Reinforcement learning from human feedback usually models preferences using a reward model that does not distinguish between people. We argue that this is unlikely to be a good design choice in contexts with high potential for disagreement, like in the training of large language models. We propose a method to specialise a reward model to a person or group of people. Our approach builds on the observation that individual preferences can be captured as a linear combination of a set of general reward features. We show how to learn such features and subsequently use them to quickly adapt the reward model to a specific individual, even if their preferences are not reflected in the training data. We present experiments with large language models comparing the proposed architecture with a non-adaptive reward model and also adaptive counterparts, including models that do in-context personalisation. Depending on how much disagreement there is in the training data, our model either significantly outperforms the baselines or matches their performance with a simpler architecture and more stable training.

The use of deep neural networks in reinforcement learning (RL) often suffers from performance degradation as model size increases. While soft mixtures of experts (SoftMoEs) have recently shown promise in mitigating this issue for online RL, the reasons behind their effectiveness remain largely unknown. In this work we provide an in-depth analysis identifying the key factors driving this performance gain. We discover the surprising result that tokenizing the encoder output, rather than the use of multiple experts, is what is behind the efficacy of SoftMoEs. Indeed, we demonstrate that even with an appropriately scaled single expert, we are able to maintain the performance gains, largely thanks to tokenization.

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.

Learning from human feedback (LHF) -- and in particular learning from pairwise preferences -- has recently become a crucial ingredient in training large language models (LLMs), and has been the subject of much research. Most recent works frame it as a reinforcement learning problem, where a reward function is learned from pairwise preference data and the LLM is treated as a policy which is adapted to maximize the rewards, often under additional regularization constraints. We propose an alternative interpretation which centers on the generative process for pairwise preferences and treats LHF as a density estimation problem. We provide theoretical and empirical results showing that for a family of generative processes defined via preference behavior distribution equations, training a reward function on pairwise preferences effectively models an annotator's implicit preference distribution. Finally, we discuss and present findings on "annotator misspecification" -- failure cases where wrong modeling assumptions are made about annotator behavior, resulting in poorly-adapted models -- suggesting that approaches that learn from pairwise human preferences could have trouble learning from a population of annotators with diverse viewpoints.

Both methods, Unlearning via Activations and Unlearning via Examples, successfully demonstrated unlearning of the forget class while having a negligible effect on the models' performance on the retain set. Importantly, this is achieved without any form of training, retraining, or fine-tuning as is usually required by other methods. The retain set test accuracy remains more or less constant for all three datasets except for a few minor fluctuations. This is a result of the fact that due to localized and context-dependent sparse updates during the initial training of the model, discrete key-representations corresponding to different classes in the dataset are well separated from each other, an important prerequisite discussed in (Träuble et al., 2023). Hence, all the information about a class can be unlearned by forgetting only a subset of the forget class training data in the case of Unlearning via Examples, making it very data-efficient.

With the success of large language models (LLMs) of code and their use as code assistants (e.g. Codex used in GitHub Copilot), techniques for introducing domain-specific knowledge in the prompt design process become important. In this work, we propose a framework called Repo-Level Prompt Generator that learns to generate example-specific prompts using prompt proposals. The prompt proposals take context from the entire repository, thereby incorporating both the structure of the repository and the context from other relevant files (e.g. imports, parent class files). Our technique doesn't require any access to the weights of the LLM, making it applicable in cases where we only have black-box access to the LLM. We conduct experiments on the task of single-line code-autocompletion using code repositories taken from Google Code archives. We demonstrate that an oracle constructed from our prompt proposals gives a remarkably high relative improvement of 36% over Codex, showing the quality of these proposals. Further, we show that when we train a model to predict a prompt proposal, we can achieve significant performance gains over Codex and other baselines. We release our code, data, and trained checkpoints at: \url{https://github.com/shrivastavadisha/repo_level_prompt_generation}.

Large language models (LLMs) have shown increasing in-context learning capabilities through scaling up model and data size. Despite this progress, LLMs are still unable to solve algorithmic reasoning problems. While providing a rationale with the final answer has led to further improvements in multi-step reasoning problems, Anil et al. 2022 showed that even simple algorithmic reasoning tasks such as parity are far from solved. In this work, we identify and study four key stages for successfully teaching algorithmic reasoning to LLMs: (1) formulating algorithms as skills, (2) teaching multiple skills simultaneously (skill accumulation), (3) teaching how to combine skills (skill composition) and (4) teaching how to use skills as tools. We show that it is possible to teach algorithmic reasoning to LLMs via in-context learning, which we refer to as algorithmic prompting. We evaluate our approach on a variety of arithmetic and quantitative reasoning tasks, and demonstrate significant boosts in performance over existing prompting techniques. In particular, for long parity, addition, multiplication and subtraction, we achieve an error reduction of approximately 10x, 9x, 5x and 2x respectively compared to the best available baselines.

Forgetting is often seen as an unwanted characteristic in both human and machine learning. However, we propose that forgetting can in fact be favorable to learning. We introduce forget-and-relearn as a powerful paradigm for shaping the learning trajectories of artificial neural networks. In this process, the forgetting step selectively removes undesirable information from the model, and the relearning step reinforces features that are consistently useful under different conditions. The forget-and-relearn framework unifies many existing iterative training algorithms in the image classification and language emergence literature, and allows us to understand the success of these algorithms in terms of the disproportionate forgetting of undesirable information. We leverage this understanding to improve upon existing algorithms by designing more targeted forgetting operations. Insights from our analysis provide a coherent view on the dynamics of iterative training in neural networks and offer a clear path towards performance improvements. Forgetting is an inescapable component of human memory. It occurs naturally as neural synapses get removed or altered over time (Wang et al., 2020), and is often thought to be an undesirable characteristic of the human mind. A well-known example is the "spacing effect", which refers to the observation that long-term recall is enhanced by spacing, rather than massing, repeated study sessions. Bjork & Allen (1970) demonstrated that the key to the spacing effect is the decreased accessibility of information in-between sessions. In this work, we study a general learning paradigm that we refer to as forget-and-relearn, and show that forgetting can also benefit learning in artificial neural networks. To generalize to unseen data, we want our models to capture generalizable concepts rather than purely statistical regularities, but these desirable solutions are a small subset of the solution space and often more difficult to learn naturally (Geirhos et al., 2020). Recently, a number of training algorithms have been proposed to improve generalization by iteratively refining the learned solution. Knowledge evolution (Taha et al., 2021) improves generalization by iteratively reinitializing one part of the network while continuously training the other. Iterative magnitude pruning (Frankle & Carbin, 2019; Frankle et al., 2019) removes weights through an iterative pruning-retraining process, and outperforms unpruned models in certain settings. Hoang et al. (2018) iteratively utilize synthetic machine translation corpus through back-translations of monolingual data.

The goal of program synthesis from examples is to find a computer program that is consistent with a given set of input-output examples. Most learning-based approaches try to find a program that satisfies all examples at once. Our work, by contrast, considers an approach that breaks the problem into two stages: (a) find programs that satisfy only one example, and (b) leverage these per-example solutions to yield a program that satisfies all examples. We introduce the Cross Aggregator neural network module based on a multi-head attention mechanism that learns to combine the cues present in these per-example solutions to synthesize a global solution. Evaluation across programs of different lengths and under two different experimental settings reveal that when given the same time budget, our technique significantly improves the success rate over PCCoder [32] and other ablation baselines.

Graph-based neural network models are producing strong results in a number of domains, in part because graphs provide flexibility to encode domain knowledge in the form of relational structure (edges) between nodes in the graph. In practice, edges are used both to represent intrinsic structure (e.g., abstract syntax trees of programs) and more abstract relations that aid reasoning for a downstream task (e.g., results of relevant program analyses). In this work, we study the problem of learning to derive abstract relations from the intrinsic graph structure. Motivated by their power in program analyses, we consider relations defined by paths on the base graph accepted by a finite-state automaton. We show how to learn these relations end-to-end by relaxing the problem into learning finite-state automata policies on a graph-based POMDP and then training these policies using implicit differentiation. The result is a differentiable Graph Finite-State Automaton (GFSA) layer that adds a new edge type (expressed as a weighted adjacency matrix) to a base graph. We demonstrate that this layer can find shortcuts in grid-world graphs and reproduce simple static analyses on Python programs. Additionally, we combine the GFSA layer with a larger graph-based model trained end-to-end on the variable misuse program understanding task, and find that using the GFSA layer leads to better performance than using hand-engineered semantic edges or other baseline methods for adding learned edge types.