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.

While generative modeling on multimodal image-text data has been actively developed with large-scale paired datasets, there have been limited attempts to generate both image and text data by a single model rather than a generation of one fixed modality conditioned on the other modality. In this paper, we explore a unified generative vision-and-language (VL) model that can produce both images and text sequences. Especially, we propose a generative VL transformer based on the non-autoregressive mask prediction, named MAGVLT, and compare it with an autoregressive generative VL transformer (ARGVLT). In comparison to ARGVLT, the proposed MAGVLT enables bidirectional context encoding, fast decoding by parallel token predictions in an iterative refinement, and extended editing capabilities such as image and text infilling. For rigorous training of our MAGVLT with image-text pairs from scratch, we combine the image-to-text, text-to-image, and joint image-and-text mask prediction tasks. Moreover, we devise two additional tasks based on the step-unrolled mask prediction and the selective prediction on the mixture of two image-text pairs. Experimental results on various downstream generation tasks of VL benchmarks show that our MAGVLT outperforms ARGVLT by a large margin even with significant inference speedup. Particularly, MAGVLT achieves competitive results on both zero-shot image-to-text and text-to-image generation tasks from MS-COCO by one moderate-sized model (fewer than 500M parameters) even without the use of monomodal data and networks.

Natural language modeling with limited training data is a challenging problem, and many algorithms make use of large-scale pretrained language models (PLMs) for this due to its great generalization ability. Among them, additive learning that incorporates a task-specific adapter on top of the fixed large-scale PLM has been popularly used in the few-shot setting. However, this added adapter is still easy to disregard the knowledge of the PLM especially for few-shot natural language generation (NLG) since an entire sequence is usually generated by only the newly trained adapter. Therefore, in this work, we develop a novel additive learning algorithm based on reinforcement learning (RL) that selectively outputs language tokens between the task-general PLM and the task-specific adapter during both training and inference. This output token selection over the two generators allows the adapter to take into account solely the task-relevant parts in sequence generation, and therefore makes it more robust to overfitting as well as more stable in RL training. In addition, to obtain the complementary adapter from the PLM for each few-shot task, we exploit a separate selecting module that is also simultaneously trained using RL. Experimental results on various few-shot NLG tasks including question answering, data-to-text generation and text summarization demonstrate that the proposed selective token generation significantly outperforms the previous additive learning algorithms based on the PLMs.

Episodic count has been widely used to design a simple yet effective intrinsic motivation for reinforcement learning with a sparse reward. However, the use of episodic count in a high-dimensional state space as well as over a long episode time requires a thorough state compression and fast hashing, which hinders rigorous exploitation of it in such hard and complex exploration environments. Moreover, the interference from task-irrelevant observations in the episodic count may cause its intrinsic motivation to overlook task-related important changes of states, and the novelty in an episodic manner can lead to repeatedly revisit the familiar states across episodes. In order to resolve these issues, in this paper, we propose a learnable hash-based episodic count, which we name LECO, that efficiently performs as a task-specific intrinsic reward in hard exploration problems. In particular, the proposed intrinsic reward consists of the episodic novelty and the task-specific modulation where the former employs a vector quantized variational autoencoder to automatically obtain the discrete state codes for fast counting while the latter regulates the episodic novelty by learning a modulator to optimize the task-specific extrinsic reward. The proposed LECO specifically enables the automatic transition from exploration to exploitation during reinforcement learning. We experimentally show that in contrast to the previous exploration methods LECO successfully solves hard exploration problems and also scales to large state spaces through the most difficult tasks in MiniGrid and DMLab environments.