Inference-time alignment enhances the performance of large language models without requiring additional training or fine-tuning but presents challenges due to balancing computational efficiency with high-quality output. Best-of-N (BoN) sampling, as a simple yet powerful approach, generates multiple responses and selects the best one, achieving improved performance but with a high computational cost. We propose TreeBoN, a novel framework that integrates a speculative tree-search strategy into Best-of-N (BoN) Sampling. TreeBoN maintains a set of parent nodes, iteratively branching and pruning low-quality responses, thereby reducing computational overhead while maintaining high output quality. Our approach also leverages token-level rewards from Direct Preference Optimization (DPO) to guide tree expansion and prune low-quality paths. We evaluate TreeBoN using AlpacaFarm, HH-RLHF, UltraFeedback, GSM8K, and TutorEval datasets, demonstrating consistent improvements. Specifically, TreeBoN achieves the highest win rate of 65% on TutorEval and around 60% win rates across other different datasets, outperforming standard BoN with the same computational cost and showcasing its scalability and alignment efficacy.

Large Language Models (LLMs) exhibit impressive capabilities but require careful alignment with human preferences. Traditional training-time methods finetune LLMs using human preference datasets but incur significant training costs and require repeated training to handle diverse user preferences. Test-time alignment methods address this by using reward models (RMs) to guide frozen LLMs without retraining. However, existing test-time approaches rely on trajectory-level RMs which are designed to evaluate complete responses, making them unsuitable for autoregressive text generation that requires computing next-token rewards from partial responses. To address this, we introduce GenARM, a test-time alignment approach that leverages the Autoregressive Reward Model--a novel reward parametrization designed to predict next-token rewards for efficient and effective autoregressive generation. Theoretically, we demonstrate that this parametrization can provably guide frozen LLMs toward any distribution achievable by traditional RMs within the KL-regularized reinforcement learning framework. Experimental results show that GenARM significantly outperforms prior test-time alignment baselines and matches the performance of training-time methods. Additionally, GenARM enables efficient weak-to-strong guidance, aligning larger LLMs with smaller RMs without the high costs of training larger models. Furthermore, GenARM supports multi-objective alignment, allowing real-time trade-offs between preference dimensions and catering to diverse user preferences without retraining.

We propose a method to open the black box of the Multi-Layer Perceptron by inferring from it a simpler and generally more accurate general additive model. The resulting model comprises non-linear univariate and bivariate partial responses derived from the original Multi-Layer Perceptron. The responses are combined using the Lasso and further optimised within a modular structure. The approach is generic and provides a constructive framework to simplify and explain the Multi-Layer Perceptron for any data set, opening the door for validation against prior knowledge. Experimental results on benchmarking datasets indicate that the partial responses are intuitive to interpret and the Area Under the Curve is competitive with Gradient Boosting, Support Vector Machines and Random Forests. The performance improvement compared with a fully connected Multi-Layer Perceptron is attributed to reduced confounding in the second stage of optimisation of the weights. The main limitation of the method is that it explicitly models only up to pairwise interactions. For many practical applications this will be optimal, but where that is not the case then this will be indicated by the performance difference compared to the original model. The streamlined model simultaneously interprets and optimises this frequently used flexible model.