T-REG: Preference Optimization with Token-Level Reward Regularization

Reinforcement learning from human feedback (RLHF) has been crucial in aligning large language models (LLMs) with human values. Traditionally, RLHF involves generating responses to a query and using a reward model to assign a reward to the entire response. However, this approach faces challenges due to its reliance on a single, sparse reward, which makes it challenging for the model to identify which parts of the sequence contribute most significantly to the final reward. Recent methods have attempted to address this limitation by introducing token-level rewards. However, these methods often rely on either a trained credit assignment model or AI annotators, raising concerns about the quality and reliability of the rewards. In this paper, we propose token-level reward regularization (T-REG), a novel approach that leverages both sequence-level and token-level rewards for preference optimization. Harnessing the self-refinement capabilities of LLMs, our method uses contrastive prompting to enable LLMs to self-generate token-level rewards. These self-generated rewards then act as reward regularization, guiding the model to more effectively distribute sequence-level rewards across tokens. This facilitates better token-level credit assignment and enhances alignment performance. Experiments on the instruction following benchmarks, including Alpaca Eval 2 and Arena-Hard, show that our method consistently outperforms baseline methods by up to 3.8% and 4.4%, respectively. We will release the code and models at https://github.com/wzhouad/T-REG.

Instructional Segment Embedding: Improving LLM Safety with Instruction Hierarchy

Large Language Models (LLMs) are susceptible to security and safety threats, such as prompt injection, prompt extraction, and harmful requests. One major cause of these vulnerabilities is the lack of an instruction hierarchy. Modern LLM architectures treat all inputs equally, failing to distinguish between and prioritize various types of instructions, such as system messages, user prompts, and data. As a result, lower-priority user prompts may override more critical system instructions, including safety protocols. Existing approaches to achieving instruction hierarchy, such as delimiters and instruction-based training, do not address this issue at the architectural level. We introduce the Instructional Segment Embedding (ISE) technique, inspired by BERT, to modern large language models, which embeds instruction priority information directly into the model. This approach enables models to explicitly differentiate and prioritize various instruction types, significantly improving safety against malicious prompts that attempt to override priority rules. Our experiments on the Structured Query and Instruction Hierarchy benchmarks demonstrate an average robust accuracy increase of up to 15.75% and 18.68%, respectively. Furthermore, we observe an improvement in instruction-following capability of up to 4.1% evaluated on AlpacaEval. Overall, our approach offers a promising direction for enhancing the safety and effectiveness of LLM architectures.

SFTMix: Elevating Language Model Instruction Tuning with Mixup Recipe

To induce desired behaviors in large language models (LLMs) for interaction-driven tasks, the instruction-tuning stage typically trains LLMs on instruction-response pairs using the next-token prediction (NTP) loss. Previous work aiming to improve instruction-tuning performance often emphasizes the need for higher-quality supervised fine-tuning (SFT) datasets, which typically involves expensive data filtering with proprietary LLMs or labor-intensive data generation by human annotators. However, these approaches do not fully leverage the datasets' intrinsic properties, resulting in high computational and labor costs, thereby limiting scalability and performance gains. In this paper, we propose SFTMix, a novel recipe that elevates instruction-tuning performance beyond the conventional NTP paradigm, without the need for well-curated datasets. Observing that LLMs exhibit uneven confidence across the semantic representation space, we argue that examples with different confidence levels should play distinct roles during the instruction-tuning process. Based on this insight, SFTMix leverages training dynamics to identify examples with varying confidence levels, then applies a Mixup-based regularization to mitigate overfitting on confident examples while propagating supervision signals to improve learning on relatively unconfident ones. This approach enables SFTMix to significantly outperform NTP across a wide range of instruction-following and healthcare domain-specific SFT tasks, demonstrating its adaptability to diverse LLM families and scalability to datasets of any size. Comprehensive ablation studies further verify the robustness of SFTMix's design choices, underscoring its versatility in consistently enhancing performance across different LLMs and datasets in broader natural language processing applications.

Score Forgetting Distillation: A Swift, Data-Free Method for Machine Unlearning in Diffusion Models

The machine learning community is increasingly recognizing the importance of fostering trust and safety in modern generative AI (GenAI) models. We posit machine unlearning (MU) as a crucial foundation for developing safe, secure, and trustworthy GenAI models. Traditional MU methods often rely on stringent assumptions and require access to real data. This paper introduces Score Forgetting Distillation (SFD), an innovative MU approach that promotes the forgetting of undesirable information in diffusion models by aligning the conditional scores of ``unsafe'' classes or concepts with those of ``safe'' ones. To eliminate the need for real data, our SFD framework incorporates a score-based MU loss into the score distillation objective of a pretrained diffusion model. This serves as a regularization term that preserves desired generation capabilities while enabling the production of synthetic data through a one-step generator. Our experiments on pretrained label-conditional and text-to-image diffusion models demonstrate that our method effectively accelerates the forgetting of target classes or concepts during generation, while preserving the quality of other classes or concepts. This unlearned and distilled diffusion not only pioneers a novel concept in MU but also accelerates the generation speed of diffusion models. Our experiments and studies on a range of diffusion models and datasets confirm that our approach is generalizable, effective, and advantageous for MU in diffusion models.

WPO: Enhancing RLHF with Weighted Preference Optimization

Reinforcement learning from human feedback (RLHF) is a promising solution to align large language models (LLMs) more closely with human values. Off-policy preference optimization, where the preference data is obtained from other models, is widely adopted due to its cost efficiency and scalability. However, off-policy preference optimization often suffers from a distributional gap between the policy used for data collection and the target policy, leading to suboptimal optimization. In this paper, we propose a novel strategy to mitigate this problem by simulating on-policy learning with off-policy preference data. Our Weighted Preference Optimization (WPO) method adapts off-policy data to resemble on-policy data more closely by reweighting preference pairs according to their probability under the current policy. This method not only addresses the distributional gap problem but also enhances the optimization process without incurring additional costs. We validate our method on instruction following benchmarks including Alpaca Eval 2 and MT-bench. WPO not only outperforms Direct Preference Optimization (DPO) by up to 5.6% on Alpaca Eval 2 but also establishes a remarkable length-controlled winning rate against GPT-4-turbo of 48.6% based on Llama-3-8B-Instruct, making it the strongest 8B model on the leaderboard. We will release the code and models at https://github.com/wzhouad/WPO.

Statistical Advantages of Perturbing Cosine Router in Sparse Mixture of Experts

The cosine router in sparse Mixture of Experts (MoE) has recently emerged as an attractive alternative to the conventional linear router. Indeed, the cosine router demonstrates favorable performance in image and language tasks and exhibits better ability to mitigate the representation collapse issue, which often leads to parameter redundancy and limited representation potentials. Despite its empirical success, a comprehensive analysis of the cosine router in sparse MoE has been lacking. Considering the least square estimation of the cosine routing sparse MoE, we demonstrate that due to the intrinsic interaction of the model parameters in the cosine router via some partial differential equations, regardless of the structures of the experts, the estimation rates of experts and model parameters can be as slow as $\mathcal{O}(1/\log^{\tau}(n))$ where $\tau > 0$ is some constant and $n$ is the sample size. Surprisingly, these pessimistic non-polynomial convergence rates can be circumvented by the widely used technique in practice to stabilize the cosine router -- simply adding noises to the $\mathbb{L}_{2}$ norms in the cosine router, which we refer to as \textit{perturbed cosine router}. Under the strongly identifiable settings of the expert functions, we prove that the estimation rates for both the experts and model parameters under the perturbed cosine routing sparse MoE are significantly improved to polynomial rates. Finally, we conduct extensive simulation studies in both synthetic and real data settings to empirically validate our theoretical results.

Switchable Decision: Dynamic Neural Generation Networks

Auto-regressive generation models achieve competitive performance across many different NLP tasks such as summarization, question answering, and classifications. However, they are also known for being slow in inference, which makes them challenging to deploy in real-time applications. We propose a switchable decision to accelerate inference by dynamically assigning computation resources for each data instance. Automatically making decisions on where to skip and how to balance quality and computation cost with constrained optimization, our dynamic neural generation networks enforce the efficient inference path and determine the optimized trade-off. Experiments across question answering, summarization, and classification benchmarks show that our method benefits from less computation cost during inference while keeping the same accuracy. Extensive experiments and ablation studies demonstrate that our method can be general, effective, and beneficial for many NLP tasks.

Language Rectified Flow: Advancing Diffusion Language Generation with Probabilistic Flows

Recent works have demonstrated success in controlling sentence attributes ($e.g.$, sentiment) and structure ($e.g.$, syntactic structure) based on the diffusion language model. A key component that drives theimpressive performance for generating high-quality samples from noise is iteratively denoise for thousands of steps. While beneficial, the complexity of starting from the noise and the learning steps has limited its implementation to many NLP real-world applications. This paper proposes Language Rectified Flow ({\ours}). Our method is based on the reformulation of the standard probabilistic flow models. Language rectified flow learns (neural) ordinary differential equation models to transport between the source distribution and the target distribution, hence providing a unified and effective solution to generative modeling and domain transfer. From the source distribution, our language rectified flow yields fast simulation and effectively decreases the inference time. Experiments on three challenging fine-grained control tasks and multiple high-quality text editing show that our method consistently outperforms its baselines. Extensive experiments and ablation studies demonstrate that our method can be general, effective, and beneficial for many NLP tasks.

Sliced Wasserstein with Random-Path Projecting Directions

Slicing distribution selection has been used as an effective technique to improve the performance of parameter estimators based on minimizing sliced Wasserstein distance in applications. Previous works either utilize expensive optimization to select the slicing distribution or use slicing distributions that require expensive sampling methods. In this work, we propose an optimization-free slicing distribution that provides a fast sampling for the Monte Carlo estimation of expectation. In particular, we introduce the random-path projecting direction (RPD) which is constructed by leveraging the normalized difference between two random vectors following the two input measures. From the RPD, we derive the random-path slicing distribution (RPSD) and two variants of sliced Wasserstein, i.e., the Random-Path Projection Sliced Wasserstein (RPSW) and the Importance Weighted Random-Path Projection Sliced Wasserstein (IWRPSW). We then discuss the topological, statistical, and computational properties of RPSW and IWRPSW. Finally, we showcase the favorable performance of RPSW and IWRPSW in gradient flow and the training of denoising diffusion generative models on images.

Preference-grounded Token-level Guidance for Language Model Fine-tuning

Aligning language models (LMs) with preferences is an important problem in natural language generation. A key challenge is that preferences are typically provided at the sequence level while LM training and generation both occur at the token level. There is, therefore, a granularity mismatch between the preference and the LM training losses, which may complicate the learning problem. In this paper, we address this issue by developing an alternate training process, where we iterate between grounding the sequence-level preference into token-level training guidance, and improving the LM with the learned guidance. For guidance learning, we design a framework that extends the pairwise-preference learning in imitation learning to both variable-length LM generation and utilizing the preference among multiple generations. For LM training, based on the amount of supervised data, we present two minimalist learning objectives that utilize the learned guidance. In experiments, our method performs competitively on two distinct representative LM tasks -- discrete-prompt generation and text summarization.