AlphaPO -- Reward shape matters for LLM alignment

Reinforcement Learning with Human Feedback (RLHF) and its variants have made huge strides toward the effective alignment of large language models (LLMs) to follow instructions and reflect human values. More recently, Direct Alignment Algorithms (DAAs) have emerged in which the reward modeling stage of RLHF is skipped by characterizing the reward directly as a function of the policy being learned. Examples include Direct Preference Optimization (DPO) and Simple Preference Optimization (SimPO). These methods often suffer from likelihood displacement, a phenomenon by which the probabilities of preferred responses are often reduced undesirably. In this paper, we argue that, for DAAs the reward (function) shape matters. We introduce AlphaPO, a new DAA method that leverages an $\alpha$-parameter to help change the shape of the reward function beyond the standard log reward. AlphaPO helps maintain fine-grained control over likelihood displacement and over-optimization. Compared to SimPO, one of the best performing DAAs, AlphaPO leads to about 7\% to 10\% relative improvement in alignment performance for the instruct versions of Mistral-7B and Llama3-8B. The analysis and results presented highlight the importance of the reward shape, and how one can systematically change it to affect training dynamics, as well as improve alignment performance.

Liger Kernel: Efficient Triton Kernels for LLM Training

Training Large Language Models (LLMs) efficiently at scale presents a formidable challenge, driven by their ever-increasing computational demands and the need for enhanced performance. In this work, we introduce Liger-Kernel, an open-sourced set of Triton kernels developed specifically for LLM training. With kernel optimization techniques like kernel operation fusing and input chunking, our kernels achieve on average a 20% increase in training throughput and a 60% reduction in GPU memory usage for popular LLMs compared to HuggingFace implementations. In addition, Liger-Kernel is designed with modularity, accessibility, and adaptability in mind, catering to both casual and expert users. Comprehensive benchmarks and integration tests are built in to ensure compatibility, performance, correctness, and convergence across diverse computing environments and model architectures. The source code is available under a permissive license at: github.com/linkedin/Liger-Kernel.

Iterative proportional scaling revisited: a modern optimization perspective

This paper revisits the classic iterative proportional scaling (IPS) from a modern optimization perspective. In contrast to the criticisms made in the literature, we show that based on a coordinate descent characterization, IPS can be slightly modified to deliver coefficient estimates, and from a majorization-minimization standpoint, IPS can be extended to handle log-affine models with features not necessarily binary-valued or nonnegative. Furthermore, some state-of-the-art optimization techniques such as block-wise computation, randomization and momentum-based acceleration can be employed to provide more scalable IPS algorithms, as well as some regularized variants of IPS for concurrent feature selection.

Indirect Gaussian Graph Learning beyond Gaussianity

This paper studies how to capture dependency graph structures from real data which may not be multivariate Gaussian. Starting from marginal loss functions not necessarily derived from probability distributions, we use an additive over-parametrization with shrinkage to incorporate variable dependencies into the criterion. An iterative Gaussian graph learning algorithm is proposed with ease in implementation. Statistical analysis shows that with the error measured in terms of a proper Bregman divergence, the estimators have fast rate of convergence. Real-life examples in different settings are given to demonstrate the efficacy of the proposed methodology.