We propose MusicRL, the first music generation system finetuned from human feedback. Appreciation of text-to-music models is particularly subjective since the concept of musicality as well as the specific intention behind a caption are user-dependent (e.g. a caption such as "upbeat work-out music" can map to a retro guitar solo or a techno pop beat). Not only this makes supervised training of such models challenging, but it also calls for integrating continuous human feedback in their post-deployment finetuning. MusicRL is a pretrained autoregressive MusicLM (Agostinelli et al., 2023) model of discrete audio tokens finetuned with reinforcement learning to maximise sequence-level rewards. We design reward functions related specifically to text-adherence and audio quality with the help from selected raters, and use those to finetune MusicLM into MusicRL-R. We deploy MusicLM to users and collect a substantial dataset comprising 300,000 pairwise preferences. Using Reinforcement Learning from Human Feedback (RLHF), we train MusicRL-U, the first text-to-music model that incorporates human feedback at scale. Human evaluations show that both MusicRL-R and MusicRL-U are preferred to the baseline. Ultimately, MusicRL-RU combines the two approaches and results in the best model according to human raters. Ablation studies shed light on the musical attributes influencing human preferences, indicating that text adherence and quality only account for a part of it. This underscores the prevalence of subjectivity in musical appreciation and calls for further involvement of human listeners in the finetuning of music generation models.

Large-scale generative models are shown to be useful for sampling meaningful candidate solutions, yet they often overlook task constraints and user preferences. Their full power is better harnessed when the models are coupled with external verifiers and the final solutions are derived iteratively or progressively according to the verification feedback. In the context of embodied AI, verification often solely involves assessing whether goal conditions specified in the instructions have been met. Nonetheless, for these agents to be seamlessly integrated into daily life, it is crucial to account for a broader range of constraints and preferences beyond bare task success (e.g., a robot should grasp bread with care to avoid significant deformations). However, given the unbounded scope of robot tasks, it is infeasible to construct scripted verifiers akin to those used for explicit-knowledge tasks like the game of Go and theorem proving. This begs the question: when no sound verifier is available, can we use large vision and language models (VLMs), which are approximately omniscient, as scalable Behavior Critics to catch undesirable robot behaviors in videos? To answer this, we first construct a benchmark that contains diverse cases of goal-reaching yet undesirable robot policies. Then, we comprehensively evaluate VLM critics to gain a deeper understanding of their strengths and failure modes. Based on the evaluation, we provide guidelines on how to effectively utilize VLM critiques and showcase a practical way to integrate the feedback into an iterative process of policy refinement. The dataset and codebase are released at: https://guansuns.github.io/pages/vlm-critic.

This paper introduces a system designed to generate explanations for the actions performed by an autonomous robot in Human-Robot Interaction (HRI). Explainability in robotics, encapsulated within the concept of an eXplainable Autonomous Robot (XAR), is a growing research area. The work described in this paper aims to take advantage of the capabilities of Large Language Models (LLMs) in performing natural language processing tasks. This study focuses on the possibility of generating explanations using such models in combination with a Retrieval Augmented Generation (RAG) method to interpret data gathered from the logs of autonomous systems. In addition, this work also presents a formalization of the proposed explanation system. It has been evaluated through a navigation test from the European Robotics League (ERL), a Europe-wide social robotics competition. Regarding the obtained results, a validation questionnaire has been conducted to measure the quality of the explanations from the perspective of technical users. The results obtained during the experiment highlight the potential utility of LLMs in achieving explanatory capabilities in robots.

Scaling laws provide important insights that can guide the design of large language models (LLMs). Existing work has primarily focused on studying scaling laws for pretraining (upstream) loss. However, in transfer learning settings, in which LLMs are pretrained on an unsupervised dataset and then finetuned on a downstream task, we often also care about the downstream performance. In this work, we study the scaling behavior in a transfer learning setting, where LLMs are finetuned for machine translation tasks. Specifically, we investigate how the choice of the pretraining data and its size affect downstream performance (translation quality) as judged by two metrics: downstream cross-entropy and BLEU score. Our experiments indicate that the size of the finetuning dataset and the distribution alignment between the pretraining and downstream data significantly influence the scaling behavior. With sufficient alignment, both downstream cross-entropy and BLEU score improve monotonically with more pretraining data. In such cases, we show that it is possible to predict the downstream BLEU score with good accuracy using a log-law. However, there are also cases where moderate misalignment causes the BLEU score to fluctuate or get worse with more pretraining, whereas downstream cross-entropy monotonically improves. By analyzing these observations, we provide new practical insights for choosing appropriate pretraining data.

In recent years, attention-based transformers have achieved tremendous success across a variety of disciplines including natural languages. A key ingredient behind their success is the generative pretraining procedure, during which these models are trained on a large text corpus in an auto-regressive manner. To shed light on this phenomenon, we propose a new framework that allows both theory and systematic experiments to study the sequential modeling capabilities of transformers through the lens of Markov chains. Inspired by the Markovianity of natural languages, we model the data as a Markovian source and utilize this framework to systematically study the interplay between the data-distributional properties, the transformer architecture, the learnt distribution, and the final model performance. In particular, we theoretically characterize the loss landscape of single-layer transformers and show the existence of global minima and bad local minima contingent upon the specific data characteristics and the transformer architecture. Backed by experiments, we demonstrate that our theoretical findings are in congruence with the empirical results. We further investigate these findings in the broader context of higher order Markov chains and deeper architectures, and outline open problems in this arena. Code is available at \url{https://github.com/Bond1995/Markov}.

Controllable text generation is a growing field within natural language generation (NLG) that focuses on producing text that meets specific constraints in real-world applications. Previous approaches, such as plug-and-play controllers (PPCs), aimed to steer the properties of generated text in a flexible manner. However, these methods often compromised the integrity of the language model's decoding process, resulting in less smooth text generation. Alternatively, other techniques utilized multiple attribute prompts to align the generated text with desired attributes, but this approach required prompt design for each attribute and was dependent on the size of the language model. This paper introduces a novel method for flexible attribute control in text generation using pre-trained language models (PLMs). The proposed approach aims to enhance the fluency of generated text by guiding the generation process with PPCs. The key idea is to dynamically adjust the distribution of generated text by modifying prompts, effectively constraining the output space of the language model and influencing the desired attribute. To enable smooth cooperation between the PLM and the PPC, our work innovatively proposes a new model fine-tuning method: Reinforcement Learning with Dynamic Adjust Feedback (RLDAF).This fine-tuning process adapts a small subset of the language model's parameters based on the generating actions taken during the PPC control process. The resulting harmonious collaboration between the PLM and PPC leads to improved smoothness in text generation during inference. Extensive experiments were conducted on the SST2 dataset, and the proposed method outperformed previous approaches in various evaluation metrics, including text fluency and attribute consistency.

CodeCompose is an AI-assisted code authoring tool powered by large language models (LLMs) that provides inline suggestions to 10's of thousands of developers at Meta. In this paper, we present how we scaled the product from displaying single-line suggestions to multi-line suggestions. This evolution required us to overcome several unique challenges in improving the usability of these suggestions for developers. First, we discuss how multi-line suggestions can have a 'jarring' effect, as the LLM's suggestions constantly move around the developer's existing code, which would otherwise result in decreased productivity and satisfaction. Second, multi-line suggestions take significantly longer to generate; hence we present several innovative investments we made to reduce the perceived latency for users. These model-hosting optimizations sped up multi-line suggestion latency by 2.5x. Finally, we conduct experiments on 10's of thousands of engineers to understand how multi-line suggestions impact the user experience and contrast this with single-line suggestions. Our experiments reveal that (i) multi-line suggestions account for 42% of total characters accepted (despite only accounting for 16% for displayed suggestions) (ii) multi-line suggestions almost doubled the percentage of keystrokes saved for users from 9% to 17%. Multi-line CodeCompose has been rolled out to all engineers at Meta, and less than 1% of engineers have opted out of multi-line suggestions.

ChatGPT is notorious for its intransparent behavior. This paper tries to shed light on this, providing an in-depth analysis of the dark personality traits and conspiracy beliefs of GPT-3.5 and GPT-4. Different psychological tests and questionnaires were employed, including the Dark Factor Test, the Mach-IV Scale, the Generic Conspiracy Belief Scale, and the Conspiracy Mentality Scale. The responses were analyzed computing average scores, standard deviations, and significance tests to investigate differences between GPT-3.5 and GPT-4. For traits that have shown to be interdependent in human studies, correlations were considered. Additionally, system roles corresponding to groups that have shown distinct answering behavior in the corresponding questionnaires were applied to examine the models' ability to reflect characteristics associated with these roles in their responses. Dark personality traits and conspiracy beliefs were not particularly pronounced in either model with little differences between GPT-3.5 and GPT-4. However, GPT-4 showed a pronounced tendency to believe in information withholding. This is particularly intriguing given that GPT-4 is trained on a significantly larger dataset than GPT-3.5. Apparently, in this case an increased data exposure correlates with a greater belief in the control of information. An assignment of extreme political affiliations increased the belief in conspiracy theories. Test sequencing affected the models' responses and the observed correlations, indicating a form of contextual memory.

Large language models (LLMs) can pass explicit bias tests but still harbor implicit biases, similar to humans who endorse egalitarian beliefs yet exhibit subtle biases. Measuring such implicit biases can be a challenge: as LLMs become increasingly proprietary, it may not be possible to access their embeddings and apply existing bias measures; furthermore, implicit biases are primarily a concern if they affect the actual decisions that these systems make. We address both of these challenges by introducing two measures of bias inspired by psychology: LLM Implicit Association Test (IAT) Bias, which is a prompt-based method for revealing implicit bias; and LLM Decision Bias for detecting subtle discrimination in decision-making tasks. Using these measures, we found pervasive human-like stereotype biases in 6 LLMs across 4 social domains (race, gender, religion, health) and 21 categories (weapons, guilt, science, career among others). Our prompt-based measure of implicit bias correlates with embedding-based methods but better predicts downstream behaviors measured by LLM Decision Bias. This measure is based on asking the LLM to decide between individuals, motivated by psychological results indicating that relative not absolute evaluations are more related to implicit biases. Using prompt-based measures informed by psychology allows us to effectively expose nuanced biases and subtle discrimination in proprietary LLMs that do not show explicit bias on standard benchmarks.

The dominance of social media has added to the channels of bullying for perpetrators. Unfortunately, cyberbullying (CB) is the most prevalent phenomenon in todays cyber world, and is a severe threat to the mental and physical health of citizens. This opens the need to develop a robust system to prevent bullying content from online forums, blogs, and social media platforms to manage the impact in our society. Several machine learning (ML) algorithms have been proposed for this purpose. However, their performances are not consistent due to high class imbalance and generalisation issues. In recent years, large language models (LLMs) like BERT and RoBERTa have achieved state-of-the-art (SOTA) results in several natural language processing (NLP) tasks. Unfortunately, the LLMs have not been applied extensively for CB detection. In our paper, we explored the use of these models for cyberbullying (CB) detection. We have prepared a new dataset (D2) from existing studies (Formspring and Twitter). Our experimental results for dataset D1 and D2 showed that RoBERTa outperformed other models.