Large language models (LLM) have emerged as a powerful tool for AI, with the key ability of in-context learning (ICL), where they can perform well on unseen tasks based on a brief series of task examples without necessitating any adjustments to the model parameters. One recent interesting mysterious observation is that models of different scales may have different ICL behaviors: larger models tend to be more sensitive to noise in the test context. This work studies this observation theoretically aiming to improve the understanding of LLM and ICL. We analyze two stylized settings: (1) linear regression with one-layer single-head linear transformers and (2) parity classification with two-layer multiple attention heads transformers (non-linear data and non-linear model). In both settings, we give closed-form optimal solutions and find that smaller models emphasize important hidden features while larger ones cover more hidden features; thus, smaller models are more robust to noise while larger ones are more easily distracted, leading to different ICL behaviors. This sheds light on where transformers pay attention to and how that affects ICL. Preliminary experimental results on large base and chat models provide positive support for our analysis.

Continual learning methods are known to suffer from catastrophic forgetting, a phenomenon that is particularly hard to counter for methods that do not store exemplars of previous tasks. Therefore, to reduce potential drift in the feature extractor, existing exemplar-free methods are typically evaluated in settings where the first task is significantly larger than subsequent tasks. Their performance drops drastically in more challenging settings starting with a smaller first task. To address this problem of feature drift estimation for exemplar-free methods, we propose to adversarially perturb the current samples such that their embeddings are close to the old class prototypes in the old model embedding space. We then estimate the drift in the embedding space from the old to the new model using the perturbed images and compensate the prototypes accordingly. We exploit the fact that adversarial samples are transferable from the old to the new feature space in a continual learning setting. The generation of these images is simple and computationally cheap. We demonstrate in our experiments that the proposed approach better tracks the movement of prototypes in embedding space and outperforms existing methods on several standard continual learning benchmarks as well as on fine-grained datasets. Code is available at https://github.com/dipamgoswami/ADC.

A hot topic of Artificial Intelligence right now is image generation from prompts. DALL-E 2 is one of the biggest names in this domain, as it allows people to create images from simple text inputs, to even more complicated ones. The company that made this possible, OpenAI, has assured everyone that visited their website that "Our mission is to ensure that artificial general intelligence benefits all humanity". A noble idea in our opinion, that also stood as the motive behind us choosing this subject. This paper analyzes the ethical implications of an AI image generative system, with an emphasis on how society is responding to it, how it probably will and how it should if all the right measures are taken.

Continual learning with deep neural networks presents challenges distinct from both the fixed-dataset and convex continual learning regimes. One such challenge is plasticity loss, wherein a neural network trained in an online fashion displays a degraded ability to fit new tasks. This problem has been extensively studied in both supervised learning and off-policy reinforcement learning (RL), where a number of remedies have been proposed. Still, plasticity loss has received less attention in the on-policy deep RL setting. Here we perform an extensive set of experiments examining plasticity loss and a variety of mitigation methods in on-policy deep RL. We demonstrate that plasticity loss is pervasive under domain shift in this regime, and that a number of methods developed to resolve it in other settings fail, sometimes even resulting in performance that is worse than performing no intervention at all. In contrast, we find that a class of ``regenerative'' methods are able to consistently mitigate plasticity loss in a variety of contexts, including in gridworld tasks and more challenging environments like Montezuma's Revenge and ProcGen.

The dominant framework for alignment of large language models (LLM), whether through reinforcement learning from human feedback or direct preference optimisation, is to learn from preference data. This involves building datasets where each element is a quadruplet composed of a prompt, two independent responses (completions of the prompt) and a human preference between the two independent responses, yielding a preferred and a dis-preferred response. Such data is typically scarce and expensive to collect. On the other hand, \emph{single-trajectory} datasets where each element is a triplet composed of a prompt, a response and a human feedback is naturally more abundant. The canonical element of such datasets is for instance an LLM's response to a user's prompt followed by a user's feedback such as a thumbs-up/down. Consequently, in this work, we propose DRO, or \emph{Direct Reward Optimisation}, as a framework and associated algorithms that do not require pairwise preferences. DRO uses a simple mean-squared objective that can be implemented in various ways. We validate our findings empirically, using T5 encoder-decoder language models, and show DRO's performance over selected baselines such as Kahneman-Tversky Optimization (KTO). Thus, we confirm that DRO is a simple and empirically compelling method for single-trajectory policy optimisation.

Applications of Generative AI (Gen AI) are expected to revolutionize a number of different areas, ranging from science & medicine to education. The potential for these seismic changes has triggered a lively debate about the potential risks of the technology, and resulted in calls for tighter regulation, in particular from some of the major tech companies who are leading in AI development. This regulation is likely to put at risk the budding field of open-source generative AI. Using a three-stage framework for Gen AI development (near, mid and long-term), we analyze the risks and opportunities of open-source generative AI models with similar capabilities to the ones currently available (near to mid-term) and with greater capabilities (long-term). We argue that, overall, the benefits of open-source Gen AI outweigh its risks. As such, we encourage the open sourcing of models, training and evaluation data, and provide a set of recommendations and best practices for managing risks associated with open-source generative AI.

To determine the safety of large language models (LLMs), AI developers must be able to assess their dangerous capabilities. But simple prompting strategies often fail to elicit an LLM's full capabilities. One way to elicit capabilities more robustly is to fine-tune the LLM to complete the task. In this paper, we investigate the conditions under which fine-tuning-based elicitation suffices to elicit capabilities. To do this, we introduce password-locked models, LLMs fine-tuned such that some of their capabilities are deliberately hidden. Specifically, these LLMs are trained to exhibit these capabilities only when a password is present in the prompt, and to imitate a much weaker LLM otherwise. Password-locked models enable a novel method of evaluating capabilities elicitation methods, by testing whether these password-locked capabilities can be elicited without using the password. We find that a few high-quality demonstrations are often sufficient to fully elicit password-locked capabilities. More surprisingly, fine-tuning can elicit other capabilities that have been locked using the same password, or even different passwords. Furthermore, when only evaluations, and not demonstrations, are available, approaches like reinforcement learning are still often able to elicit capabilities. Overall, our findings suggest that fine-tuning is an effective method of eliciting hidden capabilities of current models, but may be unreliable when high-quality demonstrations are not available, e.g. as may be the case when models' (hidden) capabilities exceed those of human demonstrators.

Effective evaluation of language models remains an open challenge in NLP. Researchers and engineers face methodological issues such as the sensitivity of models to evaluation setup, difficulty of proper comparisons across methods, and the lack of reproducibility and transparency. In this paper we draw on three years of experience in evaluating large language models to provide guidance and lessons for researchers. First, we provide an overview of common challenges faced in language model evaluation. Second, we delineate best practices for addressing or lessening the impact of these challenges on research. Third, we present the Language Model Evaluation Harness (lm-eval): an open source library for independent, reproducible, and extensible evaluation of language models that seeks to address these issues. We describe the features of the library as well as case studies in which the library has been used to alleviate these methodological concerns.

For aligning large language models (LLMs), prior work has leveraged reinforcement learning via human feedback (RLHF) or variations of direct preference optimization (DPO). While DPO offers a simpler framework based on maximum likelihood estimation, it compromises on the ability to tune language models to easily maximize non-differentiable and non-binary objectives according to the LLM designer's preferences (e.g., using simpler language or minimizing specific kinds of harmful content). These may neither align with user preferences nor even be able to be captured tractably by binary preference data. To leverage the simplicity and performance of DPO with the generalizability of RL, we propose a hybrid approach between DPO and RLHF. With a simple augmentation to the implicit reward decomposition of DPO, we allow for tuning LLMs to maximize a set of arbitrary auxiliary rewards using offline RL. The proposed method, Hybrid Preference Optimization (HPO), shows the ability to effectively generalize to both user preferences and auxiliary designer objectives, while preserving alignment performance across a range of challenging benchmarks and model sizes.

This paper explores the potential of a hybrid modeling approach that combines machine learning (ML) with conventional physics-based modeling for weather prediction beyond the medium range. It extends the work of Arcomano et al. (2022), which tested the approach for short- and medium-range weather prediction, and the work of Arcomano et al. (2023), which investigated its potential for climate modeling. The hybrid model used for the forecast experiments of the paper is based on the low-resolution, simplified parameterization atmospheric general circulation model (AGCM) SPEEDY. In addition to the hybridized prognostic variables of SPEEDY, the current version of the model has three purely ML-based prognostic variables. One of these is 6~h cumulative precipitation, another is the sea surface temperature, while the third is the heat content of the top 300 m deep layer of the ocean. The model has skill in predicting the El Ni\~no cycle and its global teleconnections with precipitation for 3-7 months depending on the season. The model captures equatorial variability of the precipitation associated with Kelvin and Rossby waves and MJO. Predictions of the precipitation in the equatorial region have skill for 15 days in the East Pacific and 11.5 days in the West Pacific. Though the model has low spatial resolution, for these tasks it has prediction skill comparable to what has been published for high-resolution, purely physics-based, conventional operational forecast models.