NeurIPS 2025 Papers

In this work, we propose a novel framework to enable diffusion models to adapt their generation quality based on real-time network bandwidth constraints. Traditional diffusion models produce high-fidelity images by performing a fixed number of denoising steps, regardless of downstream transmission limitations. However, in practical cloud-to-device scenarios, limited bandwidth often necessitates heavy compression, leading to loss of fine textures and wasted computation. To address this, we introduce a joint end-to-end training strategy where the diffusion model is conditioned on a target quality level derived from the available bandwidth. During training, the model learns to adaptively modulate the denoising process, enabling early-stop sampling that maintains perceptual quality appropriate to the target transmission condition. Our method requires minimal architectural changes and leverages a lightweight quality embedding to guide the denoising trajectory. Experimental results demonstrate that our approach significantly improves the visual fidelity of bandwidth-adapted generations compared to naive early-stopping, offering a promising solution for efficient image delivery in bandwidth-constrained environments. Code is available at: https://github.com/xzhang9308/BADiff.

Introduced by Korpelevich in 1976, the extragradient method (EG) has become a cornerstone technique for solving min-max optimization, root-finding problems, and variational inequalities (VIs). Despite its longstanding presence and significant attention within the optimization community, most works focusing on understanding its convergence guarantees assume the strong $L$-Lipschitz condition. In this work, building on the proposed assumptions by Zhang et al. [2019] for minimization and Vankov et al. [2024a] for VIs, we focus on the more relaxed $\alpha$-symmetric $(L_0, L_1)$-Lipschitz condition. This condition generalizes the standard Lipschitz assumption by allowing the Lipschitz constant to scale with the operator norm, providing a more refined characterization of problem structures in modern machine learning. Under the $\alpha$-symmetric $(L_0, L_1)$-Lipschitz condition, we propose a novel step size strategy for EG to solve root-finding problems and establish sublinear convergence rates for monotone operators and linear convergence rates for strongly monotone operators. Additionally, we prove local convergence guarantees for weak Minty operators. We supplement our analysis with experiments validating our theory and demonstrating the effectiveness and robustness of the proposed step sizes for EG.

The integration of evolutionary algorithms (EAs) with reinforcement learning (RL) has shown superior performance compared to standalone methods. However, previous research focuses on exploration in policy parameter space, while overlooking the reward function search. To bridge this gap, we propose **LaRes**, a novel hybrid framework that achieves efficient policy learning through reward function search. LaRes leverages large language models (LLMs) to generate the reward function population, guiding RL in policy learning. The reward functions are evaluated by the policy performance and improved through LLMs. To improve sample efficiency, LaRes employs a shared experience buffer that collects experiences from all policies, with each experience containing rewards from all reward functions. Upon reward function updates, the rewards of experiences are relabeled, enabling efficient use of historical data. Furthermore, we introduce a Thompson sampling-based selection mechanism that enables more efficient elite interaction. To prevent policy collapse when improving reward functions, we propose the reward scaling and parameter constraint mechanisms to efficiently coordinate reward search with policy learning. Across both initialized and non-initialized settings, LaRes consistently achieves state-of-the-art performance, outperforming strong baselines in both sample efficiency and final performance. The code is available at https://github.com/yeshenpy/LaRes.

Deep learning techniques have driven significant progress in various analytical tasks within 3D genomics in computational biology. However, a holistic understanding of 3D genomics knowledge remains underexplored. Here, we propose ***MIX-HIC***, the first multimodal foundation model of 3D genome that integrates both 3D genome structure and epigenomic tracks, which obtains unified and comprehensive semantics. For accurate heterogeneous semantic fusion, we design the cross-modal interaction and mapping blocks for robust unified representation, yielding the accurate aggregation of 3D genome knowledge. Besides, we introduce the first large-scale dataset comprising over ***1 million*** pairwise samples of Hi-C contact maps and epigenomic tracks for high-quality pre-training, enabling the exploration of functional implications in 3D genomics. Extensive experiments show that MIX-HIC significantly surpasses existing state-of-the-art methods in diverse downstream tasks. This work provides a valuable resource for advancing 3D genomics research.

Recognizing the physical states of objects and their transformations within videos is crucial for structured video understanding and enabling robust real-world applications, such as robotic manipulation. However, pretrained vision-language models often struggle to capture these nuanced dynamics and their temporal context, and specialized object state recognition frameworks may not generalize to unseen actions or objects. We introduce SAGE (State-Action Graph Embeddings), a novel framework that offers a unified model of physical state transitions by decomposing states into fine-grained, language-described visual concepts that are sharable across different objects and actions. SAGE initially leverages Large Language Models to construct a State-Action Graph, which is then multimodally refined using Vision-Language Models. Extensive experiments show that our method significantly outperforms baselines, generalizes effectively to unseen objects and actions in open-world settings. SAGE improves the prior state-of-the-art by as much as 14.6% on novel state recognition with less than 5% of its inference time.

Continual learning for large language models (LLMs) demands a precise balance between $\textbf{plasticity}$ - the ability to absorb new tasks - and $\textbf{stability}$ - the preservation of previously learned knowledge. Conventional rehearsal methods, which replay stored examples, are limited by long-term data inaccessibility; earlier pseudo-rehearsal methods require additional generation modules, while self-synthesis approaches often generate samples that poorly align with real tasks, suffer from unstable outputs, and ignore task relationships. We present $\textbf{\textit{Self-Evolving Pseudo-Rehearsal for Catastrophic Forgetting with Task Similarity}}(\textbf{SERS})$, a lightweight framework that 1) decouples pseudo-input synthesis from label creation, using semantic masking and template guidance to produce diverse, task-relevant prompts without extra modules; 2) applies label self-evolution, blending base-model priors with fine-tuned outputs to prevent over-specialization; and 3) introduces a dynamic regularizer driven by the Wasserstein distance between task distributions, automatically relaxing or strengthening constraints in proportion to task similarity. Experiments across diverse tasks on different LLMs show that our SERS reduces forgetting by over 2\% points against strong pseudo-rehearsal baselines, by ensuring efficient data utilization and wisely transferring knowledge. The code will be released at https://github.com/JerryWangJun/LLM_CL_SERS/.

Associative memory models based on Hopfield networks and self-attention based on key-value mechanisms have been popular approaches in the study of memory mechanisms in deep learning. It has been pointed out that the state update rule of the modern Hopfield network (MHN) in the adiabatic approximation is in agreement with the self-attention layer of Transformer. In this paper, we go beyond this approximation and investigate the relationship between MHN and self-attention. Our results show that the correspondence between Hopfield networks and Transformers can be established in a more generalized form by adding a new variable, the hidden state derived from the MHN, to self-attention. This new attention mechanism, modern Hopfield attention (MHA), allows the inheritance of attention scores from the input layer of the Transformer to the output layer, which greatly improves the nature of attention weights. In particular, we show both theoretically and empirically that MHA hidden states significantly improve serious problem of deep Transformers known as rank collapse and token uniformity. We also confirm that MHA can systematically improve accuracy without adding training parameters to the Vision Transformer or GPT. Our results provide a new case in which Hopfield networks can be a useful perspective for improving the Transformer architecture.

Reinforcement Learning from Human Feedback (RLHF) has emerged as a key technique for post‑training large language models. Despite its empirical success, the theoretical understanding of RLHF is still limited, as learning the KL-regularized target with only preference feedback poses additional challenges compared with canonical RL. Existing works mostly study the reward-based Bradley-Terry (BT) preference model, and extend classical designs utilizing optimism or pessimism. This work, instead, considers the general preference model (whose practical relevance has been observed recently) and obtains performance guarantees with major, order-wise improvements over existing ones. Surprisingly, these results are derived from algorithms that directly use empirical estimates (i.e., greedy sampling), as opposed to constructing optimistic or pessimistic estimates in previous works. This insight has a deep root in the unique structural property of the optimal policy class under the KL-regularized target, and we further specialize it to the BT model, highlighting the surprising sufficiency of greedy sampling in RLHF.

It has been observed that transformers with greater depth (that is, more layers) have more capabilities, but can we establish formally which capabilities are gained? We answer this question with a theoretical proof followed by an empirical study. First, we consider transformers that round to fixed precision except inside attention. We show that this subclass of transformers is expressively equivalent to the programming language $\textsf{C}$-$\textsf{RASP}$ and this equivalence preserves depth. Second, we prove that deeper $\textsf{C}$-$\textsf{RASP}$ programs are more expressive than shallower $\textsf{C}$-$\textsf{RASP}$ programs, implying that deeper transformers are more expressive than shallower transformers (within the subclass mentioned above). The same is also proven for transformers with positional encodings (like RoPE and ALiBi). These results are established by studying a temporal logic with counting operators equivalent to $\textsf{C}$-$\textsf{RASP}$. Finally, we provide empirical evidence that our theory predicts the depth required for transformers without positional encodings to length-generalize on a family of sequential dependency tasks.

3D reconstruction from in-the-wild images remains a challenging task due to inconsistent lighting conditions and transient distractors. Existing methods typically rely on heuristic strategies to handle the low-quality training data, which often struggle to produce stable and consistent reconstructions, frequently resulting in visual artifacts. In this work, we propose Asymmetric Dual 3DGS, a novel framework that leverages the stochastic nature of these artifacts: they tend to vary across different training runs due to minor randomness. Specifically, our method trains two 3D Gaussian Splatting (3DGS) models in parallel, enforcing a consistency constraint that encourages convergence on reliable scene geometry while suppressing inconsistent artifacts. To prevent the two models from collapsing into similar failure modes due to confirmation bias, we introduce a divergent masking strategy that applies two complementary masks: a multi-cue adaptive mask and a self-supervised soft mask, which leads to an asymmetric training process of the two models, reducing shared error modes. In addition, to improve the efficiency of model training, we introduce a lightweight variant called Dynamic EMA Proxy, which replaces one of the two models with a dynamically updated Exponential Moving Average (EMA) proxy, and employs an alternating masking strategy to preserve divergence. Extensive experiments on challenging real-world datasets demonstrate that our method consistently outperforms existing approaches while achieving high efficiency. Codes and trained models will be released.

In this paper, we propose a simple and efficient method for value model training on long-context reasoning traces. Compared to existing process reward models (PRMs), our method does not require a fine-grained notion of ``step,'' which is difficult to define for long-context reasoning models. By collecting a dataset of 2.5 million reasoning traces, we train a 1.5B token-level value model and apply it to DeepSeek models for improved performance with test-time compute scaling. We find that block-wise value-guided search (\texttt{VGS}) with a final weighted majority vote achieves better test-time scaling than standard methods such as majority voting or best-of-$n$. Moreover, \texttt{VGS} significantly reduces the inference FLOPs required to achieve the same performance of majority voting. Our dataset, model and codebase are open-sourced at \codeurl.

Real-time decoding of target variables from multiple simultaneously recorded neural time-series modalities, such as discrete spiking activity and continuous field potentials, is important across various neuroscience applications. However, a major challenge for doing so is that different neural modalities can have different timescales (i.e., sampling rates) and different probabilistic distributions, or can even be missing at some time-steps. Existing nonlinear models of multimodal neural activity do not address different timescales or missing samples across modalities. Further, some of these models do not allow for real-time decoding. Here, we develop a learning framework that can enable real-time recursive decoding while nonlinearly aggregating information across multiple modalities with different timescales and distributions and with missing samples. This framework consists of 1) a multiscale encoder that nonlinearly aggregates information after learning within-modality dynamics to handle different timescales and missing samples in real time, 2) a multiscale dynamical backbone that extracts multimodal temporal dynamics and enables real-time recursive decoding, and 3) modality-specific decoders to account for different probabilistic distributions across modalities. In both simulations and three distinct multiscale brain datasets, we show that our model can aggregate information across modalities with different timescales and distributions and missing samples to improve real-time target decoding. Further, our method outperforms various linear and nonlinear multimodal benchmarks in doing so.

We introduce active generation of Pareto sets (A-GPS), a new framework for online discrete black-box multi-objective optimization (MOO). A-GPS learns a generative model of the Pareto set that supports a-posteriori conditioning on user preferences. The method employs a class probability estimator (CPE) to predict non-dominance relations and to condition the generative model toward high-performing regions of the search space. We also show that this non-dominance CPE implicitly estimates the probability of hypervolume improvement (PHVI). To incorporate subjective trade-offs, A-GPS introduces _preference direction vectors_ that encode user-specified preferences in objective space. At each iteration, the model is updated using both Pareto membership and alignment with these preference directions, producing an amortized generative model capable of sampling across the Pareto front without retraining. The result is a simple yet powerful approach that achieves high-quality Pareto set approximations, avoids explicit hypervolume computation, and flexibly captures user preferences. Empirical results on synthetic benchmarks and protein design tasks demonstrate strong sample efficiency and effective preference incorporation.

Emergence is a fascinating property of large language models and neural networks more broadly: as models scale and train for longer, they sometimes develop new abilities in sudden ways. Despite initial studies, we still lack a comprehensive understanding of how and when these abilities emerge. To address this gap, we study the emergence over training of sparse attention, a critical and frequently observed attention pattern in Transformers. By combining theoretical analysis of a toy model with empirical observations on small Transformers trained on a linear regression variant, we uncover the mechanics driving sparse attention emergence and reveal that emergence timing follows power laws based on task structure, architecture, and optimizer choice. We additionally find that repetition can greatly speed up emergence. Finally, we confirm these results on a well-studied in-context associative recall task. Our findings provide a simple, theoretically grounded framework for understanding how data distributions and model design influence the learning dynamics behind one form of emergence.

We introduce DanmakuTPPBench, a comprehensive benchmark designed to advance multi-modal Temporal Point Process (TPP) modeling in the era of Large Language Models (LLMs). While TPPs have been widely studied for modeling temporal event sequences, existing datasets are predominantly unimodal, hindering progress in models that require joint reasoning over temporal, textual, and visual information. To address this gap, DanmakuTPPBench comprises two complementary components: (1) DanmakuTPP-Events, a novel dataset derived from the Bilibili video platform, where user-generated bullet comments (Danmaku) naturally form multi-modal events annotated with precise timestamps, rich textual content, and corresponding video frames; (2) DanmakuTPP-QA, a challenging question-answering dataset constructed via a novel multi-agent pipeline powered by state-of-the-art LLMs and multi-modal LLMs (MLLMs), targeting complex temporal-textual-visual reasoning. We conduct extensive evaluations using both classical TPP models and recent MLLMs, revealing significant performance gaps and limitations in current methods’ ability to model multi-modal event dynamics. Our benchmark establishes strong baselines and calls for further integration of TPP modeling into the multi-modal language modeling landscape. Project page: https://github.com/FRENKIE-CHIANG/DanmakuTPPBench.

Federated Learning (FL) confronts a significant challenge known as data heterogeneity, which impairs model performance and convergence. Existing methods have made notable progress in addressing this issue. However, improving performance in certain heterogeneity scenarios remains an overlooked question: _How robust are these methods to deploy under diverse heterogeneity scenarios?_ To answer this, we conduct comprehensive evaluations across varied heterogeneity scenarios, showing that most existing methods exhibit limited robustness. Meanwhile, insights from these experiments highlight that sharing statistical information can mitigate heterogeneity by enabling clients to update with a global perspective. Motivated by this, we propose **FedGPS** (**Fed**erated **G**oal-**P**ath **S**ynergy), a novel framework that seamlessly integrates statistical distribution and gradient information from others. Specifically, FedGPS statically modifies each client’s learning objective to implicitly model the global data distribution using surrogate information, while dynamically adjusting local update directions with gradient information from other clients at each round. Extensive experiments show that FedGPS outperforms state-of-the-art methods across diverse heterogeneity scenarios, validating its effectiveness and robustness. The code is available at: <https://github.com/CUHK-AIM-Group/FedGPS>.

We introduce IndEgo, a multimodal egocentric and exocentric dataset addressing common industrial tasks, including assembly/disassembly, logistics and organisation, inspection and repair, woodworking, and others. The dataset contains 3,460 egocentric recordings (approximately 197 hours), along with 1,092 exocentric recordings (approximately 97 hours). A key focus of the dataset is collaborative work, where two workers jointly perform cognitively and physically intensive tasks. The egocentric recordings include rich multimodal data and added context via eye gaze, narration, sound, motion, and others. We provide detailed annotations (actions, summaries, mistake annotations, narrations), metadata, processed outputs (eye gaze, hand pose, semi-dense point cloud), and benchmarks on procedural and non-procedural task understanding, Mistake Detection, and reasoning-based Question Answering. Baseline evaluations for Mistake Detection, Question Answering and collaborative task understanding show that the dataset presents a challenge for the state-of-the-art multimodal models. Our dataset is available at: https://huggingface.co/datasets/FraunhoferIPK/IndEgo

Image restoration is a fundamental task in computer vision and machine learning, which learns a mapping between the clear images and the degraded images under various conditions (e.g., blur, low-light, haze). Yet, most existing image restoration methods are highly restricted by the requirement of degraded and clear image pairs, which limits the generalization and feasibility to enormous real-world scenarios without paired images. To address this bottleneck, we propose a Degradation-aware Dynamic Schr\"{o}dinger Bridge (DDSB) for unpaired image restoration. Its general idea is to learn a Schr\"{o}dinger Bridge between clear and degraded image distribution, while at the same time emphasizing the physical degradation priors to reduce the accumulation of errors during the restoration process. A Degradation-aware Optimal Transport (DOT) learning scheme is accordingly devised. Training a degradation model to learn the inverse restoration process is particularly challenging, as it must be applicable across different stages of the iterative restoration process. A Dynamic Transport with Consistency (DTC) learning objective is further proposed to reduce the loss of image details in the early iterations and therefore refine the degradation model. Extensive experiments on multiple image degradation tasks show its state-of-the-art performance over the prior arts.

Recent advances in data-centric deep generative models have led to significant progress in solving inverse imaging problems. However, these models (e.g., diffusion models (DMs)) typically require large amounts of fully sampled (clean) training data, which is often impractical in medical and scientific settings such as dynamic imaging. On the other hand, training-data-free approaches like the Deep Image Prior (DIP) do not require clean ground-truth images but suffer from noise overfitting and can be computationally expensive as the network parameters need to be optimized for each measurement vector independently. Moreover, DIP-based methods often overlook the potential of learning a prior using a small number of sub-sampled measurements (or degraded images) available during training. In this paper, we propose **UGoDIT**—an **U**nsupervised **G**r**o**up **DI**P with **T**ransferable weights—designed for the low-data regime where only a very small number, $M$, of sub-sampled measurement vectors are available during training. Our method learns a set of transferable weights by optimizing a shared encoder and $M$ disentangled decoders. At test time, we reconstruct the unseen degraded image using a DIP network, where part of the parameters are fixed to the learned weights, while the remaining are optimized to enforce measurement consistency. We evaluate \our on both medical (multi-coil MRI) and natural (super resolution and non-linear deblurring) image recovery tasks under various settings. Compared to recent standalone DIP methods, \our provides accelerated convergence and notable improvement in reconstruction quality. Furthermore, our method achieves performance competitive with SOTA DM-based and supervised approaches, despite not requiring large amounts of clean training data. Our code is available at: https://github.com/sjames40/UGoDIT.