Large language models (LLMs) have played a pivotal role in building communicative AI to imitate human behaviors but face the challenge of efficient customization. To tackle this challenge, recent studies have delved into the realm of model editing, which manipulates specific memories of language models and changes the related language generation. However, the robustness of model editing remains an open question. This work seeks to understand the strengths and limitations of editing methods, thus facilitating robust, realistic applications of communicative AI. Concretely, we conduct extensive analysis to address the three key research questions. Q1: Can edited LLMs behave consistently resembling communicative AI in realistic situations? Q2: To what extent does the rephrasing of prompts lead LLMs to deviate from the edited knowledge memory? Q3: Which knowledge features are correlated with the performance and robustness of editing? Our experimental results uncover a substantial disparity between existing editing methods and the practical application of LLMs. On rephrased prompts that are complex and flexible but common in realistic applications, the performance of editing experiences a significant decline. Further analysis shows that more popular knowledge is memorized better, easier to recall, and more challenging to edit effectively.

Accurate solar power forecasting is crucial to integrate photovoltaic plants into the electric grid, schedule and secure the power grid safety. This problem becomes more demanding for those newly installed solar plants which lack sufficient data. Current research predominantly relies on historical solar power data or numerical weather prediction in a single-modality format, ignoring the complementary information provided in different modalities. In this paper, we propose a multi-modality fusion framework to integrate historical power data, numerical weather prediction, and satellite images, significantly improving forecast performance. We introduce a vector quantized framework that aligns modalities with varying information densities, striking a balance between integrating sufficient information and averting model overfitting. Our framework demonstrates strong zero-shot forecasting capability, which is especially useful for those newly installed plants. Moreover, we collect and release a multi-modal solar power (MMSP) dataset from real-world plants to further promote the research of multi-modal solar forecasting algorithms. Our extensive experiments show that our model not only operates with robustness but also boosts accuracy in both zero-shot forecasting and scenarios rich with training data, surpassing leading models. We have incorporated it into our eForecaster platform and deployed it for more than 300 solar plants with a capacity of over 15GW.

The aim of this study is to investigate Machine Unlearning (MU), a burgeoning field focused on addressing concerns related to neural models inadvertently retaining personal or sensitive data. Here, a novel approach is introduced to achieve precise and selective forgetting within language models. Unlike previous methodologies that adopt completely opposing training objectives, this approach aims to mitigate adverse effects on language model performance, particularly in generation tasks. Furthermore, two innovative evaluation metrics are proposed: Sensitive Information Extraction Likelihood (S-EL) and Sensitive Information Memory Accuracy (S-MA), designed to gauge the effectiveness of sensitive information elimination. To reinforce the forgetting framework, an effective method for annotating sensitive scopes is presented, involving both online and offline strategies. The online selection mechanism leverages language probability scores to ensure computational efficiency, while the offline annotation entails a robust two-stage process based on Large Language Models (LLMs).

Frequently Asked Questions (FAQs) refer to the most common inquiries about specific content. They serve as content comprehension aids by simplifying topics and enhancing understanding through succinct presentation of information. In this paper, we address FAQ generation as a well-defined Natural Language Processing (NLP) task through the development of an end-to-end system leveraging text-to-text transformation models. We present a literature review covering traditional question-answering systems, highlighting their limitations when applied directly to the FAQ generation task. We propose our system capable of building FAQs from textual content tailored to specific domains, enhancing their accuracy and relevance. We utilise self-curated algorithms for obtaining optimal representation of information to be provided as input and also for ranking the question-answer pairs to maximise human comprehension. Qualitative human evaluation showcases the generated FAQs to be well-constructed and readable, while also utilising domain-specific constructs to highlight domain-based nuances and jargon in the original content.

In this paper, we propose R$^3$: Learning Reasoning through Reverse Curriculum Reinforcement Learning (RL), a novel method that employs only outcome supervision to achieve the benefits of process supervision for large language models. The core challenge in applying RL to complex reasoning is to identify a sequence of actions that result in positive rewards and provide appropriate supervision for optimization. Outcome supervision provides sparse rewards for final results without identifying error locations, whereas process supervision offers step-wise rewards but requires extensive manual annotation. R$^3$ overcomes these limitations by learning from correct demonstrations. Specifically, R$^3$ progressively slides the start state of reasoning from a demonstration's end to its beginning, facilitating easier model exploration at all stages. Thus, R$^3$ establishes a step-wise curriculum, allowing outcome supervision to offer step-level signals and precisely pinpoint errors. Using Llama2-7B, our method surpasses RL baseline on eight reasoning tasks by $4.1$ points on average. Notebaly, in program-based reasoning on GSM8K, it exceeds the baseline by $4.2$ points across three backbone models, and without any extra data, Codellama-7B + R$^3$ performs comparable to larger models or closed-source models.

Utilitarian games such as dictator games to measure fairness have been studied in the social sciences for decades. These games have given us insight into not only how humans view fairness but also in what conditions the frequency of fairness, altruism and greed increase or decrease. While these games have traditionally been focused on humans, the rise of AI gives us the ability to study how these models play these games. AI is becoming a constant in human interaction and examining how these models portray fairness in game play can give us some insight into how AI makes decisions. Over 101 rounds of the dictator game, I conclude that AI has a strong sense of fairness that is dependant of it it deems the person it is playing with as trustworthy, framing has a strong effect on how much AI gives a recipient when designated the trustee, and there may be evidence that AI experiences inequality aversion just as humans.

We analyze the capabilities of Transformer language models on learning discrete algorithms. To this end, we introduce two new tasks demanding the composition of several discrete sub-tasks. On both training LLaMA models from scratch and prompting on GPT-4 and Gemini we measure learning compositions of learned primitives. We observe that the compositional capabilities of state-of-the-art Transformer language models are very limited and sample-wise scale worse than relearning all sub-tasks for a new algorithmic composition. We also present a theorem in complexity theory, showing that gradient descent on memorizing feedforward models can be exponentially data inefficient.

Following on recent advances in large language models (LLMs) and subsequent chat models, a new wave of large vision-language models (LVLMs) has emerged. Such models can incorporate images as input in addition to text, and perform tasks such as visual question answering, image captioning, story generation, etc. Here, we examine potential gender and racial biases in such systems, based on the perceived characteristics of the people in the input images. To accomplish this, we present a new dataset PAIRS (PArallel Images for eveRyday Scenarios). The PAIRS dataset contains sets of AI-generated images of people, such that the images are highly similar in terms of background and visual content, but differ along the dimensions of gender (man, woman) and race (Black, white). By querying the LVLMs with such images, we observe significant differences in the responses according to the perceived gender or race of the person depicted.

We introduce SPIRIT-LM, a foundation multimodal language model that freely mixes text and speech. Our model is based on a pretrained text language model that we extend to the speech modality by continuously training it on text and speech units. Speech and text sequences are concatenated as a single set of tokens, and trained with a word-level interleaving method using a small automatically-curated speech-text parallel corpus. SPIRIT-LM comes in two versions: a BASE version that uses speech semantic units and an EXPRESSIVE version that models expressivity using pitch and style units in addition to the semantic units. For both versions, the text is encoded with subword BPE tokens. The resulting model displays both the semantic abilities of text models and the expressive abilities of speech models. Additionally, we demonstrate that SPIRIT-LM is able to learn new tasks in a few-shot fashion across modalities (i.e. ASR, TTS, Speech Classification).

Offline preference optimization allows fine-tuning large models directly from offline data, and has proved effective in recent alignment practices. We propose generalized preference optimization (GPO), a family of offline losses parameterized by a general class of convex functions. GPO enables a unified view over preference optimization, encompassing existing algorithms such as DPO, IPO and SLiC as special cases, while naturally introducing new variants. The GPO framework also sheds light on how offline algorithms enforce regularization, through the design of the convex function that defines the loss. Our analysis and experiments reveal the connections and subtle differences between the offline regularization and the KL divergence regularization intended by the canonical RLHF formulation. In all, our results present new algorithmic toolkits and empirical insights to alignment practitioners.