NeurIPS 2025 Papers

Recent advances in text-to-video diffusion models have enabled high-quality video synthesis, but controllable generation remains challenging—particularly under limited data and compute. Existing fine-tuning methods often rely on external encoders or architectural modifications, which demand large datasets and are typically restricted to spatially aligned conditioning, limiting flexibility and scalability. In this work, we introduce Temporal In-Context Fine-Tuning (TIC-FT), an efficient and versatile approach for adapting pretrained video diffusion models to diverse conditional generation tasks. Our key idea is to concatenate condition and target frames along the temporal axis and insert intermediate buffer frames with progressively increasing noise levels. These buffer frames enable smooth transitions, aligning the fine-tuning process with the pretrained model’s temporal dynamics. TIC-FT requires no architectural changes and achieves strong performance with as few as 10–30 training samples. We validate our method across a range of tasks—including image-to-video and video-to-video generation—using large-scale base models such as CogVideoX-5B and Wan-14B. Extensive experiments show that TIC-FT outperforms existing baselines in both condition fidelity and visual quality, while remaining highly efficient in both training and inference.

Industrial Anomaly Detection (IAD) is an indispensable quality control technology in modern production processes. Recently, on account of the outstanding visual comprehension and cross-domain knowledge transfer capabilities of multimodal large language models (MLLMs), existing studies have explored the application of MLLMs in the IAD domain and established some multimodal IAD datasets. However, although the latest datasets contain various fundamental IAD tasks, they formulate tasks in a general question-and-answer format lacking a rigorous reasoning process, and they are relatively limited in the diversity of scenarios, which restricts their reliability in practical applications. In this paper, we propose AnomalyCoT, a multimodal Chain-of-Thought (CoT) dataset for multi-scenario IAD tasks. It consists of 37,565 IAD samples with the CoT data and is defined by challenging composite IAD tasks. Meanwhile, the CoT data for each sample provides precise coordinates of anomaly regions, thereby improving visual comprehension of defects across different types. AnomalyCoT is constructed through a systematic pipeline and involves multiple manual operations. Based on AnomalyCoT, we conducted a comprehensive evaluation of various mainstream MLLMs and fine-tuned representative models in different ways. The final results show that Gemini-2.0-flash achieved the best performance in the direct evaluation with an accuracy rate of 59.6\%, while Llama 3.2-Vision achieves the best performance after LoRA fine-tuning with an accuracy rate of 94.0\%. Among all the fine-tuned models, the average accuracy improvement reaches 36.5\%, demonstrating the potential of integrating CoT datasets in future applications within the IAD field. The code and data are available at \url{https://github.com/Zhaolutuan/AnomalyCoT}.

Image fusion is a fundamental and important task in computer vision, aiming to combine complementary information from different modalities to fuse images. In recent years, diffusion models have made significant developments in the field of image fusion. However, diffusion models often require complex computations and redundant inference time, which reduces the applicability of these methods. To address this issue, we propose RFfusion, an efficient one-step diffusion model for image fusion based on Rectified Flow. We incorporate Rectified Flow into the image fusion task to straighten the sampling path in the diffusion model, achieving one-step sampling without the need for additional training, while still maintaining high-quality fusion results. Furthermore, we propose a task-specific variational autoencoder (VAE) architecture tailored for image fusion, where the fusion operation is embedded within the latent space to further reduce computational complexity. To address the inherent discrepancy between conventional reconstruction-oriented VAE objectives and the requirements of image fusion, we introduce a two-stage training strategy. This approach facilitates the effective learning and integration of complementary information from multi-modal source images, thereby enabling the model to retain fine-grained structural details while significantly enhancing inference efficiency. Extensive experiments demonstrate that our method outperforms other state-of-the-art methods in terms of both inference speed and fusion quality.

A critical bottleneck in deep reinforcement learning (DRL) is sample inefficiency, as training high-performance agents often demands extensive environmental interactions. Model-based reinforcement learning (MBRL) mitigates this by building world models that simulate environmental dynamics and generate synthetic experience, improving sample efficiency. However, conventional world models process observations holistically, failing to decouple dynamic objects and temporal features from static backgrounds. This approach is computationally inefficient, especially for visual tasks where dynamic objects significantly influence rewards and decision-making performance. To address this, we introduce DyMoDreamer, a novel MBRL algorithm that incorporates a dynamic modulation mechanism to improve the extraction of dynamic features and enrich the temporal information. DyMoDreamer employs differential observations derived from a novel inter-frame differencing mask, explicitly encoding object-level motion cues and temporal dynamics. Dynamic modulation is modeled as stochastic categorical distributions and integrated into a recurrent state-space model (RSSM), enhancing the model's focus on reward-relevant dynamics. Experiments demonstrate that DyMoDreamer sets a new state-of-the-art on the Atari $100$k benchmark with a $156.6$\% mean human-normalized score, establishes a new record of $832$ on the DeepMind Visual Control Suite, and gains a $9.5$\% performance improvement after $1$M steps on the Crafter benchmark.

We introduce MOSDT, the first algorithm designed for multi-agent offline safe reinforcement learning (MOSRL), alongside MOSDB, the first dataset and benchmark for this domain. Different from most existing knowledge distillation-based multi-agent RL methods, we propose policy self-distillation (PSD) with a new global information reconstruction scheme by fusing the observation features of all agents, streamlining training and improving parameter efficiency. We adopt full parameter sharing across agents, significantly slashing parameter count and boosting returns up to 38.4-fold by stabilizing training. We propose a new plug-and-play cost binary embedding (CBE) module, which binarizes cumulative costs as safety signals and embeds the signals into return features for efficient information aggregation. On the strong MOSDB benchmark, MOSDT achieves state-of-the-art (SOTA) returns in 14 out of 18 tasks (across all base environments including MuJoCo, Safety Gym, and Isaac Gym) while ensuring complete safety, with only 65% of the execution parameter count of a SOTA single-agent offline safe RL method CDT. Code, dataset, and results are available at this website: https://github.com/Lucian1115/MOSDT.git

Searching through chemical space is an exceptionally challenging problem because the number of possible molecules grows combinatorially with the number of atoms. Large, autoregressive models trained on databases of chemical compounds have yielded powerful generators, but we still lack robust strategies for generating molecules with desired properties. This molecular search problem closely resembles the "alignment" problem for large language models, though for many chemical tasks we have a specific and easily evaluable reward function. Here, we introduce an algorithm called energy rank alignment (ERA) that leverages an explicit reward function to produce a gradient-based objective that we use to optimize autoregressive policies. We show theoretically that this algorithm is closely related to proximal policy optimization (PPO) and direct preference optimization (DPO), but has a minimizer that converges to an ideal Gibbs-Boltzmann distribution with the reward playing the role of an energy function. Furthermore, this algorithm is highly scalable, does not require reinforcement learning, and performs well relative to DPO when the number of preference observations per pairing is small. We deploy this approach to align molecular transformers and protein language models to generate molecules and protein sequences, respectively, with externally specified properties and find that it does so robustly, searching through diverse parts of chemical space.

Neural Combinatorial Optimisation (NCO) is a promising learning-based approach for solving Vehicle Routing Problems (VRPs) without extensive manual design. While existing constructive NCO methods typically follow an appending-based paradigm that sequentially adds unvisited nodes to partial solutions, this rigid approach often leads to suboptimal results. To overcome this limitation, we explore the idea of the insertion-based paradigm and propose Learning to Construct with Insertion-based Paradigm (L2C-Insert), a novel learning-based method for constructive NCO. Unlike traditional approaches, L2C-Insert builds solutions by strategically inserting unvisited nodes at any valid position in the current partial solution, which can significantly enhance the flexibility and solution quality. The proposed framework introduces three key components: a novel model architecture for precise insertion position prediction, an efficient training scheme for model optimization, and an advanced inference technique that fully exploits the insertion paradigm's flexibility. Extensive experiments on both synthetic and real-world instances of the Travelling Salesman Problem (TSP) and Capacitated Vehicle Routing Problem (CVRP) demonstrate that L2C-Insert consistently achieves superior performance across various problem sizes. The code is available at [https://github.com/CIAM-Group/L2C\_Insert](https://github.com/CIAM-Group/L2C\_Insert).

To fix the `bias in, bias out' problem in fair machine learning, it is important to steer feature distributions of data or internal representations of Large Language Models (LLMs) to \emph{ideal} ones that guarantee group-fair outcomes. Previous work on fair generative models and representation steering could greatly benefit from provable fairness guarantees on the model output. We define a distribution as \emph{ideal} if the minimizer of any cost-sensitive risk on it is guaranteed to have exact group-fair outcomes (e.g., demographic parity, equal opportunity)---in other words, it has no fairness-utility trade-off. We formulate an optimization program for optimal steering by finding the nearest \emph{ideal} distribution in KL-divergence, and provide efficient algorithms for it when the underlying distributions come from well-known parametric families (e.g., normal, log-normal). Empirically, our optimal steering techniques on both synthetic and real-world datasets improve fairness without diminishing utility (and sometimes even improve utility). We demonstrate affine steering of LLM representations to reduce bias in multi-class classification, e.g., occupation prediction from a short biography in Bios dataset (De-Arteaga et al.). Furthermore, we steer internal representations of LLMs towards desired outputs so that it works equally well across different groups.

Large Language Models (LLMs) have demonstrated notable capabilities across financial tasks, including financial report summarization, earnings call transcript analysis, and asset classification. However, their real-world effectiveness in managing complex fund investment remains inadequately assessed. A fundamental limitation of existing benchmarks for evaluating LLM-driven trading strategies is their reliance on historical back-testing, inadvertently enabling LLMs to "time travel"—leveraging future information embedded in their training corpora, thus resulting in possible information leakage and overly optimistic performance estimates. To address this issue, we introduce DeepFund, a live fund benchmark tool designed to rigorously evaluate LLM in real-time market conditions. Utilizing a multi-agent architecture, DeepFund connects directly with real-time stock market data—specifically data published after each model’s pretraining cutoff—to ensure fair and leakage-free evaluations. Empirical tests on nine flagship LLMs from leading global institutions across multiple investment dimensions—including ticker-level analysis, investment decision-making, portfolio management, and risk control—reveal significant practical challenges. Notably, even cutting-edge models such as DeepSeek-V3 and Claude-3.7-Sonnet incur net trading losses within DeepFund real-time evaluation environment, underscoring the present limitations of LLMs for active fund management. Our code is available at https://github.com/HKUSTDial/DeepFund.

Given $n$ i.i.d.random matrices $A_i \in \mathbb{R}^{d \times d}$ that share common expectation $\Sigma$, the objective of Differentially Private Stochastic PCA is to identify a subspace of dimension $k$ that captures the largest variance directions of $\Sigma$, while preserving differential privacy (DP) of each individual $A_i$. Existing methods either (i) require the sample size $n$ to scale super-linearly with dimension $d$, even under Gaussian assumptions on the $A_i$, or (ii) introduce excessive noise for DP even when the intrinsic randomness within $A_i$ is small.~\citet{liu2022dp} addressed these issues for sub-Gaussian data but only for estimating the top eigenvector ($k=1$) using their algorithm DP-PCA. We propose the first algorithm capable of estimating the top $k$ eigenvectors for arbitrary $k \leq d$, whilst overcoming both limitations above. For $k=1$, our algorithm matches the utility guarantees of DP-PCA, achieving near-optimal statistical error even when $n = \tilde{O}(d)$. We further provide a lower bound for general $k > 1$, matching our upper bound up to a factor of $k$, and experimentally demonstrate the advantages of our algorithm over comparable baselines.

This paper introduces Completion Pruning Policy Optimization (CPPO) to accelerate the training of reasoning models based on Group Relative Policy Optimization (GRPO). GRPO, while effective, incurs high training costs due to the need to sample multiple completions for each question. Our experiment and theoretical analysis reveal that the number of completions impacts model accuracy yet increases training time multiplicatively, and not all completions contribute equally to policy training---their contribution depends on their relative advantage. To address these issues, we propose CPPO, which prunes completions with low absolute advantages, significantly reducing the number needed for gradient calculation and updates. Additionally, we introduce a dynamic completion allocation strategy to maximize GPU utilization by incorporating additional questions, further enhancing training efficiency. Experiments show that CPPO achieves up to $7.98\times$ speedup on GSM8K and $3.48\times$ on Math while preserving or even enhancing the accuracy compared to the original GRPO. We release our code at https://github.com/lzhxmu/CPPO.

Many confounding-resistant domain generalization methods for image classification have been developed based on causal interventions. However, their reliance on strong assumptions limits their effectiveness in handling unobserved confounders. Although recent work introduces instrumental variables (IVs) to overcome this limitation, the reliance on manually predefined instruments, particularly in the context of visual data, may result in severe bias or invalidity when IV conditions are violated. To address these issues, we propose a novel approach to automatically learning Visual Instrumental Variables for confounding-resistant Domain Generalization (VIV-DG). We observe that certain non-causal visual attributes in image data naturally satisfy the basic conditions required for valid IVs. Motivated by this insight, we propose the visual instrumental variable, a novel concept that extends classical IV theory to the visual domain. Furthermore, we develop an automatic visual instrumental variable learner that enforces IV conditions on learned representations, enabling the automatic learning of valid visual instrumental variables from image data. Ultimately, VIV-DG inherits the strengths of classical IVs to mitigate unobserved confounding and avoids the significant bias caused by violations of IV conditions in predefined IVs. Extensive experiments on multiple benchmarks verify that VIV-DG achieves superior generalization ability.

Nuclear Magnetic Resonance (NMR) spectroscopy is a cornerstone technique for determining the structures of small molecules and is especially critical in the discovery of novel natural products and clinical therapeutics. Yet, interpreting NMR spectra remains a time-consuming, manual process requiring extensive domain expertise. We introduce ChefNMR (CHemical Elucidation From NMR), an end-to-end framework that directly predicts an unknown molecule's structure solely from its 1D NMR spectra and chemical formula. We frame structure elucidation as conditional generation from an atomic diffusion model built on a non-equivariant transformer architecture. To model the complex chemical groups found in natural products, we generated a dataset of simulated 1D NMR spectra for over 111,000 natural products. ChefNMR predicts the structures of challenging natural product compounds with an unsurpassed accuracy of over 65%. This work takes a significant step toward solving the grand challenge of automating small-molecule structure elucidation and highlights the potential of deep learning in accelerating molecular discovery.

Reconstructing detailed hand avatars plays a crucial role in various applications. While prior works have focused on capturing high-fidelity hand geometry, they heavily rely on high-resolution multi-view image inputs and struggle to generalize on low-resolution images. Multi-view image super-resolution methods have been proposed to enforce 3D view consistency. These methods, however, are limited to static objects/scenes with fixed resolutions and are not applicable to articulated deformable hands. In this paper, we propose SRHand (Super-Resolution Hand), the method for reconstructing detailed 3D geometry as well as textured images of hands from low-resolution images. SRHand leverages the advantages of implicit image representation with explicit hand meshes. Specifically, we introduce a geometric-aware implicit image function (GIIF) that learns detailed hand prior by upsampling the coarse input images. By jointly optimizing the implicit image function and explicit 3D hand shapes, our method preserves multi-view and pose consistency among upsampled hand images, and achieves fine-detailed 3D reconstruction (wrinkles, nails). In experiments using the InterHand2.6M and Goliath datasets, our method significantly outperforms state-of-the-art image upsampling methods adapted to hand datasets, and 3D hand reconstruction methods, quantitatively and qualitatively. The code will be publicly available.

We study the problem of infrequent exploration in linear bandits, addressing a significant yet overlooked gap between fully adaptive exploratory methods (e.g., UCB and Thompson Sampling), which explore potentially at every time step, and purely greedy approaches, which require stringent diversity assumptions to succeed. Continuous exploration can be impractical or unethical in safety-critical or costly domains, while purely greedy strategies typically fail without adequate contextual diversity. To bridge these extremes, we introduce a simple and practical framework, INFEX, explicitly designed for infrequent exploration. INFEX executes a base exploratory policy according to a given schedule while predominantly choosing greedy actions in between. Despite its simplicity, our theoretical analysis demonstrates that INFEX achieves instance-dependent regret matching standard provably efficient algorithms, provided the exploration frequency exceeds a logarithmic threshold. Additionally, INFEX is a general, modular framework that allows seamless integration of any fully adaptive exploration method, enabling wide applicability and ease of adoption. By restricting intensive exploratory computations to infrequent intervals, our approach can also enhance computational efficiency. Empirical evaluations confirm our theoretical findings, showing state-of-the-art regret performance and runtime improvements over existing methods.

Graphons, as limit objects of dense graph sequences, play a central role in the statistical analysis of network data. However, existing graphon estimation methods often struggle with scalability to large networks and resolution-independent approximation, due to their reliance on estimating latent variables or costly metrics such as the Gromov-Wasserstein distance. In this work, we propose a novel, scalable graphon estimator that directly recovers the graphon via moment matching, leveraging implicit neural representations (INRs). Our approach avoids latent variable modeling by training an INR--mapping coordinates to graphon values--to match empirical subgraph counts (i.e., moments) from observed graphs. This direct estimation mechanism yields a polynomial-time solution and crucially sidesteps the combinatorial complexity of Gromov-Wasserstein optimization. Building on foundational results, we establish a theoretical guarantee: when the observed subgraph motifs sufficiently represent those of the true graphon (a condition met with sufficiently large or numerous graph samples), the estimated graphon achieves a provable upper bound in cut distance from the ground truth. Additionally, we introduce MomentMixup, a data augmentation technique that performs mixup in the moment space to enhance graphon-based learning. Our graphon estimation method achieves strong empirical performance--demonstrating high accuracy on small graphs and superior computational efficiency on large graphs--outperforming state-of-the-art scalable estimators in 75\% of benchmark settings and matching them in the remaining cases. Furthermore, MomentMixup demonstrated improved graph classification accuracy on the majority of our benchmarks.

Cosmological simulations provide a wealth of data in the form of point clouds and directed trees. A crucial goal is to extract insights from this data that shed light on the nature and composition of the Universe. In this paper we introduce CosmoBench, a benchmark dataset curated from state-of-the-art cosmological simulations whose runs required more than 41 million core-hours and generated over two petabytes of data. CosmoBench is the largest dataset of its kind: it contains 34 thousand point clouds from simulations of dark matter halos and galaxies at three different length scales, as well as 25 thousand directed trees that record the formation history of halos on two different time scales. The data in CosmoBench can be used for multiple tasks---to predict cosmological parameters from point clouds and merger trees, to predict the velocities of individual halos and galaxies from their collective positions, and to reconstruct merger trees on finer time scales from those on coarser time scales. We provide multiple baselines on these tasks, some based on established approaches from cosmological modeling and others rooted in machine learning. For the latter, we study different approaches---from simple linear models that are minimally constrained by symmetries to much larger and more computationally-demanding models in deep learning, such as graph neural networks. We find that least-squares fits with a handful of invariant features sometimes outperform deep architectures with many more parameters and far longer training times. Still there remains tremendous potential to improve these baselines by combining machine learning and cosmological modeling in a more principled way, one that fully exploits the structure in the data. CosmoBench sets the stage for bridging cosmology and geometric deep learning at scale. We invite the community to push the frontier of scientific discovery by engaging with this challenging, high-impact dataset. The data and code are available at [this URL](https://cosmobench.streamlit.app/).

Recent advancements in multimodal large language models (MLLMs) have focused on integrating multiple modalities, yet their ability to simultaneously process and reason across different inputs remains underexplored. We introduce OmniBench, a novel benchmark designed to evaluate models’ ability to recognize, interpret, and reason across visual, acoustic, and textual inputs simultaneously. We define language models capable of such tri-modal processing as omni-language models (OLMs). OmniBench features high-quality human annotations that require integrated understanding across all modalities. Our evaluation reveals that: i) open-source OLMs show significant limitations in instruction-following and reasoning in tri-modal contexts; and ii) most baseline models perform poorly (below 50% accuracy) even with textual alternatives to image/audio inputs. To address these limitations, we develop OmniInstruct, an 96K-sample instruction tuning dataset for training OLMs. We advocate for developing more robust tri-modal integration techniques and training strategies to enhance OLM performance. Codes and data could be found at https://m-a-p.ai/OmniBench/.

Ultrasound (US) is a widely used medical imaging modality due to its real-time capabilities, non-invasive nature, and cost-effectiveness. By reducing operator dependency and enhancing access to complex anatomical regions, robotic ultrasound can help improve workflow efficiency. Recent studies have demonstrated the potential of deep reinforcement learning (DRL) and imitation learning (IL) to enable more autonomous and intelligent robotic ultrasound navigation. However, the application of learning-based robotic ultrasound to computer-assisted surgical tasks, such as anatomy reconstruction and surgical guidance, remains largely unexplored. A key bottleneck for this is the lack of realistic and efficient simulation environments tailored to these tasks. In this work, we present SonoGym, a scalable simulation platform for robotic ultrasound, enabling parallel simulation across tens to hundreds of environments. Our framework supports realistic and real-time simulation of US data from CT-derived 3D models of the anatomy through both a physics-based and a Generative Adversarial Network (GAN) approach. Our framework enables the training of DRL and recent IL agents (vision transformers and diffusion policies) for relevant tasks in robotic orthopedic surgery by integrating common robotic platforms and orthopedic end effectors. We further incorporate submodular DRL---a recent method that handles history-dependent rewards---for anatomy reconstruction and safe reinforcement learning for surgery. Our results demonstrate successful policy learning across a range of scenarios, while also highlighting the limitations of current methods in clinically relevant environments. We believe our simulation can facilitate research in robot learning approaches for such challenging robotic surgery applications. Dataset, codes and videos are publicly available at https://sonogym.github.io/.