Scaling laws are useful guides for derisking expensive training runs, as they predict performance of large models using cheaper, small-scale experiments. However, there remain gaps between current scaling studies and how language models are ultimately trained and evaluated. For instance, scaling is usually studied in the compute-optimal training regime (i.e., "Chinchilla optimal" regime). In contrast, models are often over-trained to reduce inference costs. Moreover, scaling laws mostly predict loss on next-token prediction, but models are usually compared on downstream task performance. To address both shortcomings, we create a testbed of 104 models with 0.011B to 6.9B parameters trained with various numbers of tokens on three data distributions. First, we fit scaling laws that extrapolate in both the amount of over-training and the number of model parameters. This enables us to predict the validation loss of a 1.4B parameter, 900B token run (i.e., 32$\times$ over-trained) and a 6.9B parameter, 138B token run (i.e., a compute-optimal run)$\unicode{x2014}$each from experiments that take 300$\times$ less compute. Second, we relate the perplexity of a language model to its downstream task performance by proposing a power law. We use this law to predict top-1 error averaged over downstream tasks for the two aforementioned models, using experiments that take 20$\times$ less compute. Our experiments are available at https://github.com/mlfoundations/scaling.

Trustworthy capability evaluations are crucial for ensuring the safety of AI systems, and are becoming a key component of AI regulation. However, the developers of an AI system, or the AI system itself, may have incentives for evaluations to understate the AI's actual capability. These conflicting interests lead to the problem of sandbagging $\unicode{x2013}$ which we define as "strategic underperformance on an evaluation". In this paper we assess sandbagging capabilities in contemporary language models (LMs). We prompt frontier LMs, like GPT-4 and Claude 3 Opus, to selectively underperform on dangerous capability evaluations, while maintaining performance on general (harmless) capability evaluations. Moreover, we find that models can be fine-tuned, on a synthetic dataset, to hide specific capabilities unless given a password. This behaviour generalizes to high-quality, held-out benchmarks such as WMDP. In addition, we show that both frontier and smaller models can be prompted, or password-locked, to target specific scores on a capability evaluation. Even more, we found that a capable password-locked model (Llama 3 70b) is reasonably able to emulate a less capable model (Llama 2 7b). Overall, our results suggest that capability evaluations are vulnerable to sandbagging. This vulnerability decreases the trustworthiness of evaluations, and thereby undermines important safety decisions regarding the development and deployment of advanced AI systems.

As recent multi-modality large language models (MLLMs) have shown formidable proficiency on various complex tasks, there has been increasing attention on debating whether these models could eventually mirror human intelligence. However, existing benchmarks mainly focus on evaluating solely on task performance, such as the accuracy of identifying the attribute of an object. Combining well-developed cognitive science to understand the intelligence of MLLMs beyond superficial achievements remains largely unexplored. To this end, we introduce the first cognitive-driven multi-lingual and multi-modal benchmark to evaluate the general intelligence ability of MLLMs, dubbed M3GIA. Specifically, we identify five key cognitive factors based on the well-recognized Cattell-Horn-Carrol (CHC) model of intelligence and propose a novel evaluation metric. In addition, since most MLLMs are trained to perform in different languages, a natural question arises: is language a key factor influencing the cognitive ability of MLLMs? As such, we go beyond English to encompass other languages based on their popularity, including Chinese, French, Spanish, Portuguese and Korean, to construct our M3GIA. We make sure all the data relevant to the cultural backgrounds are collected from their native context to avoid English-centric bias. We collected a significant corpus of data from human participants, revealing that the most advanced MLLM reaches the lower boundary of human intelligence in English. Yet, there remains a pronounced disparity in the other five languages assessed. We also reveals an interesting winner takes all phenomenon that are aligned with the discovery in cognitive studies. Our benchmark will be open-sourced, with the aspiration of facilitating the enhancement of cognitive capabilities in MLLMs.

Mainstream approaches to aligning large language models (LLMs) heavily rely on human preference data, particularly when models require periodic updates. The standard process for iterative alignment of LLMs involves collecting new human feedback for each update. However, the data collection process is costly and challenging to scale. To address this issue, we introduce the "TS-Align" framework, which fine-tunes a policy model using pairwise feedback data automatically mined from its outputs. This automatic mining process is efficiently accomplished through the collaboration between a large-scale teacher model and a small-scale student model. The policy fine-tuning process can be iteratively repeated using on-policy generations within our proposed teacher-student collaborative framework. Through extensive experiments, we demonstrate that our final aligned policy outperforms the base policy model with an average win rate of 69.7% across seven conversational or instruction-following datasets. Furthermore, we show that the ranking capability of the teacher is effectively distilled into the student through our pipeline, resulting in a small-scale yet effective reward model for policy model alignment.

The rise of automation has provided an opportunity to achieve higher efficiency in manufacturing processes, yet it often compromises the flexibility required to promptly respond to evolving market needs and meet the demand for customization. Human-robot collaboration attempts to tackle these challenges by combining the strength and precision of machines with human ingenuity and perceptual understanding. In this paper, we conceptualize and propose an implementation framework for an autonomous, machine learning-based manipulator that incorporates human-in-the-loop principles and leverages Extended Reality (XR) to facilitate intuitive communication and programming between humans and robots. Furthermore, the conceptual framework foresees human involvement directly in the robot learning process, resulting in higher adaptability and task generalization. The paper highlights key technologies enabling the proposed framework, emphasizing the importance of developing the digital ecosystem as a whole. Additionally, we review the existent implementation approaches of XR in human-robot collaboration, showcasing diverse perspectives and methodologies. The challenges and future outlooks are discussed, delving into the major obstacles and potential research avenues of XR for more natural human-robot interaction and integration in the industrial landscape.

Most public instruction finetuning datasets are relatively small compared to the closed source datasets used to train industry models. To study questions about finetuning at scale, such as curricula and learning rate cooldown schedules, there is a need for industrial-scale datasets. However, this scale necessitates a data generation process that is almost entirely automated. In this work, we study methods for generating large instruction datasets from a single prompt. With little human oversight, we get LLMs to write diverse sets of instruction examples ranging from simple completion tasks to complex multi-turn dialogs across a variety of subject areas. When finetuning a Llama-3 8B base model, our dataset meets or exceeds both WizardLM and Ultrachat on both knowledge-intensive leaderboard tasks as well as conversational evaluations. We release our dataset, the "generator" prompts that created it, and our finetuned model checkpoints.

The advancements in disentangled representation learning significantly enhance the accuracy of counterfactual predictions by granting precise control over instrumental variables, confounders, and adjustable variables. An appealing method for achieving the independent separation of these factors is mutual information minimization, a task that presents challenges in numerous machine learning scenarios, especially within high-dimensional spaces. To circumvent this challenge, we propose the Self-Distilled Disentanglement framework, referred to as $SD^2$. Grounded in information theory, it ensures theoretically sound independent disentangled representations without intricate mutual information estimator designs for high-dimensional representations. Our comprehensive experiments, conducted on both synthetic and real-world datasets, confirms the effectiveness of our approach in facilitating counterfactual inference in the presence of both observed and unobserved confounders.

Last year, LinkedIn began experimenting with AI-powered tools for job seekers on its platform. Now the company has added a bunch of new capabilities for its premium subscribers who are #OpentoWork, including personalized resumé, AI-assisted cover letters and more conversational job searches. The changes are meant to speed up some of the most tedious aspects of looking for a new role. For example, the revamped job search feature now allows you to look for roles with queries like "find me a marketing job that's fully remote and pays at least 100,000 a year," or "find business development roles in biotech." Those are all relatively simple descriptions but anyone who has searched for jobs on LinkedIn (without the help of AI) knows that it can often be a struggle to narrow down job listings with keywords.

Many burned-out workers have likely dreamed of hiring a career coach or résumé writer. Now, LinkedIn is introducing chats with generative AI career experts based on real people. Other new AI tools within the platform will help people write résumés and cover letters or evaluate their qualifications for jobs posted. LinkedIn has ramped up its generative AI tools in the past year and is moving to incorporate the tech into even more of its offerings. On Thursday, the career site announced new features like a pilot for AI-powered expert advice, an interactive chat to break down information in LinkedIn courses, and more AI features that can be used to search for and apply for jobs for its premium users in English.

Drug design is the process of identifying molecules responsible for medicinal activity. Suppose we want to automate drug design with machine learning. To do so, we would like to automatically learn programs which explain why a molecule is active or inactive. For instance, as illustrated in the figure above, a program might determine that a molecule is active if it contains a hydrogen atom with a charge greater than 0.2C, and located within 0.1 angstroms of a carbon atom. Discovering this program involves identifying the numerical values 0.2 and 0.1.