Direct alignment from preferences (DAP) has emerged as a promising paradigm for aligning large language models (LLMs) to human desiderata from pre-collected, offline preference datasets. While recent studies indicate that existing offline DAP methods can directly benefit from online training samples, we highlight the need to develop specific online DAP algorithms to fully harness the power of online training. Specifically, we identify that the learned LLM should adhere to the proximity of the behavior LLM, which collects the training samples. To this end, we propose online Preference Optimization in proximity to the Behavior LLM (BPO), emphasizing the importance of constructing a proper trust region for LLM alignment.We conduct extensive experiments to validate the effectiveness and applicability of our approach by integrating it with various DAP methods, resulting in significant performance improvements across a wide range of tasks when training with the same amount of preference data. Even when only introducing one additional data collection phase, our online BPO improves its offline DAP baseline from 72.0% to 80.2% on TL;DR and from 82.2% to 89.1% on Anthropic Helpfulness in terms of win rate against human reference text.

While most existing works on LLM prompting techniques focus only on how to select a better set of data samples inside one single prompt input (In-Context Learning or ICL), why can not we design and leverage multiple prompts together to further improve the LLM’s performance? In this work, we propose In-Context Sampling (ICS), a low-resource LLM prompting technique to produce confident predictions by optimizing the construction of multiple ICL prompt inputs. Extensive experiments with three open-source LLMs (FlanT5-XL, Mistral-7B, and Mixtral-8x7B) on four NLI datasets (e-SNLI, Multi-NLI, ANLI, and Contract-NLI) and one QA dataset (CommonsenseQA) illustrate that ICS can consistently enhance LLMs’ performance. An in-depth evaluation with three data similarity-based ICS strategies suggests that these strategies can further elevate LLM’s performance, which sheds light on a new yet promising future research direction.

Recent studies show that text-to-image (T2I) models are vulnerable to adversarial attacks, especially with noun perturbations in text prompts. In this study, we investigate the impact of adversarial attacks on different POS tags within text prompts on the images generated by T2I models. We create a high-quality dataset for realistic POS tag token swapping and perform gradient-based attacks to find adversarial suffixes that mislead T2I models into generating images with altered tokens. Our empirical results show that the attack success rate (ASR) varies significantly among different POS tag categories, with nouns, proper nouns, and adjectives being the easiest to attack. We explore the mechanism behind the steering effect of adversarial suffixes, finding that the number of critical tokens and information fusion vary among POS tags, while features like suffix transferability are consistent across categories.

Reinforcement learning with natural language action space often suffers from the curse of dimensionality due to the combinatorial nature of the natural language. Previous research leverages pretrained language models to capture action semantics and reduce the size of the action space. However, since pretrained models are typically trained on general corpora, there can be an unpredictable mismatch between the priors encoded in pretrained models and the characteristics of the specific RL environment. To address this issue, we propose Mutual-Information Regularized Policy Optimization, MIPO. MIPO enables implicit and dynamic reduction of the action space. Starting from the prior provided by the pretrained language model, our method dynamically adjusts the prior during the learning process based on the guidance of mutual information regularization. Theoretically, we demonstrate that this policy optimization process leads to the monotonic improvement on the mutual-information regularized RL objective. Empirically, we conduct experiments in various environments and demonstrate the effectiveness of MIPO.

Recent studies reveal that Large Language Models (LLMs) face challenges in balancing safety with utility, particularly when processing long texts for NLP tasks like summarization and translation. Despite defenses against malicious short questions, the ability of LLMs to safely handle dangerous long content, such as manuals teaching illicit activities, remains unclear. Our work aims to develop robust defenses for LLMs in processing malicious documents alongside benign NLP task queries. We introduce a defense dataset comprised of safety-related examples and propose single-task and mixed-task losses for instruction tuning. Our empirical results demonstrate that LLMs can significantly enhance their capacity to safely manage dangerous content with appropriate instruction tuning. Additionally, strengthening the defenses of tasks most susceptible to misuse is effective in protecting LLMs against processing harmful information. We also observe that trade-offs between utility and safety exist in defense strategies, where Llama2, utilizing our proposed approach, displays a significantly better balance compared to Llama1.