., accuracy and model size) of a candidate architecture during the search, which however is complicated and inefficient. In contrast, we aim to learn an efficient Pareto classifier to simplify the search process of NAS by transforming the complex multi-objective NAS task into a simple Pareto-dominance classification task. To this end, we propose a classification-wise Pareto evolution approach for one-shot NAS, where an online classifier is trained to predict the dominance relationship between the candidate and constructed reference architectures, instead of using surrogates to fit the objective functions. The main contribution of this study is to change supernet adaption into a Pareto classifier. Besides, we design two adaptive schemes to select the reference set of architectures for constructing classification boundary and regulate the rate of positive samples over negative ones, respectively. We compare the proposed evolution approach with state-of-the-art approaches on widely-used benchmark datasets, and experimental results indicate that the proposed approach outperforms other approaches and have found a number of neural architectures with different model sizes ranging from 2M to 6M under diverse objectives and constraints.

Despite the recent advances in model compression techniques for Model compression techniques such as quantization have been deep neural networks, deploying such models on ultra-low-power successfully applied to Convolutional Neural Networks (CNNs), embedded devices still proves challenging. In particular, quantization allowing them to be deployed on embedded devices with limited schemes for Gated Recurrent Units (GRU) are difficult to computational resources. Remarkably, the quantization of RNNs tune due to their dependence on an internal state, preventing them has not been explored as extensively, potentially due to the additional from fully benefiting from sub-8bit quantization. In this work, we complexity introduced by their recurrent nature. Among the propose a modular integer quantization scheme for GRUs where the most notable works, [1] propose binary, ternary, and quaternary bit width of each operator can be selected independently. We then quantization schemes for RNNs and evaluate it on sentiment analysis, employ Genetic Algorithms (GA) to explore the vast search space [11] combines structural pruning and 8-bit quantization to of possible bit widths, simultaneously optimizing for model size optimize LSTMs for speech enhancement on a Cortex-M7 embedded and accuracy. We evaluate our methods on four different sequential platform, [20] presents quantization schemes for the standard tasks and demonstrate that mixed-precision solutions exceed LSTM and its variants, based on fixed-point arithmetic, evaluating homogeneous-precision ones in terms of Pareto efficiency. Our them on speech recognition; finally [26] employs mixed-precision results show a model size reduction between 25% and 55% while FP16 and 8-bit integer quantization to deploy speech enhancement maintaining an accuracy comparable with the 8-bit homogeneous models based on LSTMs or GRUs on a RISC-V embedded target.

The efficient collection of samples is an important factor in outdoor information gathering applications on account of high sampling costs such as time, energy, and potential destruction to the environment. Utilization of available a-priori data can be a powerful tool for increasing efficiency. However, the relationships of this data with the quantity of interest are often not known ahead of time, limiting the ability to leverage this knowledge for improved planning efficiency. To this end, this work combines transfer learning and active learning through a Multi-Task Gaussian Process and an information-based objective function. Through this combination it can explore the space of hypothetical inter-quantity relationships and evaluate these hypotheses in real-time, allowing this new knowledge to be immediately exploited for future plans. The performance of the proposed method is evaluated against synthetic data and is shown to evaluate multiple hypotheses correctly. Its effectiveness is also demonstrated on real datasets. The technique is able to identify and leverage hypotheses which show a medium or strong correlation to reduce prediction error by a factor of 1.4--3.4 within the first 7 samples, and poor hypotheses are quickly identified and rejected eventually having no adverse effect.