Wei Xiong

Yong Lin, Hangyu Lin, Wei Xiong et al.

LLMs acquire a wide range of abilities during pre-training, but aligning LLMs under Reinforcement Learning with Human Feedback (RLHF) can lead to forgetting pretrained abilities, which is also known as the alignment tax. To investigate alignment tax, we conducted experiments with existing RLHF algorithms using OpenLLaMA-3B, which revealed a pronounced alignment tax in NLP tasks. Whereas, despite various techniques to mitigate forgetting, they are often at odds with the RLHF performance, leading to a trade-off between alignment performance and forgetting mitigation, leading to an alignment-forgetting trade-off. In this paper we show that model averaging, which simply interpolates between pre and post RLHF model weights, surprisingly achieves the most strongest alignment-forgetting Pareto front among a wide range of competing methods. To understand its effectiveness, we offer theoretical insights into model averaging, revealing that it enhances performance Pareto front by increasing feature diversity on the layers where tasks share overlapped feature spaces. Empirical evidence corroborates our analysis by showing the benefits of averaging low-level transformer layers. Building on the analysis and the observation that averaging different layers of the transformer leads to significantly different alignment-forgetting trade-offs, we propose Heterogeneous Model Averaging (HMA) to Heterogeneously find various combination ratios of model layers. HMA seeks to maximize the alignment performance while incurring minimal alignment tax. Moreover, we validate HMA's performance across a range of RLHF algorithms over OpenLLaMA-3B and further extend our findings to Mistral-7B which is evaluated by open-sourced preference model and GPT4. Code available here: https://github.com/avalonstrel/Mitigating-the-Alignment-Tax-of-RLHF.git.

Shizhe Diao, Rui Pan, Hanze Dong et al.

Foundation models have demonstrated a great ability to achieve general human-level intelligence far beyond traditional approaches. As the technique keeps attracting attention from the AI community, an increasing number of foundation models are becoming publicly accessible. However, a significant shortcoming of most of these models lies in their performance in specialized-domain and task-specific applications, necessitating domain- and task-aware fine-tuning to develop effective scientific language models. As the number of available foundation models and specialized tasks keeps growing, the job of training scientific language models becomes highly nontrivial. In this paper, we initiate steps to tackle this issue. We introduce an extensible and lightweight toolkit, LMFlow, which aims to simplify the domain- and task-aware finetuning of general foundation models. LMFlow offers a complete finetuning workflow for a foundation model to support specialized training with limited computing resources. Furthermore, it supports continuous pretraining, instruction tuning, parameter-efficient finetuning, alignment tuning, inference acceleration, long context generalization, model customization, and even multimodal finetuning, along with carefully designed and extensible APIs. This toolkit has been thoroughly tested and is available at https://github.com/OptimalScale/LMFlow.

Yawei Li, Kai Zhang, Radu Timofte et al. · eth-zurich, tencent-ai

This paper reviews the NTIRE 2022 challenge on efficient single image super-resolution with focus on the proposed solutions and results. The task of the challenge was to super-resolve an input image with a magnification factor of $\times$4 based on pairs of low and corresponding high resolution images. The aim was to design a network for single image super-resolution that achieved improvement of efficiency measured according to several metrics including runtime, parameters, FLOPs, activations, and memory consumption while at least maintaining the PSNR of 29.00dB on DIV2K validation set. IMDN is set as the baseline for efficiency measurement. The challenge had 3 tracks including the main track (runtime), sub-track one (model complexity), and sub-track two (overall performance). In the main track, the practical runtime performance of the submissions was evaluated. The rank of the teams were determined directly by the absolute value of the average runtime on the validation set and test set. In sub-track one, the number of parameters and FLOPs were considered. And the individual rankings of the two metrics were summed up to determine a final ranking in this track. In sub-track two, all of the five metrics mentioned in the description of the challenge including runtime, parameter count, FLOPs, activations, and memory consumption were considered. Similar to sub-track one, the rankings of five metrics were summed up to determine a final ranking. The challenge had 303 registered participants, and 43 teams made valid submissions. They gauge the state-of-the-art in efficient single image super-resolution.

Hanze Dong, Wei Xiong, Deepanshu Goyal et al.

Generative foundation models are susceptible to implicit biases that can arise from extensive unsupervised training data. Such biases can produce suboptimal samples, skewed outcomes, and unfairness, with potentially serious consequences. Consequently, aligning these models with human ethics and preferences is an essential step toward ensuring their responsible and effective deployment in real-world applications. Prior research has primarily employed Reinforcement Learning from Human Feedback (RLHF) to address this problem, where generative models are fine-tuned with RL algorithms guided by a human-feedback-informed reward model. However, the inefficiencies and instabilities associated with RL algorithms frequently present substantial obstacles to the successful alignment, necessitating the development of a more robust and streamlined approach. To this end, we introduce a new framework, Reward rAnked FineTuning (RAFT), designed to align generative models effectively. Utilizing a reward model and a sufficient number of samples, our approach selects the high-quality samples, discarding those that exhibit undesired behavior, and subsequently enhancing the model by fine-tuning on these filtered samples. Our studies show that RAFT can effectively improve the model performance in both reward learning and other automated metrics in both large language models and diffusion models.

Wei Xiong, Neil Yeung, Shubo Wang et al. · amazon-science

Objective and Impact Statement. We adopt a deep learning model for bone osteolysis prediction on computed tomography (CT) images of murine breast cancer bone metastases. Given the bone CT scans at previous time steps, the model incorporates the bone-cancer interactions learned from the sequential images and generates future CT images. Its ability of predicting the development of bone lesions in cancer-invading bones can assist in assessing the risk of impending fractures and choosing proper treatments in breast cancer bone metastasis. Introduction. Breast cancer often metastasizes to bone, causes osteolytic lesions, and results in skeletal related events (SREs) including severe pain and even fatal fractures. Although current imaging techniques can detect macroscopic bone lesions, predicting the occurrence and progression of bone lesions remains a challenge. Methods. We adopt a temporal variational auto-encoder (T-VAE) model that utilizes a combination of variational auto-encoders and long short-term memory networks to predict bone lesion emergence on our micro-CT dataset containing sequential images of murine tibiae. Given the CT scans of murine tibiae at early weeks, our model can learn the distribution of their future states from data. Results. We test our model against other deep learning-based prediction models on the bone lesion progression prediction task. Our model produces much more accurate predictions than existing models under various evaluation metrics. Conclusion. We develop a deep learning framework that can accurately predict and visualize the progression of osteolytic bone lesions. It will assist in planning and evaluating treatment strategies to prevent SREs in breast cancer patients.

Wei Xiong, Han Zhong, Chengshuai Shi et al.

Offline reinforcement learning (RL) aims at learning an optimal strategy using a pre-collected dataset without further interactions with the environment. While various algorithms have been proposed for offline RL in the previous literature, the minimax optimality has only been (nearly) established for tabular Markov decision processes (MDPs). In this paper, we focus on offline RL with linear function approximation and propose a new pessimism-based algorithm for offline linear MDP. At the core of our algorithm is the uncertainty decomposition via a reference function, which is new in the literature of offline RL under linear function approximation. Theoretical analysis demonstrates that our algorithm can match the performance lower bound up to logarithmic factors. We also extend our techniques to the two-player zero-sum Markov games (MGs), and establish a new performance lower bound for MGs, which tightens the existing result, and verifies the nearly minimax optimality of the proposed algorithm. To the best of our knowledge, these are the first computationally efficient and nearly minimax optimal algorithms for offline single-agent MDPs and MGs with linear function approximation.

Han Zhong, Wei Xiong, Sirui Zheng et al.

We study sample efficient reinforcement learning (RL) under the general framework of interactive decision making, which includes Markov decision process (MDP), partially observable Markov decision process (POMDP), and predictive state representation (PSR) as special cases. Toward finding the minimum assumption that empowers sample efficient learning, we propose a novel complexity measure, generalized eluder coefficient (GEC), which characterizes the fundamental tradeoff between exploration and exploitation in online interactive decision making. In specific, GEC captures the hardness of exploration by comparing the error of predicting the performance of the updated policy with the in-sample training error evaluated on the historical data. We show that RL problems with low GEC form a remarkably rich class, which subsumes low Bellman eluder dimension problems, bilinear class, low witness rank problems, PO-bilinear class, and generalized regular PSR, where generalized regular PSR, a new tractable PSR class identified by us, includes nearly all known tractable POMDPs and PSRs. Furthermore, in terms of algorithm design, we propose a generic posterior sampling algorithm, which can be implemented in both model-free and model-based fashion, under both fully observable and partially observable settings. The proposed algorithm modifies the standard posterior sampling algorithm in two aspects: (i) we use an optimistic prior distribution that biases towards hypotheses with higher values and (ii) a loglikelihood function is set to be the empirical loss evaluated on the historical data, where the choice of loss function supports both model-free and model-based learning. We prove that the proposed algorithm is sample efficient by establishing a sublinear regret upper bound in terms of GEC. In summary, we provide a new and unified understanding of both fully observable and partially observable RL.

Chenlu Ye, Wei Xiong, Quanquan Gu et al.

Despite the significant interest and progress in reinforcement learning (RL) problems with adversarial corruption, current works are either confined to the linear setting or lead to an undesired $\tilde{O}(\sqrt{T}ζ)$ regret bound, where $T$ is the number of rounds and $ζ$ is the total amount of corruption. In this paper, we consider the contextual bandit with general function approximation and propose a computationally efficient algorithm to achieve a regret of $\tilde{O}(\sqrt{T}+ζ)$. The proposed algorithm relies on the recently developed uncertainty-weighted least-squares regression from linear contextual bandit and a new weighted estimator of uncertainty for the general function class. In contrast to the existing analysis that heavily relies on the linear structure, we develop a novel technique to control the sum of weighted uncertainty, thus establishing the final regret bounds. We then generalize our algorithm to the episodic MDP setting and first achieve an additive dependence on the corruption level $ζ$ in the scenario of general function approximation. Notably, our algorithms achieve regret bounds either nearly match the performance lower bound or improve the existing methods for all the corruption levels and in both known and unknown $ζ$ cases.

Wei Xiong, Han Zhong, Chengshuai Shi et al.

Existing studies on provably efficient algorithms for Markov games (MGs) almost exclusively build on the "optimism in the face of uncertainty" (OFU) principle. This work focuses on a different approach of posterior sampling, which is celebrated in many bandits and reinforcement learning settings but remains under-explored for MGs. Specifically, for episodic two-player zero-sum MGs, a novel posterior sampling algorithm is developed with general function approximation. Theoretical analysis demonstrates that the posterior sampling algorithm admits a $\sqrt{T}$-regret bound for problems with a low multi-agent decoupling coefficient, which is a new complexity measure for MGs, where $T$ denotes the number of episodes. When specialized to linear MGs, the obtained regret bound matches the state-of-the-art results. To the best of our knowledge, this is the first provably efficient posterior sampling algorithm for MGs with frequentist regret guarantees, which enriches the toolbox for MGs and promotes the broad applicability of posterior sampling.

Wei Xiong, Chengshuai Shi, Jiaming Shen et al.

Recent studies have shown that large language models' (LLMs) mathematical problem-solving capabilities can be enhanced by integrating external tools, such as code interpreters, and employing multi-turn Chain-of-Thought (CoT) reasoning. While current methods focus on synthetic data generation and Supervised Fine-Tuning (SFT), this paper studies the complementary direct preference learning approach to further improve model performance. However, existing direct preference learning algorithms are originally designed for the single-turn chat task, and do not fully address the complexities of multi-turn reasoning and external tool integration required for tool-integrated mathematical reasoning tasks. To fill in this gap, we introduce a multi-turn direct preference learning framework, tailored for this context, that leverages feedback from code interpreters and optimizes trajectory-level preferences. This framework includes multi-turn DPO and multi-turn KTO as specific implementations. The effectiveness of our framework is validated through training of various language models using an augmented prompt set from the GSM8K and MATH datasets. Our results demonstrate substantial improvements: a supervised fine-tuned Gemma-1.1-it-7B model's performance increased from 77.5% to 83.9% on GSM8K and from 46.1% to 51.2% on MATH. Similarly, a Gemma-2-it-9B model improved from 84.1% to 86.3% on GSM8K and from 51.0% to 54.5% on MATH.

Zhongyi Qiu, Kangyi Qiu, Hanjia Lyu et al.

Emojis have become an integral part of digital communication, enriching text by conveying emotions, tone, and intent. Existing emoji recommendation methods are primarily evaluated based on their ability to match the exact emoji a user chooses in the original text. However, they ignore the essence of users' behavior on social media in that each text can correspond to multiple reasonable emojis. To better assess a model's ability to align with such real-world emoji usage, we propose a new semantics preserving evaluation framework for emoji recommendation, which measures a model's ability to recommend emojis that maintain the semantic consistency with the user's text. To evaluate how well a model preserves semantics, we assess whether the predicted affective state, demographic profile, and attitudinal stance of the user remain unchanged. If these attributes are preserved, we consider the recommended emojis to have maintained the original semantics. The advanced abilities of Large Language Models (LLMs) in understanding and generating nuanced, contextually relevant output make them well-suited for handling the complexities of semantics preserving emoji recommendation. To this end, we construct a comprehensive benchmark to systematically assess the performance of six proprietary and open-source LLMs using different prompting techniques on our task. Our experiments demonstrate that GPT-4o outperforms other LLMs, achieving a semantics preservation score of 79.23%. Additionally, we conduct case studies to analyze model biases in downstream classification tasks and evaluate the diversity of the recommended emojis.

Tianqi Liu, Wei Xiong, Jie Ren et al.

Reward models (RMs) play a pivotal role in aligning large language models (LLMs) with human preferences. However, traditional RM training, which relies on response pairs tied to specific prompts, struggles to disentangle prompt-driven preferences from prompt-independent artifacts, such as response length and format. In this work, we expose a fundamental limitation of current RM training methods, where RMs fail to effectively distinguish between contextual signals and irrelevant artifacts when determining preferences. To address this, we introduce a causal framework that learns preferences independent of these artifacts and propose a novel data augmentation technique designed to eliminate them. Extensive experiments show that our approach successfully filters out undesirable artifacts, yielding a more robust reward model (RRM). Our RRM improves the performance of a pairwise reward model trained on Gemma-2-9b-it, on RewardBench, increasing accuracy from 80.61% to 84.15%. Additionally, we train two DPO policies using both the RM and RRM, demonstrating that the RRM significantly enhances DPO-aligned policies, improving MT-Bench scores from 7.27 to 8.31 and length-controlled win-rates in AlpacaEval-2 from 33.46% to 52.49%.

Jing Shi, Wei Xiong, Zhe Lin et al.

Recent advances in personalized image generation allow a pre-trained text-to-image model to learn a new concept from a set of images. However, existing personalization approaches usually require heavy test-time finetuning for each concept, which is time-consuming and difficult to scale. We propose InstantBooth, a novel approach built upon pre-trained text-to-image models that enables instant text-guided image personalization without any test-time finetuning. We achieve this with several major components. First, we learn the general concept of the input images by converting them to a textual token with a learnable image encoder. Second, to keep the fine details of the identity, we learn rich visual feature representation by introducing a few adapter layers to the pre-trained model. We train our components only on text-image pairs without using paired images of the same concept. Compared to test-time finetuning-based methods like DreamBooth and Textual-Inversion, our model can generate competitive results on unseen concepts concerning language-image alignment, image fidelity, and identity preservation while being 100 times faster.

Xuanchang Zhang, Wei Xiong, Lichang Chen et al.

In this paper, we study format biases in reinforcement learning from human feedback (RLHF). We observe that many widely-used preference models, including human evaluators, GPT-4, and top-ranking models on the RewardBench benchmark, exhibit strong biases towards specific format patterns, such as lists, links, bold text, and emojis. Furthermore, large language models (LLMs) can exploit these biases to achieve higher rankings on popular benchmarks like AlpacaEval and LMSYS Chatbot Arena. One notable example of this is verbosity bias, where current preference models favor longer responses that appear more comprehensive, even when their quality is equal to or lower than shorter, competing responses. However, format biases beyond verbosity remain largely underexplored in the literature. In this work, we extend the study of biases in preference learning beyond the commonly recognized length bias, offering a comprehensive analysis of a wider range of format biases. Additionally, we show that with a small amount of biased data (less than 1%), we can inject significant bias into the reward model. Moreover, these format biases can also be easily exploited by downstream alignment algorithms, such as best-of-n sampling and online iterative DPO, as it is usually easier to manipulate the format than to improve the quality of responses. Our findings emphasize the need to disentangle format and content both for designing alignment algorithms and evaluating models.

Wei Xiong, Xiaomeng Huang, Ziyang Zhang et al.

The lacking of analytic solutions of diverse partial differential equations (PDEs) gives birth to a series of computational techniques for numerical solutions. Although numerous latest advances are accomplished in developing neural operators, a kind of neural-network-based PDE solver, these solvers become less accurate and explainable while learning long-term behaviors of non-linear PDE families. In this paper, we propose the Koopman neural operator (KNO), a new neural operator, to overcome these challenges. With the same objective of learning an infinite-dimensional mapping between Banach spaces that serves as the solution operator of the target PDE family, our approach differs from existing models by formulating a non-linear dynamic system of equation solution. By approximating the Koopman operator, an infinite-dimensional operator governing all possible observations of the dynamic system, to act on the flow mapping of the dynamic system, we can equivalently learn the solution of a non-linear PDE family by solving simple linear prediction problems. We validate the KNO in mesh-independent, long-term, and5zero-shot predictions on five representative PDEs (e.g., the Navier-Stokes equation and the Rayleigh-B{é}nard convection) and three real dynamic systems (e.g., global water vapor patterns and western boundary currents). In these experiments, the KNO exhibits notable advantages compared with previous state-of-the-art models, suggesting the potential of the KNO in supporting diverse science and engineering applications (e.g., PDE solving, turbulence modelling, and precipitation forecasting).

Wei Xiong, Yanfei Xiang, Hao Wu et al.

Ocean modeling is a powerful tool for simulating the physical, chemical, and biological processes of the ocean, which is the foundation for marine science research and operational oceanography. Modern numerical ocean modeling mainly consists of governing equations and numerical algorithms. Nonlinear instability, computational expense, low reusability efficiency and high coupling costs have gradually become the main bottlenecks for the further development of numerical ocean modeling. Recently, artificial intelligence-based modeling in scientific computing has shown revolutionary potential for digital twins and scientific simulations, but the bottlenecks of numerical ocean modeling have not been further solved. Here, we present AI-GOMS, a large AI-driven global ocean modeling system, for accurate and efficient global ocean daily prediction. AI-GOMS consists of a backbone model with the Fourier-based Masked Autoencoder structure for basic ocean variable prediction and lightweight fine-tuning models incorporating regional downscaling, wave decoding, and biochemistry coupling modules. AI-GOMS has achieved the best performance in 30 days of prediction for the global ocean basic variables with 15 depth layers at 1/4° spatial resolution. Beyond the good performance in statistical metrics, AI-GOMS realizes the simulation of mesoscale eddies in the Kuroshio region at 1/12° spatial resolution and ocean stratification in the tropical Pacific Ocean. AI-GOMS provides a new backbone-downstream paradigm for Earth system modeling, which makes the system transferable, scalable and reusable.

Hanze Dong, Wei Xiong, Bo Pang et al.

We present the workflow of Online Iterative Reinforcement Learning from Human Feedback (RLHF) in this technical report, which is widely reported to outperform its offline counterpart by a large margin in the recent large language model (LLM) literature. However, existing open-source RLHF projects are still largely confined to the offline learning setting. In this technical report, we aim to fill in this gap and provide a detailed recipe that is easy to reproduce for online iterative RLHF. In particular, since online human feedback is usually infeasible for open-source communities with limited resources, we start by constructing preference models using a diverse set of open-source datasets and use the constructed proxy preference model to approximate human feedback. Then, we discuss the theoretical insights and algorithmic principles behind online iterative RLHF, followed by a detailed practical implementation. Our trained LLM achieves impressive performance on LLM chatbot benchmarks, including AlpacaEval-2, Arena-Hard, and MT-Bench, as well as other academic benchmarks such as HumanEval and TruthfulQA. We have shown that supervised fine-tuning (SFT) and iterative RLHF can obtain state-of-the-art performance with fully open-source datasets. Further, we have made our models, curated datasets, and comprehensive step-by-step code guidebooks publicly available. Please refer to https://github.com/RLHFlow/RLHF-Reward-Modeling and https://github.com/RLHFlow/Online-RLHF for more detailed information.

Chengshuai Shi, Wei Xiong, Cong Shen et al.

Existing theoretical studies on offline reinforcement learning (RL) mostly consider a dataset sampled directly from the target task. In practice, however, data often come from several heterogeneous but related sources. Motivated by this gap, this work aims at rigorously understanding offline RL with multiple datasets that are collected from randomly perturbed versions of the target task instead of from itself. An information-theoretic lower bound is derived, which reveals a necessary requirement on the number of involved sources in addition to that on the number of data samples. Then, a novel HetPEVI algorithm is proposed, which simultaneously considers the sample uncertainties from a finite number of data samples per data source and the source uncertainties due to a finite number of available data sources. Theoretical analyses demonstrate that HetPEVI can solve the target task as long as the data sources collectively provide a good data coverage. Moreover, HetPEVI is demonstrated to be optimal up to a polynomial factor of the horizon length. Finally, the study is extended to offline Markov games and offline robust RL, which demonstrates the generality of the proposed designs and theoretical analyses.

Yi Huang, Jiancheng Huang, Yifan Liu et al.

Denoising diffusion models have emerged as a powerful tool for various image generation and editing tasks, facilitating the synthesis of visual content in an unconditional or input-conditional manner. The core idea behind them is learning to reverse the process of gradually adding noise to images, allowing them to generate high-quality samples from a complex distribution. In this survey, we provide an exhaustive overview of existing methods using diffusion models for image editing, covering both theoretical and practical aspects in the field. We delve into a thorough analysis and categorization of these works from multiple perspectives, including learning strategies, user-input conditions, and the array of specific editing tasks that can be accomplished. In addition, we pay special attention to image inpainting and outpainting, and explore both earlier traditional context-driven and current multimodal conditional methods, offering a comprehensive analysis of their methodologies. To further evaluate the performance of text-guided image editing algorithms, we propose a systematic benchmark, EditEval, featuring an innovative metric, LMM Score. Finally, we address current limitations and envision some potential directions for future research. The accompanying repository is released at https://github.com/SiatMMLab/Awesome-Diffusion-Model-Based-Image-Editing-Methods.

Wei Xiong, Muyuan Ma, Xiaomeng Huang et al.

Numerous physics theories are rooted in partial differential equations (PDEs). However, the increasingly intricate physics equations, especially those that lack analytic solutions or closed forms, have impeded the further development of physics. Computationally solving PDEs by classic numerical approaches suffers from the trade-off between accuracy and efficiency and is not applicable to the empirical data generated by unknown latent PDEs. To overcome this challenge, we present KoopmanLab, an efficient module of the Koopman neural operator family, for learning PDEs without analytic solutions or closed forms. Our module consists of multiple variants of the Koopman neural operator (KNO), a kind of mesh-independent neural-network-based PDE solvers developed following dynamic system theory. The compact variants of KNO can accurately solve PDEs with small model sizes while the large variants of KNO are more competitive in predicting highly complicated dynamic systems govern by unknown, high-dimensional, and non-linear PDEs. All variants are validated by mesh-independent and long-term prediction experiments implemented on representative PDEs (e.g., the Navier-Stokes equation and the Bateman-Burgers equation in fluid mechanics) and ERA5 (i.e., one of the largest high-resolution global-scale climate data sets in earth physics). These demonstrations suggest the potential of KoopmanLab to be a fundamental tool in diverse physics studies related to equations or dynamic systems.

Han Zhong, Zikang Shan, Guhao Feng et al. · pku

In the classical Reinforcement Learning from Human Feedback (RLHF) framework, Proximal Policy Optimization (PPO) is employed to learn from sparse, sentence-level rewards -- a challenging scenario in traditional deep reinforcement learning. Despite the great successes of PPO in the alignment of large language models, its open-source implementation is still largely sub-optimal. To address these issues, we introduce a framework that models RLHF problems as a Markov decision process (MDP), enabling the capture of fine-grained token-wise information. Under this framework, we introduce an algorithm Reinforced Token Optimization (\texttt{RTO}), which learns the token-wise reward function from preference data and performs policy optimization based on this learned token-wise reward signal. Theoretically, \texttt{RTO} is proven to have the capability of finding the near-optimal policy sample-efficiently. For its practical implementation, \texttt{RTO} innovatively integrates Direct Preference Optimization (DPO) and PPO. DPO, originally derived from sparse sentence rewards, surprisingly provides us with a token-wise characterization of response quality, which is seamlessly incorporated into our subsequent PPO training stage. Extensive experiments demonstrate that \texttt{RTO} performs better than PPO and other direct preference learning algorithms. In particular, RTO outperforms PPO by 7.5 points on the AlpacaEval 2 benchmark and by 4.1 points on Arena-Hard. Our code and models are available at \href{https://github.com/zkshan2002/RTO}{https://github.com/zkshan2002/RTO}.

Yuheng Zhang, Chenlu Ye, Shuowei Jin et al.

Reinforcement learning, including reinforcement learning with verifiable rewards (RLVR), has emerged as a powerful approach for LLM post-training. Central to these approaches is the design of the importance sampling (IS) ratio used in off-policy policy-gradient estimation. Existing methods face a fundamental bias-variance dilemma: token-level IS ratios, as adopted by PPO (Schulman et al., 2017) and GRPO (Shao et al., 2024), introduce bias by ignoring prefix state distribution mismatch; full sequence ratios provide exact trajectory-level correction but suffer from high variance due to the multiplicative accumulation of per-token ratios, while GSPO (Zheng et al., 2025) improves numerical stability via length normalization at the cost of deviating from the exact full-sequence IS correction. In this work, we identify the cumulative token IS ratio, the product of per-token ratios up to position $t$, as a theoretically principled solution to this dilemma. We prove that, under the token-level policy-gradient formulation, this ratio provides an unbiased prefix correction for each token-level gradient term and has strictly lower variance than the full sequence ratio. Building on this insight, we propose CTPO (Cumulative Token Policy Optimization), which combines the cumulative token IS ratio with position-adaptive clipping that scales log-space clip bounds according to the natural $\sqrt{t}$ growth of the cumulative log-ratio. This yields more consistent regularization across token positions. We implement and evaluate CTPO in the tool-integrated reasoning setting on several challenging mathematical reasoning benchmarks, achieving the best average performance across both model scales compared with strong GRPO and GSPO baselines. Code will be available at https://github.com/horizon-llm/CTPO.

Zhexiao Xiong, Wei Xiong, Jing Shi et al.

Recent approaches in text-to-image customization have primarily focused on preserving the identity of the input subject, but often fail to control the spatial location and size of objects. We introduce GroundingBooth, which achieves zero-shot, instance-level spatial grounding on both foreground subjects and background objects in the text-to-image customization task. Our proposed grounding module and subject-grounded cross-attention layer enable the creation of personalized images with accurate layout alignment, identity preservation, and strong text-image coherence. In addition, our model seamlessly supports personalization with multiple subjects. Our model shows strong results in both layout-guided image synthesis and text-to-image customization tasks. The project page is available at https://groundingbooth.github.io.

Hao Wu, Yuxuan Liang, Wei Xiong et al.

Efficiently modeling spatio-temporal (ST) physical processes and observations presents a challenging problem for the deep learning community. Many recent studies have concentrated on meticulously reconciling various advantages, leading to designed models that are neither simple nor practical. To address this issue, this paper presents a systematic study on existing shortcomings faced by off-the-shelf models, including lack of local fidelity, poor prediction performance over long time-steps,low scalability, and inefficiency. To systematically address the aforementioned problems, we propose an EarthFarseer, a concise framework that combines parallel local convolutions and global Fourier-based transformer architectures, enabling dynamically capture the local-global spatial interactions and dependencies. EarthFarseer also incorporates a multi-scale fully convolutional and Fourier architectures to efficiently and effectively capture the temporal evolution. Our proposal demonstrates strong adaptability across various tasks and datasets, with fast convergence and better local fidelity in long time-steps predictions. Extensive experiments and visualizations over eight human society physical and natural physical datasets demonstrates the state-of-the-art performance of EarthFarseer. We release our code at https://github.com/easylearningscores/EarthFarseer.

Xudong Xie, Yuzhe Li, Yang Liu et al.

Accurate text segmentation results are crucial for text-related generative tasks, such as text image generation, text editing, text removal, and text style transfer. Recently, some scene text segmentation methods have made significant progress in segmenting regular text. However, these methods perform poorly in scenarios containing artistic text. Therefore, this paper focuses on the more challenging task of artistic text segmentation and constructs a real artistic text segmentation dataset. One challenge of the task is that the local stroke shapes of artistic text are changeable with diversity and complexity. We propose a decoder with the layer-wise momentum query to prevent the model from ignoring stroke regions of special shapes. Another challenge is the complexity of the global topological structure. We further design a skeleton-assisted head to guide the model to focus on the global structure. Additionally, to enhance the generalization performance of the text segmentation model, we propose a strategy for training data synthesis, based on the large multi-modal model and the diffusion model. Experimental results show that our proposed method and synthetic dataset can significantly enhance the performance of artistic text segmentation and achieve state-of-the-art results on other public datasets.

Chi-Jui Ho, Harsh Bhakta, Wei Xiong et al.

High-dimensional hyperspectral imaging (HSI) enables the visualization of ultrafast molecular dynamics and complex, heterogeneous spectra. However, applying this capability to resolve spatially varying vibrational couplings in two-dimensional infrared (2DIR) spectroscopy, a type of coherent multidimensional spectroscopy (CMDS), necessitates prohibitively long data acquisition, driven by dense Nyquist sampling requirements and the need for extensive signal accumulation. To address this challenge, we introduce a physics-informed neural representation approach that efficiently reconstructs dense spatially-resolved 2DIR hyperspectral images from sparse experimental measurements. In particular, we used a multilayer perceptron (MLP) to model the relationship between the sub-sampled 4D coordinates and their corresponding spectral intensities, and recover densely sampled 4D spectra from limited observations. The reconstruction results demonstrate that our method, using a fraction of the samples, faithfully recovers both oscillatory and non-oscillatory spectral dynamics in experimental measurement. Moreover, we develop a loss-aware adaptive sampling method to progressively introduce potentially informative samples for iterative data collection while conducting experiments. Experimental results show that the proposed approach achieves high-fidelity spectral recovery using only $1/32$ of the sampling budget, as opposed to exhaustive sampling, effectively reducing total experiment time by up to 32-fold. This framework offers a scalable solution for accelerating any experiments with hypercube data, including multidimensional spectroscopy and hyperspectral imaging, paving the way for rapid chemical imaging of transient biological and material systems.

Yongsheng Yu, Ziyun Zeng, Zhiyuan Xiao et al.

Recent video editing models have converged on a unified conditioning design: a single diffusion transformer jointly consumes text, source video, and reference images, and one set of weights covers replacement, removal, style transfer, and reference-driven insertion. The design is flexible, but it assumes that the user already provides model-ready text, reference images, and spatial grounding for local edits, which real requests often omit. We present Aurora, an agentic video editing framework that pairs a tool-augmented vision-language model (VLM) agent with a unified video diffusion transformer. The VLM agent maps a raw user request to a structured edit plan aligned with the transformer's conditioning channels, thereby resolving textual and visual underspecification before generation. We train the VLM agent with supervised data for complete edit planning and reference-image selection, together with preference pairs for robust tool use and instruction refinement. We introduce AgentEdit-Bench to evaluate agent-enhanced video editing under textual and visual underspecification. Experiments on AgentEdit-Bench and two existing video editing benchmarks show that Aurora improves over instruction-only baselines and that the VLM agent transfers to compatible frozen video editing models. Project page: https://yeates.github.io/Aurora-Page

Xuewei Li, Zewen Shang, Zhiqiang Liu et al.

Wind speed prediction is critical to the management of wind power generation. Due to the large range of wind speed fluctuations and wake effect, there may also be strong correlations between long-distance wind turbines. This difficult-to-extract feature has become a bottleneck for improving accuracy. History and future time information includes the trend of airflow changes, whether this dynamic information can be utilized will also affect the prediction effect. In response to the above problems, this paper proposes Windformer. First, Windformer divides the wind turbine cluster into multiple non-overlapping windows and calculates correlations inside the windows, then shifts the windows partially to provide connectivity between windows, and finally fuses multi-channel features based on detailed and global information. To dynamically model the change process of wind speed, this paper extracts time series in both history and future directions simultaneously. Compared with other current-advanced methods, the Mean Square Error (MSE) of Windformer is reduced by 0.5\% to 15\% on two datasets from NERL.

Jiarui Yao, Yifan Hao, Hanning Zhang et al.

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.

Zhexiao Xiong, Yizhi Song, Liu He et al.

Video Diffusion Models (VDMs) offer a promising approach for simulating dynamic scenes and environments, with broad applications in robotics and media generation. However, existing models often generate temporally incoherent content that violates basic physical intuition, significantly limiting their practical applicability. We propose PhysAlign, an efficient framework for physics-coherent image-to-video (I2V) generation that explicitly addresses this limitation. To overcome the critical scarcity of physics-annotated videos, we first construct a fully controllable synthetic data generation pipeline based on rigid-body simulation, yielding a highly-curated dataset with accurate, fine-grained physics and 3D annotations. Leveraging this data, PhysAlign constructs a unified physical latent space by coupling explicit 3D geometry constraints with a Gram-based spatio-temporal relational alignment that extracts kinematic priors from video foundation models. Extensive experiments demonstrate that PhysAlign significantly outperforms existing VDMs on tasks requiring complex physical reasoning and temporal stability, without compromising zero-shot visual quality. PhysAlign shows the potential to bridge the gap between raw visual synthesis and rigid-body kinematics, establishing a practical paradigm for genuinely physics-grounded video generation. The project page is available at https://physalign.github.io/PhysAlign.

Wei Xiong, Jiangtong Li, Jie Li et al.

Electroencephalography (EEG) foundation models are poised to significantly advance brain signal analysis by learning robust representations from large-scale, unlabeled datasets. However, their rapid proliferation has outpaced the development of standardized evaluation benchmarks, which complicates direct model comparisons and hinders systematic scientific progress. This fragmentation fosters scientific inefficiency and obscures genuine architectural advancements. To address this critical gap, we introduce EEG-FM-Bench, the first comprehensive benchmark for the systematic and standardized evaluation of EEG foundation models (EEG-FMs). Our contributions are threefold: (1) we curate a diverse suite of downstream tasks and datasets from canonical EEG paradigms, implementing standardized processing and evaluation protocols within a unified open-source framework; (2) we benchmark prominent state-of-the-art foundation models to establish comprehensive baseline results for a clear comparison of the current landscape; (3) we perform qualitative analyses of the learned representations to provide insights into model behavior and inform future architectural design. Through extensive experiments, we find that fine-grained spatio-temporal feature interaction, multitask unified training and neuropsychological priors would contribute to enhancing model performance and generalization capabilities. By offering a unified platform for fair comparison and reproducible research, EEG-FM-Bench seeks to catalyze progress and guide the community toward the development of more robust and generalizable EEG-FMs. Code is released at https://github.com/xw1216/EEG-FM-Bench.

Wei Xiong, Junming Lin, Jiangtong Li et al.

While foundation models excel in text, image, and video domains, the critical biological signals, particularly electroencephalography(EEG), remain underexplored. EEG benefits neurological research with its high temporal resolution, operational practicality, and safety profile. However, low signal-to-noise ratio, inter-subject variability, and cross-paradigm differences hinder the generalization of current models. Existing methods often employ simplified strategies, such as a single loss function or a channel-temporal joint representation module, and suffer from a domain gap between pretraining and evaluation tasks that compromises efficiency and adaptability. To address these limitations, we propose the Adaptive Large Foundation model for EEG signal representation(ALFEE) framework, a novel hybrid transformer architecture with two learning stages for robust EEG representation learning. ALFEE employs a hybrid attention that separates channel-wise feature aggregation from temporal dynamics modeling, enabling robust EEG representation with variable channel configurations. A channel encoder adaptively compresses variable channel information, a temporal encoder captures task-guided evolution, and a hybrid decoder reconstructs signals in both temporal and frequency domains. During pretraining, ALFEE optimizes task prediction, channel and temporal mask reconstruction, and temporal forecasting to enhance multi-scale and multi-channel representation. During fine-tuning, a full-model adaptation with a task-specific token dictionary and a cross-attention layer boosts performance across multiple tasks. After 25,000 hours of pretraining, extensive experimental results on six downstream EEG tasks demonstrate the superior performance of ALFEE over existing models. Our ALFEE framework establishes a scalable foundation for biological signal analysis with implementation at https://github.com/xw1216/ALFEE.

Wei Xiong, Hanze Dong, Chenlu Ye et al.

This paper studies the alignment process of generative models with Reinforcement Learning from Human Feedback (RLHF). We first identify the primary challenges of existing popular methods like offline PPO and offline DPO as lacking in strategical exploration of the environment. Then, to understand the mathematical principle of RLHF, we consider a standard mathematical formulation, the reverse-KL regularized contextual bandit for RLHF. Despite its widespread practical application, a rigorous theoretical analysis of this formulation remains open. We investigate its behavior in three distinct settings -- offline, online, and hybrid -- and propose efficient algorithms with finite-sample theoretical guarantees. Moving towards practical applications, our framework, with a robust approximation of the information-theoretical policy improvement oracle, naturally gives rise to several novel RLHF algorithms. This includes an iterative version of the Direct Preference Optimization (DPO) algorithm for online settings, and a multi-step rejection sampling strategy for offline scenarios. Our empirical evaluations on real-world alignment experiment of large language model demonstrate that these proposed methods significantly surpass existing strong baselines, such as DPO and Rejection Sampling Optimization (RSO), showcasing the connections between solid theoretical foundations and their potent practical implementations.

Sonakshi Gupta, Akhlak Mahmood, Wei Xiong et al.

Polymer literature contains a large and growing body of experimental knowledge, yet much of it is buried in unstructured text and inconsistent terminology, making systematic retrieval and reasoning difficult. Existing tools typically extract narrow, study-specific facts in isolation, failing to preserve the cross-study context required to answer broader scientific questions. Retrieval-augmented generation (RAG) offers a promising way to overcome this limitation by combining large language models (LLMs) with external retrieval, but its effectiveness depends strongly on how domain knowledge is represented. In this work, we develop two retrieval pipelines: a dense semantic vector-based approach (VectorRAG) and a graph-based approach (GraphRAG). Using over 1,000 polyhydroxyalkanoate (PHA) papers, we construct context-preserving paragraph embeddings and a canonicalized structured knowledge graph supporting entity disambiguation and multi-hop reasoning. We evaluate these pipelines through standard retrieval metrics, comparisons with general state-of-the-art systems such as GPT and Gemini, and qualitative validation by a domain chemist. The results show that GraphRAG achieves higher precision and interpretability, while VectorRAG provides broader recall, highlighting complementary trade-offs. Expert validation further confirms that the tailored pipelines, particularly GraphRAG, produce well-grounded, citation-reliable responses with strong domain relevance. By grounding every statement in evidence, these systems enable researchers to navigate the literature, compare findings across studies, and uncover patterns that are difficult to extract manually. More broadly, this work establishes a practical framework for building materials science assistants using curated corpora and retrieval design, reducing reliance on proprietary models while enabling trustworthy literature analysis at scale.

Wei Xiong, Chenlu Ye, Baohao Liao et al.

Reinforcement learning applied to large language models (LLMs) for reasoning tasks is often bottlenecked by unstable gradient estimates due to fixed and uniform sampling of responses across prompts. Prior work such as GVM-RAFT addresses this by dynamically allocating inference budget per prompt to minimize stochastic gradient variance under a budget constraint. Inspired by this insight, we propose Reinforce-Ada, an adaptive sampling framework for online RL post-training of LLMs that continuously reallocates sampling effort to the prompts with the greatest uncertainty or learning potential. Unlike conventional two-stage allocation methods, Reinforce-Ada interleaves estimation and sampling in an online successive elimination process, and automatically stops sampling for a prompt once sufficient signal is collected. To stabilize updates, we form fixed-size groups with enforced reward diversity and compute advantage baselines using global statistics aggregated over the adaptive sampling phase. Empirical results across multiple model architectures and reasoning benchmarks show that Reinforce-Ada accelerates convergence and improves final performance compared to GRPO, especially when using the balanced sampling variant. Our work highlights the central role of variance-aware, adaptive data curation in enabling efficient and reliable reinforcement learning for reasoning-capable LLMs. Code is available at https://github.com/RLHFlow/Reinforce-Ada.

Kaizhen Sun, Jinghang Li, Kuan Dai et al.

Time-to-Collision (TTC) estimation lies in the core of the forward collision warning (FCW) functionality, which is key to all Automatic Emergency Braking (AEB) systems. Although the success of solutions using frame-based cameras (e.g., Mobileye's solutions) has been witnessed in normal situations, some extreme cases, such as the sudden variation in the relative speed of leading vehicles and the sudden appearance of pedestrians, still pose significant risks that cannot be handled. This is due to the inherent imaging principles of frame-based cameras, where the time interval between adjacent exposures introduces considerable system latency to AEB. Event cameras, as a novel bio-inspired sensor, offer ultra-high temporal resolution and can asynchronously report brightness changes at the microsecond level. To explore the potential of event cameras in the above-mentioned challenging cases, we propose EvTTC, which is, to the best of our knowledge, the first multi-sensor dataset focusing on TTC tasks under high-relative-speed scenarios. EvTTC consists of data collected using standard cameras and event cameras, covering various potential collision scenarios in daily driving and involving multiple collision objects. Additionally, LiDAR and GNSS/INS measurements are provided for the calculation of ground-truth TTC. Considering the high cost of testing TTC algorithms on full-scale mobile platforms, we also provide a small-scale TTC testbed for experimental validation and data augmentation. All the data and the design of the testbed are open sourced, and they can serve as a benchmark that will facilitate the development of vision-based TTC techniques.

Xuange Cui, Wei Xiong, Songlin Wang

In this paper, we propose a robust multilingual model to improve the quality of search results. Our model not only leverage the processed class-balanced dataset, but also benefit from multitask pre-training that leads to more general representations. In pre-training stage, we adopt mlm task, classification task and contrastive learning task to achieve considerably performance. In fine-tuning stage, we use confident learning, exponential moving average method (EMA), adversarial training (FGM) and regularized dropout strategy (R-Drop) to improve the model's generalization and robustness. Moreover, we use a multi-granular semantic unit to discover the queries and products textual metadata for enhancing the representation of the model. Our approach obtained competitive results and ranked top-8 in three tasks. We release the source code and pre-trained models associated with this work.

Haoxiang Wang, Yong Lin, Wei Xiong et al.

Fine-grained control over large language models (LLMs) remains a significant challenge, hindering their adaptability to diverse user needs. While Reinforcement Learning from Human Feedback (RLHF) shows promise in aligning LLMs, its reliance on scalar rewards often limits its ability to capture diverse user preferences in real-world applications. To address this limitation, we introduce the Directional Preference Alignment (DPA) framework. Unlike the scalar-reward RLHF, DPA incorporates multi-objective reward modeling to represent diverse preference profiles. Additionally, DPA models user preferences as directions (i.e., unit vectors) in the reward space to achieve user-dependent preference control. Our method involves training a multi-objective reward model and then fine-tuning the LLM with a preference-conditioned variant of Rejection Sampling Finetuning (RSF), an RLHF method adopted by Llama 2. This method enjoys a better performance trade-off across various reward objectives. In comparison with the scalar-reward RLHF, DPA offers users intuitive control over LLM generation: they can arithmetically specify their desired trade-offs (e.g., more helpfulness with less verbosity). We also validate the effectiveness of DPA with real-world alignment experiments on Mistral-7B. Our method provides straightforward arithmetic control over the trade-off between helpfulness and verbosity while maintaining competitive performance with strong baselines such as Direct Preference Optimization (DPO).

Wei Xiong, Jiarui Yao, Yuhui Xu et al. · salesforce

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.

Renjie Pi, Tianyang Han, Wei Xiong et al.

Multimodal Large Language Models (MLLMs) excel in generating responses based on visual inputs. However, they often suffer from a bias towards generating responses similar to their pretraining corpus, overshadowing the importance of visual information. We treat this bias as a "preference" for pretraining statistics, which hinders the model's grounding in visual input. To mitigate this issue, we propose Bootstrapped Preference Optimization (BPO), which conducts preference learning with datasets containing negative responses bootstrapped from the model itself. Specifically, we propose the following two strategies: 1) using distorted image inputs to the MLLM for eliciting responses that contain signified pretraining bias; 2) leveraging text-based LLM to explicitly inject erroneous but common elements into the original response. Those undesirable responses are paired with original annotated responses from the datasets to construct the preference dataset, which is subsequently utilized to perform preference learning. Our approach effectively suppresses pretrained LLM bias, enabling enhanced grounding in visual inputs. Extensive experimentation demonstrates significant performance improvements across multiple benchmarks, advancing the state-of-the-art in multimodal conversational systems.

Yizhi Song, Zhifei Zhang, Zhe Lin et al.

Generative object compositing emerges as a promising new avenue for compositional image editing. However, the requirement of object identity preservation poses a significant challenge, limiting practical usage of most existing methods. In response, this paper introduces IMPRINT, a novel diffusion-based generative model trained with a two-stage learning framework that decouples learning of identity preservation from that of compositing. The first stage is targeted for context-agnostic, identity-preserving pretraining of the object encoder, enabling the encoder to learn an embedding that is both view-invariant and conducive to enhanced detail preservation. The subsequent stage leverages this representation to learn seamless harmonization of the object composited to the background. In addition, IMPRINT incorporates a shape-guidance mechanism offering user-directed control over the compositing process. Extensive experiments demonstrate that IMPRINT significantly outperforms existing methods and various baselines on identity preservation and composition quality.

Mengwei Ren, Wei Xiong, Jae Shin Yoon et al.

Portrait harmonization aims to composite a subject into a new background, adjusting its lighting and color to ensure harmony with the background scene. Existing harmonization techniques often only focus on adjusting the global color and brightness of the foreground and ignore crucial illumination cues from the background such as apparent lighting direction, leading to unrealistic compositions. We introduce Relightful Harmonization, a lighting-aware diffusion model designed to seamlessly harmonize sophisticated lighting effect for the foreground portrait using any background image. Our approach unfolds in three stages. First, we introduce a lighting representation module that allows our diffusion model to encode lighting information from target image background. Second, we introduce an alignment network that aligns lighting features learned from image background with lighting features learned from panorama environment maps, which is a complete representation for scene illumination. Last, to further boost the photorealism of the proposed method, we introduce a novel data simulation pipeline that generates synthetic training pairs from a diverse range of natural images, which are used to refine the model. Our method outperforms existing benchmarks in visual fidelity and lighting coherence, showing superior generalization in real-world testing scenarios, highlighting its versatility and practicality.

Wei Xiong, Hanning Zhang, Chenlu Ye et al.

We study self-rewarding reasoning large language models (LLMs), which can simultaneously generate step-by-step reasoning and evaluate the correctness of their outputs during the inference time-without external feedback. This integrated approach allows a single model to independently guide its reasoning process, offering computational advantages for model deployment. We particularly focus on the representative task of self-correction, where models autonomously detect errors in their responses, revise outputs, and decide when to terminate iterative refinement loops. To enable this, we propose a two-staged algorithmic framework for constructing self-rewarding reasoning models using only self-generated data. In the first stage, we employ sequential rejection sampling to synthesize long chain-of-thought trajectories that incorporate both self-rewarding and self-correction mechanisms. Fine-tuning models on these curated data allows them to learn the patterns of self-rewarding and self-correction. In the second stage, we further enhance the models' ability to assess response accuracy and refine outputs through reinforcement learning with rule-based signals. Experiments with Llama-3 and Qwen-2.5 demonstrate that our approach surpasses intrinsic self-correction capabilities and achieves performance comparable to systems that rely on external reward models.

Chenlu Ye, Wei Xiong, Yuheng Zhang et al.

We investigate Reinforcement Learning from Human Feedback (RLHF) in the context of a general preference oracle. In particular, we do not assume the existence of a reward function and an oracle preference signal drawn from the Bradley-Terry model as most of the prior works do. We consider a standard mathematical formulation, the reverse-KL regularized minimax game between two LLMs for RLHF under general preference oracle. The learning objective of this formulation is to find a policy so that it is consistently preferred by the KL-regularized preference oracle over any competing LLMs. We show that this framework is strictly more general than the reward-based one, and propose sample-efficient algorithms for both the offline learning from a pre-collected preference dataset and online learning where we can query the preference oracle along the way of training. Empirical studies verify the effectiveness of the proposed framework.

Rijie Xi, Weikang Xu, Wei Xiong et al.

Accurately reconstructing Gene Regulatory Networks (GRNs) is crucial for understanding gene functions and disease mechanisms. Single-cell RNA sequencing (scRNA-seq) technology provides vast data for computational GRN reconstruction. Since GRNs are ideally modeled as signed directed graphs to capture activation/inhibition relationships, the most intuitive and reasonable approach is to design feature extractors based on the topological structure of GRNs to extract structural features, then combine them with biological characteristics for research. However, traditional spectral graph convolution struggles with this representation. Thus, we propose MSGRNLink, a novel framework that explicitly models GRNs as signed directed graphs and employs magnetic signed Laplacian convolution. Experiments across simulated and real datasets demonstrate that MSGRNLink outperforms all baseline models in AUROC. Parameter sensitivity analysis and ablation studies confirmed its robustness and the importance of each module. In a bladder cancer case study, MSGRNLink predicted more known edges and edge signs than benchmark models, further validating its biological relevance.

Wei Xiong, Wenting Zhao, Weizhe Yuan et al.

As models increasingly leverage multi-step reasoning strategies to solve complex problems, supervising the logical validity of these intermediate steps has become a critical research challenge. Process reward models address this by providing step-by-step feedback, but current approaches have two major drawbacks: they typically function as classifiers without providing explanations, and their reliance on supervised fine-tuning with static datasets limits generalization. Inspired by recent advances, we reframe stepwise reward modeling from a classification task to a reasoning task itself. We thus propose a generative judge that reasons about the policy model's reasoning steps (i.e., meta-reasons), outputting thinking tokens before delivering a final verdict. Our model, StepWiser, is trained by reinforcement learning using relative outcomes of rollouts. We show it provides (i) better judgment accuracy on intermediate steps than existing methods; (ii) can be used to improve the policy model at training time; and (iii) improves inference-time search.

Jing Gu, Nanxuan Zhao, Wei Xiong et al.

Effective editing of personal content holds a pivotal role in enabling individuals to express their creativity, weaving captivating narratives within their visual stories, and elevate the overall quality and impact of their visual content. Therefore, in this work, we introduce SwapAnything, a novel framework that can swap any objects in an image with personalized concepts given by the reference, while keeping the context unchanged. Compared with existing methods for personalized subject swapping, SwapAnything has three unique advantages: (1) precise control of arbitrary objects and parts rather than the main subject, (2) more faithful preservation of context pixels, (3) better adaptation of the personalized concept to the image. First, we propose targeted variable swapping to apply region control over latent feature maps and swap masked variables for faithful context preservation and initial semantic concept swapping. Then, we introduce appearance adaptation, to seamlessly adapt the semantic concept into the original image in terms of target location, shape, style, and content during the image generation process. Extensive results on both human and automatic evaluation demonstrate significant improvements of our approach over baseline methods on personalized swapping. Furthermore, SwapAnything shows its precise and faithful swapping abilities across single object, multiple objects, partial object, and cross-domain swapping tasks. SwapAnything also achieves great performance on text-based swapping and tasks beyond swapping such as object insertion.

Zhongyi Qiu, Hanjia Lyu, Wei Xiong et al.

Social media enables dynamic user engagement with trending topics, and recent research has explored the potential of large language models (LLMs) for response generation. While some studies investigate LLMs as agents for simulating user behavior on social media, their focus remains on practical viability and scalability rather than a deeper understanding of how well LLM aligns with human behavior. This paper analyzes LLMs' ability to simulate social media engagement through action guided response generation, where a model first predicts a user's most likely engagement action-retweet, quote, or rewrite-towards a trending post before generating a personalized response conditioned on the predicted action. We benchmark GPT-4o-mini, O1-mini, and DeepSeek-R1 in social media engagement simulation regarding a major societal event discussed on X. Our findings reveal that zero-shot LLMs underperform BERT in action prediction, while few-shot prompting initially degrades the prediction accuracy of LLMs with limited examples. However, in response generation, few-shot LLMs achieve stronger semantic alignment with ground truth posts.

Heyang Zhao, Chenlu Ye, Wei Xiong et al.

Recent advances in Reinforcement Learning from Human Feedback (RLHF) have shown that KL-regularization plays a pivotal role in improving the efficiency of RL fine-tuning for large language models (LLMs). Despite its empirical advantage, the theoretical difference between KL-regularized RL and standard RL remains largely under-explored. While there is a recent line of work on the theoretical analysis of KL-regularized objective in decision making \citep{xiong2024iterative, xie2024exploratory,zhao2024sharp}, these analyses either reduce to the traditional RL setting or rely on strong coverage assumptions. In this paper, we propose an optimism-based KL-regularized online contextual bandit algorithm, and provide a novel analysis of its regret. By carefully leveraging the benign optimization landscape induced by the KL-regularization and the optimistic reward estimation, our algorithm achieves an $\mathcal{O}\big(η\log (N_{\mathcal R} T)\cdot d_{\mathcal R}\big)$ logarithmic regret bound, where $η, N_{\mathcal R},T,d_{\mathcal R}$ denote the KL-regularization parameter, the cardinality of the reward function class, number of rounds, and the complexity of the reward function class. Furthermore, we extend our algorithm and analysis to reinforcement learning by developing a novel decomposition over transition steps and also obtain a similar logarithmic regret bound.

Bowen Jin, Jinsung Yoon, Zhen Qin et al.

Large Language Models (LLMs) have revolutionized artificial intelligence with capabilities in reasoning, coding, and communication, driving innovation across industries. Their true potential depends on effective alignment to ensure correct, trustworthy and ethical behavior, addressing challenges like misinformation, hallucinations, bias and misuse. While existing Reinforcement Learning (RL)-based alignment methods are notoriously complex, direct optimization approaches offer a simpler alternative. In this work, we introduce a novel direct optimization approach for LLM alignment by drawing on established Information Retrieval (IR) principles. We present a systematic framework that bridges LLM alignment and IR methodologies, mapping LLM generation and reward models to IR's retriever-reranker paradigm. Building on this foundation, we propose LLM Alignment as Retriever Preference Optimization (LarPO), a new alignment method that enhances overall alignment quality. Extensive experiments validate LarPO's effectiveness with 38.9 % and 13.7 % averaged improvement on AlpacaEval2 and MixEval-Hard respectively. Our work opens new avenues for advancing LLM alignment by integrating IR foundations, offering a promising direction for future research.