PORT: Preference Optimization on Reasoning Traces

Preference optimization methods have been successfully applied to improve not only the alignment of large language models (LLMs) with human values, but also specific natural language tasks such as summarization and stylistic continuations. This paper proposes using preference optimization methods on Chain-of-Thought steps in order to improve the reasoning performances of language models. While the chosen answers are obtained from datasets that include reasoning traces, we propose two complementary schemes for generating rejected answers: digit corruption, and weak LLM prompting. Our approach leads to increased accuracy on the GSM8K, AQuA-RAT, and ARC benchmarks for Falcon2-11B and Mistral-7B. For example, the approach can lead to up to a relative 8.47% increase in accuracy on the GSM8K benchmark without any extra annotations. This work suggests that spending resources on creating more datasets of reasoning traces would further boost LLM performances on informal reasoning tasks.

Investigating Regularization of Self-Play Language Models

This paper explores the effects of various forms of regularization in the context of language model alignment via self-play. While both reinforcement learning from human feedback (RLHF) and direct preference optimization (DPO) require to collect costly human-annotated pairwise preferences, the self-play fine-tuning (SPIN) approach replaces the rejected answers by data generated from the previous iterate. However, the SPIN method presents a performance instability issue in the learning phase, which can be mitigated by playing against a mixture of the two previous iterates. In the same vein, we propose in this work to address this issue from two perspectives: first, by incorporating an additional Kullback-Leibler (KL) regularization to stay at the proximity of the reference policy; second, by using the idea of fictitious play which smoothens the opponent policy across all previous iterations. In particular, we show that the KL-based regularizer boils down to replacing the previous policy by its geometric mixture with the base policy inside of the SPIN loss function. We finally discuss empirical results on MT-Bench as well as on the Hugging Face Open LLM Leaderboard.

Self-Supervised Pretraining for Heterogeneous Hypergraph Neural Networks

Recently, pretraining methods for the Graph Neural Networks (GNNs) have been successful at learning effective representations from unlabeled graph data. However, most of these methods rely on pairwise relations in the graph and do not capture the underling higher-order relations between entities. Hypergraphs are versatile and expressive structures that can effectively model higher-order relationships among entities in the data. Despite the efforts to adapt GNNs to hypergraphs (HyperGNN), there are currently no fully self-supervised pretraining methods for HyperGNN on heterogeneous hypergraphs. In this paper, we present SPHH, a novel self-supervised pretraining framework for heterogeneous HyperGNNs. Our method is able to effectively capture higher-order relations among entities in the data in a self-supervised manner. SPHH is consist of two self-supervised pretraining tasks that aim to simultaneously learn both local and global representations of the entities in the hypergraph by using informative representations derived from the hypergraph structure. Overall, our work presents a significant advancement in the field of self-supervised pretraining of HyperGNNs, and has the potential to improve the performance of various graph-based downstream tasks such as node classification and link prediction tasks which are mapped to hypergraph configuration. Our experiments on two real-world benchmarks using four different HyperGNN models show that our proposed SPHH framework consistently outperforms state-of-the-art baselines in various downstream tasks. The results demonstrate that SPHH is able to improve the performance of various HyperGNN models in various downstream tasks, regardless of their architecture or complexity, which highlights the robustness of our framework.