Optimizing Chain-of-Thought Reasoners via Gradient Variance Minimization in Rejection Sampling and RL

Chain-of-thought (CoT) reasoning in large language models (LLMs) can be formalized as a latent variable problem, where the model needs to generate intermediate reasoning steps. While prior approaches such as iterative reward-ranked fine-tuning (RAFT) have relied on such formulations, they typically apply uniform inference budgets across prompts, which fails to account for variability in difficulty and convergence behavior. This work identifies the main bottleneck in CoT training as inefficient stochastic gradient estimation due to static sampling strategies. We propose GVM-RAFT, a prompt-specific Dynamic Sample Allocation Strategy designed to minimize stochastic gradient variance under a computational budget constraint. The method dynamically allocates computational resources by monitoring prompt acceptance rates and stochastic gradient norms, ensuring that the resulting gradient variance is minimized. Our theoretical analysis shows that the proposed dynamic sampling strategy leads to accelerated convergence guarantees under suitable conditions. Experiments on mathematical reasoning show that GVM-RAFT achieves a 2-4x speedup and considerable accuracy improvements over vanilla RAFT. The proposed dynamic sampling strategy is general and can be incorporated into other reinforcement learning algorithms, such as GRPO, leading to similar improvements in convergence and test accuracy. Our code is available at https://github.com/RLHFlow/GVM.

RM-R1: Reward Modeling as Reasoning

Reward modeling is essential for aligning large language models (LLMs) with human preferences, especially through reinforcement learning from human feedback (RLHF). To provide accurate reward signals, a reward model (RM) should stimulate deep thinking and conduct interpretable reasoning before assigning a score or a judgment. However, existing RMs either produce opaque scalar scores or directly generate the prediction of a preferred answer, making them struggle to integrate natural language critiques, thus lacking interpretability. Inspired by recent advances of long chain-of-thought (CoT) on reasoning-intensive tasks, we hypothesize and validate that integrating reasoning capabilities into reward modeling significantly enhances RM's interpretability and performance. In this work, we introduce a new class of generative reward models -- Reasoning Reward Models (ReasRMs) -- which formulate reward modeling as a reasoning task. We propose a reasoning-oriented training pipeline and train a family of ReasRMs, RM-R1. The training consists of two key stages: (1) distillation of high-quality reasoning chains and (2) reinforcement learning with verifiable rewards. RM-R1 improves LLM rollouts by self-generating reasoning traces or chat-specific rubrics and evaluating candidate responses against them. Empirically, our models achieve state-of-the-art or near state-of-the-art performance of generative RMs across multiple comprehensive reward model benchmarks, outperforming much larger open-weight models (e.g., Llama3.1-405B) and proprietary ones (e.g., GPT-4o) by up to 13.8%. Beyond final performance, we perform thorough empirical analysis to understand the key ingredients of successful ReasRM training. To facilitate future research, we release six ReasRM models along with code and data at https://github.com/RM-R1-UIUC/RM-R1.

Multi-Step Consistency Models: Fast Generation with Theoretical Guarantees

Consistency models have recently emerged as a compelling alternative to traditional SDE based diffusion models, offering a significant acceleration in generation by producing high quality samples in very few steps. Despite their empirical success, a proper theoretic justification for their speed up is still lacking. In this work, we provide the analysis which bridges this gap, showing that given a consistency model which can map the input at a given time to arbitrary timestamps along the reverse trajectory, one can achieve KL divergence of order $ O(\varepsilon^2) $ using only $ O\left(\log\left(\frac{d}{\varepsilon}\right)\right) $ iterations with constant step size, where d is the data dimension. Additionally, under minimal assumptions on the data distribution an increasingly common setting in recent diffusion model analyses we show that a similar KL convergence guarantee can be obtained, with the number of steps scaling as $ O\left(d \log\left(\frac{d}{\varepsilon}\right)\right) $. Going further, we also provide a theoretical analysis for estimation of such consistency models, concluding that accurate learning is feasible using small discretization steps, both in smooth and non smooth settings. Notably, our results for the non smooth case yield best in class convergence rates compared to existing SDE or ODE based analyses under minimal assumptions.

MIMIC-\RNum{4}-Ext-22MCTS: A 22 Millions-Event Temporal Clinical Time-Series Dataset with Relative Timestamp for Risk Prediction

Clinical risk prediction based on machine learning algorithms plays a vital role in modern healthcare. A crucial component in developing a reliable prediction model is collecting high-quality time series clinical events. In this work, we release such a dataset that consists of 22,588,586 Clinical Time Series events, which we term MIMIC-\RNum{4}-Ext-22MCTS. Our source data are discharge summaries selected from the well-known yet unstructured MIMIC-IV-Note \cite{Johnson2023-pg}. We then extract clinical events as short text span from the discharge summaries, along with the timestamps of these events as temporal information. The general-purpose MIMIC-IV-Note pose specific challenges for our work: it turns out that the discharge summaries are too lengthy for typical natural language models to process, and the clinical events of interest often are not accompanied with explicit timestamps. Therefore, we propose a new framework that works as follows: 1) we break each discharge summary into manageably small text chunks; 2) we apply contextual BM25 and contextual semantic search to retrieve chunks that have a high potential of containing clinical events; and 3) we carefully design prompts to teach the recently released Llama-3.1-8B \cite{touvron2023llama} model to identify or infer temporal information of the chunks. We show that the obtained dataset is so informative and transparent that standard models fine-tuned on our dataset are achieving significant improvements in healthcare applications. In particular, the BERT model fine-tuned based on our dataset achieves 10\% improvement in accuracy on medical question answering task, and 3\% improvement in clinical trial matching task compared with the classic BERT. The GPT-2 model, fine-tuned on our dataset, produces more clinically reliable results for clinical questions.

Capturing Conditional Dependence via Auto-regressive Diffusion Models

Diffusion models have demonstrated appealing performance in both image and video generation. However, many works discover that they struggle to capture important, high-level relationships that are present in the real world. For example, they fail to learn physical laws from data, and even fail to understand that the objects in the world exist in a stable fashion. This is due to the fact that important conditional dependence structures are not adequately captured in the vanilla diffusion models. In this work, we initiate an in-depth study on strengthening the diffusion model to capture the conditional dependence structures in the data. In particular, we examine the efficacy of the auto-regressive (AR) diffusion models for such purpose and develop the first theoretical results on the sampling error of AR diffusion models under (possibly) the mildest data assumption. Our theoretical findings indicate that, compared with typical diffusion models, the AR variant produces samples with a reduced gap in approximating the data conditional distribution. On the other hand, the overall inference time of the AR-diffusion models is only moderately larger than that for the vanilla diffusion models, making them still practical for large scale applications. We also provide empirical results showing that when there is clear conditional dependence structure in the data, the AR diffusion models captures such structure, whereas vanilla DDPM fails to do so. On the other hand, when there is no obvious conditional dependence across patches of the data, AR diffusion does not outperform DDPM.

STCL:Curriculum learning Strategies for deep learning image steganography models

Aiming at the problems of poor quality of steganographic images and slow network convergence of image steganography models based on deep learning, this paper proposes a Steganography Curriculum Learning training strategy (STCL) for deep learning image steganography models. So that only easy images are selected for training when the model has poor fitting ability at the initial stage, and gradually expand to more difficult images, the strategy includes a difficulty evaluation strategy based on the teacher model and an knee point-based training scheduling strategy. Firstly, multiple teacher models are trained, and the consistency of the quality of steganographic images under multiple teacher models is used as the difficulty score to construct the training subsets from easy to difficult. Secondly, a training control strategy based on knee points is proposed to reduce the possibility of overfitting on small training sets and accelerate the training process. Experimental results on three large public datasets, ALASKA2, VOC2012 and ImageNet, show that the proposed image steganography scheme is able to improve the model performance under multiple algorithmic frameworks, which not only has a high PSNR, SSIM score, and decoding accuracy, but also the steganographic images generated by the model under the training of the STCL strategy have a low steganography analysis scores. You can find our code at \href{https://github.com/chaos-boops/STCL}{https://github.com/chaos-boops/STCL}.

CLPSTNet: A Progressive Multi-Scale Convolutional Steganography Model Integrating Curriculum Learning

In recent years, a large number of works have introduced Convolutional Neural Networks (CNNs) into image steganography, which transform traditional steganography methods such as hand-crafted features and prior knowledge design into steganography methods that neural networks autonomically learn information embedding. However, due to the inherent complexity of digital images, issues of invisibility and security persist when using CNN models for information embedding. In this paper, we propose Curriculum Learning Progressive Steganophy Network (CLPSTNet). The network consists of multiple progressive multi-scale convolutional modules that integrate Inception structures and dilated convolutions. The module contains multiple branching pathways, starting from a smaller convolutional kernel and dilatation rate, extracting the basic, local feature information from the feature map, and gradually expanding to the convolution with a larger convolutional kernel and dilatation rate for perceiving the feature information of a larger receptive field, so as to realize the multi-scale feature extraction from shallow to deep, and from fine to coarse, allowing the shallow secret information features to be refined in different fusion stages. The experimental results show that the proposed CLPSTNet not only has high PSNR , SSIM metrics and decoding accuracy on three large public datasets, ALASKA2, VOC2012 and ImageNet, but also the steganographic images generated by CLPSTNet have low steganalysis scores.You can find our code at \href{https://github.com/chaos-boops/CLPSTNet}{https://github.com/chaos-boops/CLPSTNet}.

MMHCL: Multi-Modal Hypergraph Contrastive Learning for Recommendation

The burgeoning presence of multimodal content-sharing platforms propels the development of personalized recommender systems. Previous works usually suffer from data sparsity and cold-start problems, and may fail to adequately explore semantic user-product associations from multimodal data. To address these issues, we propose a novel Multi-Modal Hypergraph Contrastive Learning (MMHCL) framework for user recommendation. For a comprehensive information exploration from user-product relations, we construct two hypergraphs, i.e. a user-to-user (u2u) hypergraph and an item-to-item (i2i) hypergraph, to mine shared preferences among users and intricate multimodal semantic resemblance among items, respectively. This process yields denser second-order semantics that are fused with first-order user-item interaction as complementary to alleviate the data sparsity issue. Then, we design a contrastive feature enhancement paradigm by applying synergistic contrastive learning. By maximizing/minimizing the mutual information between second-order (e.g. shared preference pattern for users) and first-order (information of selected items for users) embeddings of the same/different users and items, the feature distinguishability can be effectively enhanced. Compared with using sparse primary user-item interaction only, our MMHCL obtains denser second-order hypergraphs and excavates more abundant shared attributes to explore the user-product associations, which to a certain extent alleviates the problems of data sparsity and cold-start. Extensive experiments have comprehensively demonstrated the effectiveness of our method. Our code is publicly available at: https://github.com/Xu107/MMHCL.

EarthGPT-X: Enabling MLLMs to Flexibly and Comprehensively Understand Multi-Source Remote Sensing Imagery

Recent advances in the visual-language area have developed natural multi-modal large language models (MLLMs) for spatial reasoning through visual prompting. However, due to remote sensing (RS) imagery containing abundant geospatial information that differs from natural images, it is challenging to effectively adapt natural spatial models to the RS domain. Moreover, current RS MLLMs are limited in overly narrow interpretation levels and interaction manner, hindering their applicability in real-world scenarios. To address those challenges, a spatial MLLM named EarthGPT-X is proposed, enabling a comprehensive understanding of multi-source RS imagery, such as optical, synthetic aperture radar (SAR), and infrared. EarthGPT-X offers zoom-in and zoom-out insight, and possesses flexible multi-grained interactive abilities. Moreover, EarthGPT-X unifies two types of critical spatial tasks (i.e., referring and grounding) into a visual prompting framework. To achieve these versatile capabilities, several key strategies are developed. The first is the multi-modal content integration method, which enhances the interplay between images, visual prompts, and text instructions. Subsequently, a cross-domain one-stage fusion training strategy is proposed, utilizing the large language model (LLM) as a unified interface for multi-source multi-task learning. Furthermore, by incorporating a pixel perception module, the referring and grounding tasks are seamlessly unified within a single framework. In addition, the experiments conducted demonstrate the superiority of the proposed EarthGPT-X in multi-grained tasks and its impressive flexibility in multi-modal interaction, revealing significant advancements of MLLM in the RS field.

A Minimalist Approach to LLM Reasoning: from Rejection Sampling to Reinforce

Reinforcement learning (RL) has become a prevailing approach for fine-tuning large language models (LLMs) on complex reasoning tasks. Among recent methods, GRPO stands out for its empirical success in training models such as DeepSeek-R1, yet the sources of its effectiveness remain poorly understood. In this work, we revisit GRPO from a reinforce-like algorithm perspective and analyze its core components. Surprisingly, we find that a simple rejection sampling baseline, RAFT, which trains only on positively rewarded samples, yields competitive performance than GRPO and PPO. Our ablation studies reveal that GRPO's main advantage arises from discarding prompts with entirely incorrect responses, rather than from its reward normalization. Motivated by this insight, we propose Reinforce-Rej, a minimal extension of policy gradient that filters both entirely incorrect and entirely correct samples. Reinforce-Rej improves KL efficiency and stability, serving as a lightweight yet effective alternative to more complex RL algorithms. We advocate RAFT as a robust and interpretable baseline, and suggest that future advances should focus on more principled designs for incorporating negative samples, rather than relying on them indiscriminately. Our findings provide guidance for future work in reward-based LLM post-training.