NeurIPS 2025 论文列表

Traditional Chinese Medicine (TCM), as an effective alternative medicine, has been receiving increasing attention. In recent years, the rapid development of large language models (LLMs) tailored for TCM has highlighted the urgent need for an objective and comprehensive evaluation framework to assess their performance on real-world tasks. However, existing evaluation datasets are limited in scope and primarily text-based, lacking a unified and standardized multimodal question-answering (QA) benchmark. To address this issue, we introduce TCM-Ladder, the first comprehensive multimodal QA dataset specifically designed for evaluating large TCM language models. The dataset covers multiple core disciplines of TCM, including fundamental theory, diagnostics, herbal formulas, internal medicine, surgery, pharmacognosy, and pediatrics. In addition to textual content, TCM-Ladder incorporates various modalities such as images and videos. The dataset was constructed using a combination of automated and manual filtering processes and comprises over 52,000 questions. These questions include single-choice, multiple-choice, fill-in-the-blank, diagnostic dialogue, and visual comprehension tasks. We trained a reasoning model on TCM-Ladder and conducted comparative experiments against nine state-of-the-art general domain and five leading TCM-specific LLMs to evaluate their performance on the dataset. Moreover, we propose Ladder-Score, an evaluation method specifically designed for TCM question answering that effectively assesses answer quality in terms of terminology usage and semantic expression. To the best of our knowledge, this is the first work to systematically evaluate mainstream general domain and TCM-specific LLMs on a unified multimodal benchmark. The datasets and leaderboard are publicly available at https://tcmladder.com and will be continuously updated. The source code is available at https://github.com/orangeshushu/TCM-Ladder.

The immense model sizes of large language models (LLMs) challenge deployment on memory-limited consumer GPUs. Although model compression and parameter offloading are common strategies to address memory limitations, compression can degrade quality, and offloading maintains quality but suffers from slow inference. Speculative decoding presents a promising avenue to accelerate parameter offloading, utilizing a fast draft model to propose multiple draft tokens, which are then verified by the target LLM in parallel with a single forward pass. This method reduces the time-consuming data transfers in forward passes that involve offloaded weight transfers. Existing methods often rely on pretrained weights of the same family, but require additional training to align with custom-trained models. Moreover, approaches that involve draft model training usually yield only modest speedups. This limitation arises from insufficient alignment with the target model, preventing higher token acceptance lengths. To address these challenges and achieve greater speedups, we propose SubSpec, a plug-and-play method to accelerate parameter offloading that is lossless and training-free. SubSpec constructs a highly aligned draft model by generating low-bit quantized substitute layers from offloaded target LLM portions. Additionally, our method shares the remaining GPU-resident layers and the KV-Cache, further reducing memory overhead and enhance alignment. SubSpec achieves a high average acceptance length, delivering 9.1$\times$ speedup for Qwen2.5 7B on MT-Bench (8GB VRAM limit) and an average of 12.5$\times$ speedup for Qwen2.5 32B on popular generation benchmarks (24GB VRAM limit).

Vision-language models (VLMs) have recently been integrated into multiple instance learning (MIL) frameworks to address the challenge of few-shot, weakly supervised classification of whole slide images (WSIs). A key trend involves leveraging multi-scale information to better represent hierarchical tissue structures. However, existing methods often face two key limitations: (1) insufficient modeling of interactions within the same modalities across scales (e.g., 5x and 20x) and (2) inadequate alignment between visual and textual modalities on the same scale. To address these gaps, we propose HiVE-MIL, a hierarchical vision-language framework that constructs a unified graph consisting of (1) parent–child links between coarse (5x) and fine (20x) visual/textual nodes to capture hierarchical relationships, and (2) heterogeneous intra-scale edges linking visual and textual nodes on the same scale. To further enhance semantic consistency, HiVE-MIL incorporates a two-stage, text-guided dynamic filtering mechanism that removes weakly correlated patch–text pairs, and introduces a hierarchical contrastive loss to align textual semantics across scales. Extensive experiments on TCGA breast, lung, and kidney cancer datasets demonstrate that HiVE-MIL consistently outperforms both traditional MIL and recent VLM-based MIL approaches, achieving gains of up to 4.1% in macro F1 under 16-shot settings. Our results demonstrate the value of jointly modeling hierarchical structure and multimodal alignment for efficient and scalable learning from limited pathology data. The code is available at https://github.com/bryanwong17/HiVE-MIL.

Recent text-to-image models have revolutionized image generation, but they still struggle with maintaining concept consistency across generated images. While existing works focus on character consistency, they often overlook the crucial role of scenes in storytelling, which restricts their creativity in practice. This paper introduces scene-oriented story generation, addressing two key challenges: (i) scene planning, where current methods fail to ensure scene-level narrative coherence by relying solely on text descriptions, and (ii) scene consistency, which remains largely unexplored in terms of maintaining scene consistency across multiple stories. We propose SceneDecorator, a training-free framework that employs VLM-Guided Scene Planning to ensure narrative coherence across different scenes in a ``global-to-local'' manner, and Long-Term Scene-Sharing Attention to maintain long-term scene consistency and subject diversity across generated stories. Extensive experiments demonstrate the superior performance of SceneDecorator, highlighting its potential to unleash creativity in the fields of arts, films, and games.

The fine-tuning of pre-trained models has become ubiquitous in generative AI, computer vision, and robotics. Although much attention has been paid to improving the efficiency of fine-tuning model, there has been less scholarship around fine-tuning specifically for improved model performance. To remedy this gap, we present PROFIT, one of the first optimizers designed to incrementally fine-tune converged models on new tasks and/or datasets. Unlike traditional optimizers such as SGD or Adam, which make minimal assumptions due to random initializations, PROFIT takes the properties of a converged model into account explicitly to regularize the optimization process. Employing a temporal gradient-orthogonalization process, PROFIT outperforms fine-tuning methods in various tasks, from image classification to multimodal language model training to large-scale motion prediction. Moreover, PROFIT is encapsulated as a modular optimizer, which makes it easy to integrate directly into any training pipeline with minimal engineering effort.

Reinforcement learning (RL) has become an indispensable post-training step for unlocking the full potential of Large Language Models (LLMs). Its core motivation is to incentivize the model’s inference trajectory via a reward model, effectively balancing the exploration–exploitation trade-off in scenarios where collecting exhaustive input–output ground-truth pairs is infeasible. This motivation naturally extends to visual generation, where perfect alignment between an image and a textual prompt is inherently ambiguous and often unattainable. However, existing visual generative models are not yet ready for RL due to the following two fundamental drawbacks that undermine the foundations of RL: 1) For diffusion-based models, the actual generation trajectories of sampled images cannot be reliably rewarded, as diffusion inversion is notoriously difficult. 2) For autoregressive (AR) models, we show that the widely used spatial visual tokens do not satisfy the Bellman equation and thus violate the policy improvement theorem of RL. To this end, we propose to use Selftok (Self-consistency Tokenizer), which represents each image as a sequential 1D stream of discrete, autoregressive tokens. Together with language, we train a pure AR vision-language model (VLM) for visual generation. Impressively, without using any text-image training pairs, a simple policy gradient algorithm applied to Selftok tokens significantly boosts visual generation performance, surpassing existing models by a large margin. Implementation details are provided in the Appendix.

In large-scale recommendation systems, multimodal (MM) content is increasingly introduced to enhance the generalization of ID features. The rise of Multimodal Large Language Models (MLLMs) enables the construction of unified user and item representations. However, the semantic distribution gap between MM and ID representations leads to \textit{convergence inconsistency} during joint training: the ID branch converges quickly, while the MM branch requires more epochs, thus limiting overall performance. To address this, we propose a two-stage framework including MM representation learning and joint training optimization. First, we fine-tune the MLLM to generate unified user and item representations, and introduce collaborative signals by post-aligning user ID representations to alleviate semantic differences. Then, we propose an Adaptive Gradient Masking (AGM) training strategy to dynamically regulate parameter updates between ID and MLLM branches. AGM estimates the contribution of each representation with mutual information, and applies non-uniform gradient masking at the sub-network level to balance optimization. We provide theoretical analysis of AGM's effectiveness and further introduce an unbiased variant, AGM*, to enhance training stability. Experiments on offline and online A/B tests validate the effectiveness of our approach in mitigating convergence inconsistency and improving performance.

Real-world applications are stretching context windows to hundreds of thousand of tokens while Large Language Models (LLMs) swell from billions to trillions of parameters. This dual expansion send compute and memory costs skyrocketing, making $\textit{token compression}$ indispensable. We introduce Vision Centric Token Compression ($\textbf{Vist}$), a $\textit{slow–fast}$ compression framework that mirrors human reading: the $\textit{fast}$ path renders distant tokens into images, letting a $\textbf{frozen, lightweight vision encoder}$ skim the low-salience context; the $\textit{slow}$ path feeds the proximal window into the LLM for fine-grained reasoning. A Probability-Informed Visual Enhancement (PVE) objective masks high-frequency tokens during training, steering the Resampler to concentrate on semantically rich regions—just as skilled reader gloss over function words. On eleven in-context learning benchmarks, $\textbf{Vist}$ achieves the same accuracy with 2.3$\times$ fewer tokens, cutting FLOPs by 16\% and memory by 50\%. This method delivers remarkable results, outperforming the strongest text encoder-based compression method CEPE by $\textbf{7.6}$\% on average over benchmarks like TriviaQA, NQ, PopQA, NLUI, and CLIN, setting a new standard for token efficiency in LLMs. The source code will be released.

Language models (LMs) and their extension, vision-language models (VLMs), have achieved remarkable performance across various tasks. However, they still struggle with complex reasoning tasks that require multimodal or multilingual real world knowledge. To support such capabilities, an external memory system that can efficiently provide relevant multimodal information is essential. Existing approaches generally concatenate image and text tokens into a long sequence as memory, which, however, may drastically increase context length and even degrade performance. In contrast, we propose using continuous memory-a compact set of dense embeddings-to more effectively and efficiently represent multimodal and multilingual knowledge. Our key insight is that a VLM can serve as its own continuous memory encoder. We empirically show that this design improves performance on complex multimodal reasoning tasks. Building on this, we introduce a data-efficient and parameter-efficient method to fine-tune the VLM into a memory encoder, requiring only 1.2\% of the model’s parameters and a small corpus of 15.6K self-synthesized samples. Our approach CoMEM utilizes VLM's original capabilities to encode arbitrary multimodal and multilingual knowledge into just 8 continuous embeddings. Since the inference-time VLM remains frozen, our memory module is plug-and-play and can be flexibly integrated as needed. Extensive experiments across eight multimodal reasoning benchmarks demonstrate the effectiveness of our approach. Code and data is publicly released here https://github.com/WenyiWU0111/CoMEM.

While Multimodal Large Language Models (MLLMs) have advanced GUI navigation agents, current approaches face limitations in cross-domain generalization and effective history utilization. We present a reasoning-enhanced framework that systematically integrates structured reasoning, action prediction, and history summarization. The structured reasoning component generates coherent Chain-of-Thought analyses combining progress estimation and decision reasoning, which inform both immediate action predictions and compact history summaries for future steps. Based on this framework, we train a GUI agent, GUI-Rise, through supervised fine-tuning on pseudo-labeled trajectories and reinforcement learning with Group Relative Policy Optimization (GRPO). This framework employs specialized rewards, including a history-aware objective, directly linking summary quality to subsequent action performance. Comprehensive evaluations on standard benchmarks demonstrate state-of-the-art results under identical training data conditions, with particularly strong performance in out-of-domain scenarios. These findings validate our framework's ability to maintain robust reasoning and generalization across diverse GUI navigation tasks.

The emergence of advanced multimodal large language models (MLLMs) has significantly enhanced AI assistants' ability to process complex information across modalities. Recently, egocentric videos, by directly capturing user focus, actions, and context in an unified coordinate, offer an exciting opportunity to enable proactive and personalized AI user experiences with MLLMs. However, existing benchmarks overlook the crucial role of gaze as an indicator of user intent. To address this gap, we introduce EgoGazeVQA, an egocentric gaze-guided video question answering benchmark that leverages gaze information to improve the understanding of longer daily-life videos. EgoGazeVQA consists of gaze-based QA pairs generated by MLLMs and refined by human annotators. Our experiments reveal that existing MLLMs struggle to accurately interpret user intentions using only global visual tokens. In contrast, our gaze-guided intent prompting methods significantly enhance performance by integrating spatial, temporal, and intent-related cues. We further conduct experiments on gaze-related fine-tuning and analyze how gaze estimation accuracy impacts prompting effectiveness. These results underscore the value of gaze for more personalized and effective AI assistants in egocentric settings.

While image generation techniques are now capable of producing high-quality images that respect prompts which span multiple sentences, the task of text-guided image editing remains a challenge. Even edit requests that consist of only a few words often fail to be executed correctly. We explore three strategies to enhance performance on a wide range of image editing tasks: supervised fine-tuning (SFT), reinforcement learning (RL), and Chain-of-Thought (CoT) reasoning. In order to study all these components in one consistent framework, we adopt an autoregressive multimodal model that processes textual and visual tokens in a unified manner. We find RL combined with a large multi-modal LLM verifier to be the most effective of these strategies. As a result, we release **EARL**: **E**diting with **A**utoregression and **RL**, a strong RL-based image editing model that performs competitively on a diverse range of edits compared to strong baselines, despite using much less training data. Thus, EARL pushes the frontier of autoregressive multimodal models on image editing. We release our code, training data, and trained models at [https://github.com/mair-lab/EARL](https://github.com/mair-lab/EARL).

Large Language Models (LLMs) are predominantly evaluated on Standard American English (SAE), often overlooking the diversity of global English varieties. This narrow focus may raise fairness concerns as degraded performance on non-standard varieties can lead to unequal benefits for users worldwide. Therefore, it is critical to extensively evaluate the linguistic robustness of LLMs on multiple non-standard English varieties. We introduce Trans-EnV, a framework that automatically transforms SAE datasets into multiple English varieties to evaluate the linguistic robustness. Our framework combines (1) linguistics expert knowledge to curate variety-specific features and transformation guidelines from linguistic literature and corpora, and (2) LLM-based transformations to ensure both linguistic validity and scalability. Using Trans-EnV, we transform six benchmark datasets into 38 English varieties and evaluate seven state-of-the-art LLMs. Our results reveal significant performance disparities, with accuracy decreasing by up to 46.3% on non-standard varieties. These findings highlight the importance of comprehensive linguistic robustness evaluation across diverse English varieties. Each construction of Trans-EnV was validated through rigorous statistical testing and consultation with a researcher in the field of second language acquisition, ensuring its linguistic validity. Our code and datasets are publicly available.

Driven by the applications in autonomous driving, robotics, and augmented reality, 3D object annotation is a critical task compared to 2D annotation, such as spatial complexity, occlusion, and viewpoint inconsistency. The existing methods relying on single models often struggle with these issues. In this paper, we introduce Tri-MARF, a novel framework that integrates tri-modal inputs (i.e., 2D multi-view images, text descriptions, and 3D point clouds) with multi-agent collaboration to enhance the 3D annotation process. Our Tri-MARF consists of three specialized agents: a vision-language model agent that generates multi-view descriptions, an information aggregation agent that selects optimal descriptions, and a gating agent that aligns text descriptions with 3D geometries for more refined captioning. Extensive experiments on the Objaverse-LVIS, Objaverse-XL, and ABO datasets demonstrate the superiority of our Tri-MARF, which achieves a CLIPScore of 88.7 (compared to 78.6–82.4 for other SOTA methods), retrieval accuracy of 45.2/43.8 (ViLT R@5), and an impressive throughput of 12,000 objects per hour on a single NVIDIA A100 GPU.

LLMs have demonstrated unprecedented capabilities in natural language processing, yet their practical deployment remains hindered by persistent factuality and faithfulness hallucinations. While existing methods address these hallucination types independently, they inadvertently induce performance trade-offs, as interventions targeting one type often exacerbate the other. Through empirical and theoretical analysis of activation space dynamics in LLMs, we reveal that these hallucination categories share overlapping subspaces within neural representations, presenting an opportunity for concurrent mitigation. To harness this insight, we propose SPACE, a unified framework that jointly enhances factuality and faithfulness by editing shared activation subspaces. SPACE establishes a geometric foundation for shared subspace existence through dual-task feature modeling, then identifies and edits these subspaces via a hybrid probe strategy combining spectral clustering and attention head saliency scoring. Experimental results across multiple benchmark datasets demonstrate the superiority of our approach.

Preference optimization for diffusion models aims to align them with human preferences for images. Previous methods typically use Vision-Language Models (VLMs) as pixel-level reward models to approximate human preferences. However, when used for step-level preference optimization, these models face challenges in handling noisy images of different timesteps and require complex transformations into pixel space. In this work, we show that pre-trained diffusion models are naturally suited for step-level reward modeling in the noisy latent space, as they are explicitly designed to process latent images at various noise levels. Accordingly, we propose the **Latent Reward Model (LRM)**, which repurposes components of the diffusion model to predict preferences of latent images at arbitrary timesteps. Building on LRM, we introduce **Latent Preference Optimization (LPO)**, a step-level preference optimization method conducted directly in the noisy latent space. Experimental results indicate that LPO significantly improves the model's alignment with general, aesthetic, and text-image alignment preferences, while achieving a 2.5-28x training speedup over existing preference optimization methods.

Multimodal models often over-rely on dominant modalities, failing to achieve optimal performance. While prior work focuses on modifying training objectives or optimization procedures, data-centric solutions remain underexplored. We propose MIDAS, a novel data augmentation strategy that generates misaligned samples with semantically inconsistent cross-modal information, labeled using unimodal confidence scores to compel learning from contradictory signals. However, this confidence-based labeling can still favor the more confident modality. To address this within our misaligned samples, we introduce weak-modality weighting, which dynamically increases the loss weight of the least confident modality, thereby helping the model fully utilize weaker modality. Furthermore, when misaligned features exhibit greater similarity to the aligned features, these misaligned samples pose a greater challenge, thereby enabling the model to better distinguish between classes. To leverage this, we propose hard-sample weighting, which prioritizes such semantically ambiguous misaligned samples. Experiments on multiple multimodal classification benchmarks demonstrate that MIDAS significantly outperforms related baselines in addressing modality imbalance.

Natural disasters cause devastating damage to communities and infrastructure every year. Effective disaster response is hampered by the difficulty of accessing affected areas during and after events. Remote sensing has allowed us to monitor natural disasters in a remote way. More recently there have been advances in computer vision and deep learning that help automate satellite imagery analysis, However, they remain limited by their narrow focus on specific disaster types, reliance on manual expert interpretation, and lack of datasets with sufficient temporal granularity or natural language annotations for tracking disaster progression. We present MONITRS, a novel multimodal dataset of $\sim$10,000 FEMA disaster events with temporal satellite imagery with natural language annotations from news articles, accompanied by geotagged locations, and question-answer pairs. We demonstrate that fine-tuning existing MLLMs on our dataset yields significant performance improvements for disaster monitoring tasks, establishing a new benchmark for machine learning-assisted disaster response systems.

Referring Image Segmentation (RIS) aims to segment the target object in an image given a natural language expression. While recent methods leverage pre-trained vision backbones and more training corpus to achieve impressive results, they predominantly focus on simple expressions—short, clear noun phrases like “red car” or “left girl”. This simplification often reduces RIS to a key word/concept matching problem, limiting the model’s ability to handle referential ambiguity in expressions. In this work, we identify two challenging real-world scenarios: object-distracting expressions, which involve multiple entities with contextual cues, and category-implicit expressions, where the object class is not explicitly stated. To address the challenges, we propose a novel framework, SaFiRe, which mimics the human two-phase cognitive process—first forming a global understanding, then refining it through detail-oriented inspection. This is naturally supported by Mamba’s scan-then-update property, which aligns with our phased design and enables efficient multi-cycle refinement with linear complexity. We further introduce aRefCOCO, a new benchmark designed to evaluate RIS models under ambiguous referring expressions. Extensive experiments on both standard and proposed datasets demonstrate the superiority of SaFiRe over state-of-the-art baselines.