Despite the notable advancements of existing prompting methods, such as In-Context Learning and Chain-of-Thought for Large Language Models (LLMs), they still face challenges related to various biases. Traditional debiasing methods primarily focus on the model training stage, including approaches based on data augmentation and reweighting, yet they struggle with the complex biases inherent in LLMs. To address such limitations, the causal relationship behind the prompting methods is uncovered using a structural causal model, and a novel causal prompting method based on front-door adjustment is proposed to effectively mitigate LLMs biases. In specific, causal intervention is achieved by designing the prompts without accessing the parameters and logits of LLMs. The chain-of-thought generated by LLM is employed as the mediator variable and the causal effect between input prompts and output answers is calculated through front-door adjustment to mitigate model biases. Moreover, to accurately represent the chain-of-thoughts and estimate the causal effects, contrastive learning is used to fine-tune the encoder of chain-of-thought by aligning its space with that of the LLM. Experimental results show that the proposed causal prompting approach achieves excellent performance across seven natural language processing datasets on both open-source and closed-source LLMs.

In this paper, we propose a pre-trained foundation model \textbf{FMint} (\textbf{F}oundation \textbf{M}odel based on \textbf{In}i\textbf{t}ialization), designed to speed up large-scale simulations of various differential equations with high accuracy via error correction. Human-designed simulation algorithms excel at capturing the fundamental physics of engineering problems, but often need to balance the trade-off between accuracy and efficiency. While deep learning methods offer innovative solutions across numerous scientific fields, they frequently fall short in domain-specific knowledge. FMint bridges these gaps through conditioning on the initial coarse solutions obtained from conventional human-designed algorithms, and trained to obtain refined solutions for various differential equations. Based on the backbone of large language models, we adapt the in-context learning scheme to learn a universal error correction method for dynamical systems from given prompted sequences of coarse solutions. The model is pre-trained on a corpus of 600K ordinary differential equations (ODEs), and we conduct extensive experiments on both in-distribution and out-of-distribution tasks. FMint outperforms various baselines on large-scale simulation, and demonstrates its capability in generalization to unseen ODEs. Our approach achieves an accuracy improvement of 1 to 2 orders of magnitude over state-of-the-art dynamical system simulators, and delivers a 5X speedup compared to traditional numerical algorithms.

While large language models (LLMs) have achieved significant success in various applications, they often struggle with hallucinations, especially in scenarios that require deep and responsible reasoning. These issues could be partially mitigate by integrating external knowledge graphs (KG) in LLM reasoning. However, the method of their incorporation is still largely unexplored. In this paper, we propose a retrieval-exploration interactive method, FiDelis to handle intermediate steps of reasoning grounded by KGs. Specifically, we propose Path-RAG module for recalling useful intermediate knowledge from KG for LLM reasoning. We incorporate the logic and common-sense reasoning of LLMs and topological connectivity of KGs into the knowledge retrieval process, which provides more accurate recalling performance. Furthermore, we propose to leverage deductive reasoning capabilities of LLMs as a better criterion to automatically guide the reasoning process in a stepwise and generalizable manner. Deductive verification serve as precise indicators for when to cease further reasoning, thus avoiding misleading the chains of reasoning and unnecessary computation. Extensive experiments show that our method, as a training-free method with lower computational cost and better generality outperforms the existing strong baselines in three benchmarks.

The advent of Large Language Models (LLMs) has revolutionized language understanding and human-like text generation, drawing interest from many other fields with this question in mind: What else are the LLMs capable of? Despite their widespread adoption, ongoing research continues to explore new ways to integrate LLMs into diverse systems. This paper explores new techniques to harness the power of LLMs for 6G (6th Generation) wireless communication technologies, a domain where automation and intelligent systems are pivotal. The inherent adaptability of LLMs to domain-specific tasks positions them as prime candidates for enhancing wireless systems in the 6G landscape. We introduce a novel Reinforcement Learning (RL) based framework that leverages LLMs for network deployment in wireless communications. Our approach involves training an RL agent, utilizing LLMs as its core, in an urban setting to maximize coverage. The agent's objective is to navigate the complexities of urban environments and identify the network parameters for optimal area coverage. Additionally, we integrate LLMs with Convolutional Neural Networks (CNNs) to capitalize on their strengths while mitigating their limitations. The Deep Deterministic Policy Gradient (DDPG) algorithm is employed for training purposes. The results suggest that LLM-assisted models can outperform CNN-based models in some cases while performing at least as well in others.

Large language models (LLMs) excel at few-shot in-context learning (ICL) -- learning from a few examples provided in context at inference, without any weight updates. Newly expanded context windows allow us to investigate ICL with hundreds or thousands of examples -- the many-shot regime. Going from few-shot to many-shot, we observe significant performance gains across a wide variety of generative and discriminative tasks. While promising, many-shot ICL can be bottlenecked by the available amount of human-generated examples. To mitigate this limitation, we explore two new settings: Reinforced and Unsupervised ICL. Reinforced ICL uses model-generated chain-of-thought rationales in place of human examples. Unsupervised ICL removes rationales from the prompt altogether, and prompts the model only with domain-specific questions. We find that both Reinforced and Unsupervised ICL can be quite effective in the many-shot regime, particularly on complex reasoning tasks. Finally, we demonstrate that, unlike few-shot learning, many-shot learning is effective at overriding pretraining biases, can learn high-dimensional functions with numerical inputs, and performs comparably to fine-tuning. Our analysis also reveals the limitations of next-token prediction loss as an indicator of downstream ICL performance.

Time series prediction is crucial for understanding and forecasting complex dynamics in various domains, ranging from finance and economics to climate and healthcare. Based on Transformer architecture, one approach involves encoding multiple variables from the same timestamp into a single temporal token to model global dependencies. In contrast, another approach embeds the time points of individual series into separate variate tokens. The former method faces challenges in learning variate-centric representations, while the latter risks missing essential temporal information critical for accurate forecasting. In our work, we introduce GridTST, a model that combines the benefits of two approaches using innovative multi-directional attentions based on a vanilla Transformer. We regard the input time series data as a grid, where the $x$-axis represents the time steps and the $y$-axis represents the variates. A vertical slicing of this grid combines the variates at each time step into a \textit{time token}, while a horizontal slicing embeds the individual series across all time steps into a \textit{variate token}. Correspondingly, a \textit{horizontal attention mechanism} focuses on time tokens to comprehend the correlations between data at various time steps, while a \textit{vertical}, variate-aware \textit{attention} is employed to grasp multivariate correlations. This combination enables efficient processing of information across both time and variate dimensions, thereby enhancing the model's analytical strength. % We also integrate the patch technique, segmenting time tokens into subseries-level patches, ensuring that local semantic information is retained in the embedding. The GridTST model consistently delivers state-of-the-art performance across various real-world datasets.

Aligning with human preference datasets has been critical to the success of large language models (LLMs). Reinforcement learning from human feedback (RLHF) employs a costly reward model to provide feedback for on-policy sampling responses. Recently, offline methods that directly fit responses with binary preferences in the dataset have emerged as alternatives. However, existing methods do not explicitly model preference strength information, which is crucial for distinguishing different response pairs. To overcome this limitation, we propose Online Self-Preferring (OSP) language models to learn from self-generated response pairs and self-judged preference strengths. For each prompt and corresponding self-generated responses, we introduce a ranked pairing method to construct multiple response pairs with preference strength information. We then propose the soft-preference cross-entropy loss to leverage such information. Empirically, we demonstrate that leveraging preference strength is crucial for avoiding overfitting and enhancing alignment performance. OSP achieves state-of-the-art alignment performance across various metrics in two widely used human preference datasets. OSP is parameter-efficient and more robust than the dominant online method, RLHF when limited offline data are available and generalizing to out-of-domain tasks. Moreover, OSP language models established by LLMs with proficiency in self-preferring can efficiently self-improve without external supervision.

Scaling laws have driven remarkable progress across machine learning domains like language modeling and computer vision. However, the exploration of scaling laws in embodied AI and robotics has been limited, despite the rapidly increasing usage of machine learning in this field. This paper presents the first study to quantify scaling laws for Robot Foundation Models (RFMs) and the use of LLMs in robotics tasks. Through a meta-analysis spanning 198 research papers, we analyze how key factors like compute, model size, and training data quantity impact model performance across various robotic tasks. Our findings confirm that scaling laws apply to both RFMs and LLMs in robotics, with performance consistently improving as resources increase. The power law coefficients for RFMs closely match those of LLMs in robotics, resembling those found in computer vision and outperforming those for LLMs in the language domain. We also note that these coefficients vary with task complexity, with familiar tasks scaling more efficiently than unfamiliar ones, emphasizing the need for large and diverse datasets. Furthermore, we highlight the absence of standardized benchmarks in embodied AI. Most studies indicate diminishing returns, suggesting that significant resources are necessary to achieve high performance, posing challenges due to data and computational limitations. Finally, as models scale, we observe the emergence of new capabilities, particularly related to data and model size.

Safety alignment of large language models (LLMs) has been gaining increasing attention. However, current safety-aligned LLMs suffer from the fragile and imbalanced safety mechanisms, which can still be induced to generate unsafe responses, exhibit over-safety by rejecting safe user inputs, and fail to preserve general utility after safety alignment. To this end, we propose a novel post safety alignment (PSA) method to address these inherent and emerging safety challenges, including safety enhancement, over-safety mitigation, and utility preservation. In specific, we introduce \textsc{SafePatching}, a novel framework for comprehensive and efficient PSA, where two distinct safety patches are developed on the harmful data to enhance safety and mitigate over-safety concerns, and then seamlessly integrated into the target LLM backbone without compromising its utility. Extensive experiments show that \textsc{SafePatching} achieves a more comprehensive and efficient PSA than baseline methods. It even enhances the utility of the backbone, further optimizing the balance between being helpful and harmless in current aligned LLMs. Also, \textsc{SafePatching} demonstrates its superiority in continual PSA scenarios.

The ever-growing computational complexity of Large Language Models (LLMs) necessitates efficient deployment strategies. The current state-of-the-art approaches for Post-training Quantization (PTQ) often require calibration to achieve the desired accuracy. This paper presents AdpQ, a novel zero-shot adaptive PTQ method for LLMs that achieves the state-of-the-art performance in low-precision quantization (e.g. 3-bit) without requiring any calibration data. Inspired by Adaptive LASSO regression model, our proposed approach tackles the challenge of outlier activations by separating salient weights using an adaptive soft-thresholding method. Guided by Adaptive LASSO, this method ensures that the quantized weights distribution closely follows the originally trained weights and eliminates the need for calibration data entirely, setting our method apart from popular approaches such as SpQR and AWQ. Furthermore, our method offers an additional benefit in terms of privacy preservation by eliminating any calibration or training data. We also delve deeper into the information-theoretic underpinnings of the proposed method. We demonstrate that it leverages the Adaptive LASSO to minimize the Kullback-Leibler divergence between the quantized weights and the originally trained weights. This minimization ensures the quantized model retains the Shannon information content of the original model to a great extent, guaranteeing efficient deployment without sacrificing accuracy or information. Our results achieve the same accuracy as the existing methods on various LLM benchmarks while the quantization time is reduced by at least 10x, solidifying our contribution to efficient and privacy-preserving LLM deployment.