Large language models (LLMs), despite their extensive pretraining on diverse datasets, require effective alignment to human preferences for practical and reliable deployment. Conventional alignment methods typically employ off-policy learning and depend on human-annotated datasets, which limits their broad applicability and introduces distribution shift issues during training. To address these challenges, we propose Self-Generated Preference Optimization based on Self-Improver (SGPO), an innovative alignment framework that leverages an on-policy self-improving mechanism. Specifically, the improver refines responses from a policy model to self-generate preference data for direct preference optimization (DPO) of the policy model. Here, the improver and policy are unified into a single model, and in order to generate higher-quality preference data, this self-improver learns to make incremental yet discernible improvements to the current responses by referencing supervised fine-tuning outputs. Experimental results on AlpacaEval 2.0 and Arena-Hard show that the proposed SGPO significantly improves performance over DPO and baseline self-improving methods without using external preference data.

Neural solvers have shown significant potential in solving the Traveling Salesman Problem (TSP), yet current approaches face significant challenges. Supervised learning (SL)-based solvers require large amounts of high-quality labeled data, while reinforcement learning (RL)-based solvers, though less dependent on such data, often suffer from inefficiencies. To address these limitations, we propose LocalEscaper, a novel weakly-supervised learning framework for large-scale TSP. LocalEscaper effectively combines the advantages of both SL and RL, enabling effective training on datasets with low-quality labels. To further enhance solution quality, we introduce a regional reconstruction strategy, which mitigates the problem of local optima, a common issue in existing local reconstruction methods. Additionally, we propose a linear-complexity attention mechanism that reduces computational overhead, enabling the efficient solution of large-scale TSPs without sacrificing performance. Experimental results on both synthetic and real-world datasets demonstrate that LocalEscaper outperforms existing neural solvers, achieving state-of-the-art results. Notably, it sets a new benchmark for scalability and efficiency, solving TSP instances with up to 50,000 cities.

Large language models (LLMs) have been used to generate formal proofs of mathematical theorems in proofs assistants such as Lean. However, we often want to optimize a formal proof with respect to various criteria, depending on its downstream use. For example, we may want a proof to adhere to a certain style, or to be readable, concise, or modularly structured. Having suitably optimized proofs is also important for learning tasks, especially since human-written proofs may not optimal for that purpose. To this end, we study a new problem of automated proof optimization: rewriting a proof so that it is correct and optimizes for an arbitrary criterion, such as length or readability. As a first method for automated proof optimization, we present ImProver, a large-language-model agent that rewrites proofs to optimize arbitrary user-defined metrics in Lean. We find that naively applying LLMs to proof optimization falls short, and we incorporate various improvements into ImProver, such as the use of symbolic Lean context in a novel Chain-of-States technique, as well as error-correction and retrieval. We test ImProver on rewriting real-world undergraduate, competition, and research-level mathematics theorems, finding that ImProver is capable of rewriting proofs so that they are substantially shorter, more modular, and more readable.

Site-specific weather forecasts are essential to accurate prediction of power demand and are consequently of great interest to energy operators. However, weather forecasts from current numerical weather prediction (NWP) models lack the fine-scale detail to capture all important characteristics of localised real-world sites. Instead they provide weather information representing a rectangular gridbox (usually kilometres in size). Even after post-processing and bias correction, area-averaged information is usually not optimal for specific sites. Prior work on site optimised forecasts has focused on linear methods, weighted consensus averaging, time-series methods, and others. Recent developments in machine learning (ML) have prompted increasing interest in applying ML as a novel approach towards this problem. In this study, we investigate the feasibility of optimising forecasts at sites by adopting the popular machine learning model gradient boosting decision tree, supported by the Python version of the XGBoost package. Regression trees have been trained with historical NWP and site observations as training data, aimed at predicting temperature and dew point at multiple site locations across Australia. We developed a working ML framework, named 'Multi-SiteBoost' and initial testing results show a significant improvement compared with gridded values from bias-corrected NWP models. The improvement from XGBoost is found to be comparable with non-ML methods reported in literature. With the insights provided by SHapley Additive exPlanations (SHAP), this study also tests various approaches to understand the ML predictions and increase the reliability of the forecasts generated by ML.

Several recent advances in AI systems (e.g., Tree-of-Thoughts and Program-Aided Language Models) solve problems by providing a "scaffolding" program that structures multiple calls to language models to generate better outputs. A scaffolding program is written in a programming language such as Python. In this work, we use a language-model-infused scaffolding program to improve itself. We start with a seed "improver" that improves an input program according to a given utility function by querying a language model several times and returning the best solution. We then run this seed improver to improve itself. Across a small set of downstream tasks, the resulting improved improver generates programs with significantly better performance than its seed improver. A variety of self-improvement strategies are proposed by the language model, including beam search, genetic algorithms, and simulated annealing. Since the language models themselves are not altered, this is not full recursive self-improvement. Nonetheless, it demonstrates that a modern language model, GPT-4 in our proof-of-concept experiments, is capable of writing code that can call itself to improve itself. We consider concerns around the development of self-improving technologies and evaluate the frequency with which the generated code bypasses a sandbox. A language model can be queried to optimize virtually any objective describable in natural language. However, a program that makes multiple, structured calls to a language model can often produce outputs with higher objective values (Yao et al., 2022; 2023; Zelikman et al., 2023; Chen et al., 2022). We refer to these as "scaffolding" programs, typically written (by humans) in a programming language such as Python. Our key observation is that, for any distribution over optimization problems and any fixed language model, the design of a scaffolding program is itself an optimization problem. In this work, we introduce the Self-Taught Optimizer (STOP), a method in which code that applies a language model to improve arbitrary solutions is applied recursively to improve itself. Our approach begins with an initial seed'improver' scaffolding program that uses the language model to improve a solution to some downstream task.

Machine Learning (ML) is more than just training models, the whole workflow must be considered. Once deployed, a ML model needs to be watched and constantly supervised and debugged to guarantee its validity and robustness in unexpected situations. Debugging in ML aims to identify (and address) the model weaknesses in not trivial contexts. Several techniques have been proposed to identify different types of model weaknesses, such as bias in classification, model decay, adversarial attacks, etc., yet there is not a generic framework that allows them to work in a collaborative, modular, portable, iterative way and, more importantly, flexible enough to allow both human- and machine-driven techniques. In this paper, we propose a novel containerized directed graph framework to support and accelerate end-to-end ML workflow management, supervision, and debugging. The framework allows defining and deploying ML workflows in containers, tracking their metadata, checking their behavior in production, and improving the models by using both learned and human-provided knowledge. We demonstrate these capabilities by integrating in the framework two hybrid systems to detect data drift distribution which identify the samples that are far from the latent space of the original distribution, ask for human intervention, and whether retrain the model or wrap it with a filter to remove the noise of corrupted data at inference time. We test these systems on MNIST-C, CIFAR-10-C, and FashionMNIST-C datasets, obtaining promising accuracy results with the help of human involvement.