NeurIPS 2025 论文列表

Recent advances in language modeling and vision stem from training large models on diverse, multi‑task data. This paradigm has had limited impact in value-based reinforcement learning (RL), where improvements are often driven by small models trained in a single-task context. This is because in multi-task RL sparse rewards and gradient conflicts make optimization of temporal difference brittle. Practical workflows for generalist policies therefore avoid online training, instead cloning expert trajectories or distilling collections of single‑task policies into one agent. In this work, we show that the use of high-capacity value models trained via cross-entropy and conditioned on learnable task embeddings addresses the problem of task interference in online RL, allowing for robust and scalable multi‑task training. We test our approach on 7 multi-task benchmarks with over 280 unique tasks, spanning high degree-of-freedom humanoid control and discrete vision-based RL. We find that, despite its simplicity, the proposed approach leads to state-of-the-art single and multi-task performance, as well as sample-efficient transfer to new tasks.

Model merging integrates multiple expert models with diverse capabilities into a unified framework, facilitating collaborative learning. However, most existing methods assume simultaneous access to all models, which is often impractical in real-world scenarios where models are received sequentially. While some studies have investigated continual model merging (CMM)--which involves sequentially merging multiple models--the challenge of balancing prior knowledge (stability) and incorporating new tasks (plasticity) remains unresolved. This paper, for the first time, formally defines the stability and plasticity of CMM from the perspective of orthogonal projection. Subsequently, we analyze the relationships among the spaces spanned by task data, historical gradients, and accumulated gradients. Building on this, we propose a data-free \textbf{D}ual \textbf{O}rthogonal \textbf{P}rojection (DOP) method, which eliminates data dependence and mitigates interference between the merged model and models for old and new tasks by projecting their parameter differences onto their respective approximate data spaces. Finally, to solve potential conflicts between stability and plasticity, we reformulate DOP as a multi-objective optimization problem and employ a multi-gradient descent algorithm to obtain a Pareto-optimal solution. Extensive experiments across multiple architectures and task configurations validate that our approach significantly outperforms state-of-the-art CMM methods.

Foundation models trained at scale exhibit remarkable emergent behaviors, learning new capabilities beyond their initial training objectives. We find such emergent behaviors in biological vision models via large-scale contrastive vision-language training. To achieve this, we first curate TreeOfLife-200M, comprising 214 million images of living organisms, the largest and most diverse biological organism image dataset to date. We then train BioCLIP 2 on TreeOfLife-200M to distinguish different species. Despite the narrow training objective, BioCLIP 2 yields extraordinary accuracy when applied to various biological visual tasks such as habitat classification and trait prediction. We identify emergent properties in the learned embedding space of BioCLIP 2. At the inter-species level, the embedding distribution of different species aligns closely with functional and ecological meanings (e.g., beak sizes and habitats). At the intra-species level, instead of being diminished, the intra-species variations (e.g., life stages and sexes) are preserved and better separated in subspaces orthogonal to inter-species distinctions. We provide formal proof and analyses to explain why hierarchical supervision and contrastive objectives encourage these emergent properties. Crucially, our results reveal that these properties become increasingly significant with larger-scale training data, leading to a biologically meaningful embedding space.

Prompt tuning has emerged as a key technique for adapting large pre-trained Decision Transformers (DTs) in offline Reinforcement Learning (RL), particularly in multi-task and few-shot settings. The Prompting Decision Transformer (PDT) enables task generalization via trajectory prompts sampled uniformly from expert demonstrations -- without accounting for prompt informativeness. In this work, we propose a bandit-based prompt-tuning method that learns to construct optimal trajectory prompts from demonstration data at inference time. We devise a structured bandit architecture operating in the trajectory prompt space, achieving linear rather than combinatorial scaling with prompt size. Additionally, we show that the pre-trained PDT itself can serve as a powerful feature extractor for the bandit, enabling efficient reward modeling across various environments. We theoretically establish regret bounds and demonstrate empirically that our method consistently enhances performance across a wide range of tasks, high-dimensional environments, and out-of-distribution scenarios, outperforming existing baselines in prompt tuning.

Learning-to-learn or meta-learning focuses on developing algorithms that leverage prior experience to quickly acquire new skills or adapt to novel environments. A crucial component of meta-learning is representation learning, which aims to construct data representations capable of transferring knowledge across multiple tasks—a critical advantage in data-scarce settings. We study how representation learning can improve the efficiency of bandit problems. We consider $T$ $d$-dimensional linear bandits that share a common low-dimensional linear representation. We provide provably fast, sample-efficient algorithms to address the two key problems in meta-learning: (1) learning a common set of features from multiple related bandit tasks and (2) transferring this knowledge to new, unseen bandit tasks. We validated the theoretical results through numerical experiments using real-world and synthetic datasets, comparing them against benchmark algorithms.

Evolutionary multi-view classification (EMVC) methods have gained wide recognition due to their adaptive mechanisms. Fitness evaluation (FE), which aims to calculate the classification performance of each individual in the population and provide reliable performance ranking for subsequent operations, is a core step in such methods. Its accuracy directly determines the correctness of the evolutionary direction. However, when FE fails to correctly reflect the superiority-inferiority relationship among individuals, it will lead to confusion in individual performance ranking, which in turn misleads the evolutionary direction and results in trapping into local optima. This paper is the first to identify the aforementioned issue in the field of EMVC and call it as fitness evaluation bias (FEB). FEB may be caused by a variety of factors, and this paper approaches the issue from the perspective of view information content: existing methods generally adopt joint training strategies, which restrict the exploration of key information in views with low information content. This makes it difficult for multi-view model (MVM) to achieve optimal performance during convergence, which in turn leads to FE failing to accurately reflect individual performance rankings and ultimately triggering FEB. To address this issue, we propose an evolutionary multi-view classification via eliminating individual fitness bias (EFB-EMVC) method, which alleviates the FEB issue by introducing evolutionary navigators for each MVM, thereby providing more accurate individual ranking. Experimental results fully verify the effectiveness of the proposed method in alleviating the FEB problem, and the EMVC method equipped with this strategy exhibits more superior performance compared with the original EMVC method. (The code is available at https://github.com/LiShuailzn/Neurips-2025-EFB-EMVC)

Modern foundation models are trained on diverse datasets to enhance generalization across tasks and domains. A central challenge in this process is determining how to effectively mix and sample data from multiple sources. This naturally leads to a multi-task learning (MTL) perspective. While prior work in MTL has emphasized mitigating gradient conflicts, we observe that large-scale pretraining scenarios—such as multilingual or multi-domain training—often exhibit little to no gradient conflict. Motivated by this observation, we propose $\textbf{PiKE}$ ($\textbf{P}$ositive gradient $\textbf{i}$nteraction-based $\textbf{K}$-task weights $\textbf{E}$stimator), an adaptive data mixing algorithm that dynamically adjusts sampling weights during training. PiKE exploits non-conflicting gradient interactions to minimize a near-tight upper bound on the average loss decrease at each step, while incurring negligible computational overhead. We provide theoretical convergence guarantees and show that PiKE outperforms static and non-adaptive mixing baselines. Furthermore, we extend PiKE to promote balanced learning across tasks. Extensive experiments on large-scale language model pretraining confirm that PiKE achieves faster convergence and improved downstream performance compared to existing approaches.

The exponential growth of data-driven systems and AI technologies has intensified the demand for high-quality web-sourced datasets. While existing datasets have proven valuable, conventional web data collection approaches face significant limitations in terms of human effort and scalability. Current data collecting solutions fall into two categories: wrapper-based methods that struggle with adaptability and reproducibility, and large language model (LLM)-based approaches that incur substantial computational and financial costs. To address these challenges, we propose AutoData, a novel multi-agent system for Automated web Data collection, that requires minimal human intervention, i.e., only necessitating a natural language instruction specifying the desired dataset. In addition, AutoData is designed for a robust multi-agent architecture, featuring a novel oriented message hypergraph coordinated by a central task manager, to efficiently organize agents across research and development squads. Besides, we introduce a novel hypergraph cache system to advance the multi-agent collaboration process that enables efficient automated data collection and mitigates the token cost issues prevalent in existing LLM-based systems. Moreover, we introduce Instruct2DS, a new benchmark dataset supporting live data collection from web sources across three domains: academic, finance, and sports. Comprehensive evaluations over Instruct2DS and three existing benchmark datasets demonstrate AutoData's superior performance compared to baseline methods. Case studies on challenging tasks such as picture book collection and paper extraction from surveys further validate its applicability.

Casting complex inputs into tractable representations is a critical step across various fields. Diverse embedding models emerge from differences in architectures, loss functions, input modalities and datasets, each capturing unique aspects of the input. Multi-teacher distillation leverages this diversity to enrich representations but often remains tailored to specific tasks. We introduce a task-agnostic framework based on a ``majority vote" objective function. We demonstrate that this function is bounded by the mutual information between the student and the teachers' embeddings, leading to a task-agnostic distillation loss that eliminates dependence on task-specific labels or prior knowledge. Comprehensive evaluations across text, vision models, and molecular modeling show that our method effectively leverages teacher diversity, resulting in representations enabling better performance for a wide range of downstream tasks such as classification, clustering, or regression. Additionally, we train and release state-of-the-art embedding models, enhancing downstream performance in various modalities.

Preference-based data often appear complex and noisy but may conceal underlying homogeneous structures. This paper introduces a novel framework of ranking structure recognition for preference-based data. We first develop an approach to identify dynamic ranking groups by incorporating temporal penalties into a spectral estimation for the celebrated Bradley-Terry model. To detect structural changes, we introduce an innovative objective function and present a practicable algorithm based on dynamic programming. Theoretically, we establish the consistency of ranking group recognition by exploiting properties of a random 'design matrix' induced by a reversible Markov chain. We also tailor a group inverse technique to quantify the uncertainty in item ability estimates. Additionally, we prove the consistency of structure change recognition, ensuring the robustness of the proposed framework. Experiments on both synthetic and real-world datasets demonstrate the practical utility and interpretability of our approach.

View missing remains a significant challenge in graph-based multi-view semi-supervised learning, hindering their real-world applications. To address this issue, traditional methods introduce a missing indicator matrix and focus on mining partial structure among existing samples in each view for label propagation (LP). However, we argue that these disregarded missing samples sometimes induce discontinuous local structures, i.e., sub-clusters, breaking the fundamental smoothness assumption in LP. Consequently, such a Sub-Cluster Problem (SCP) would distort graph fusion and degrade classification performance. To alleviate SCP, we propose a novel incomplete multi-view semi-supervised learning method, termed AGF-TI. Firstly, we design an adversarial graph fusion scheme to learn a robust consensus graph against the distorted local structure through a min-max framework. By stacking all similarity matrices into a tensor, we further recover the incomplete structure from the high-order consistency information based on the low-rank tensor learning. Additionally, the anchor-based strategy is incorporated to reduce the computational complexity. An efficient alternative optimization algorithm combining a reduced gradient descent method is developed to solve the formulated objective, with theoretical convergence. Extensive experimental results on various datasets validate the superiority of our proposed AGF-TI as compared to state-of-the-art methods. Code is available at https://github.com/ZhangqiJiang07/AGF_TI.

Meta-learning has been demonstrated to be useful to improve the sampling efficiency of Bayesian optimization (BO) and surrogate-assisted evolutionary algorithms (SAEAs) when solving expensive optimization problems (EOPs). Existing studies mainly focus on either combinations of existing meta-learning modeling methods with optimization algorithms, or the development of meta-learning acquisition functions for specific meta BO. However, the meta-learning models used in the literature are not designed for optimization purpose, and the generalization ability of meta-learning acquisition functions is limited. In this work, we develop a novel architecture of meta-learning model for optimization purpose and propose a generalized few-shot evolutionary optimization (FSEO) framework to solve EOPs. We focus on the scenario of expensive multi-objective EOPs (EMOPs) in the context of few-shot optimization as there are few studies on it and its high requirement on surrogate modeling performance. The surrogates in FSEO framework combines neural network with Gaussian Processes (GPs), their network parameters and some parameters of GPs represent task-independent experience and are meta-learned across related optimization tasks, the remaining GPs parameters are task-specific parameters that represent unique features of the target task. We demonstrate that our FSEO framework is able to improve the sampling efficiency of existing SAEAs on EMOPs.

Non-stationary multi-armed bandits (nsMAB) enable agents to adapt to changing environments by incorporating mechanisms to detect and respond to shifts in reward distributions, making them well-suited for dynamic settings. However, existing approaches typically assume that reward feedback is available at every round—an assumption that overlooks many real-world scenarios where feedback is limited. In this paper, we take a significant step forward by introducing a new model of *constrained feedback in non-stationary multi-armed bandits* (ConFee-nsMAB), where the availability of reward feedback is restricted. We propose the first prior-free algorithm—that is, one that does not require prior knowledge of the degree of non-stationarity—that achieves near-optimal dynamic regret in this setting. Specifically, our algorithm attains a dynamic regret of $\tilde {\mathcal{O}}({K^{1/3} V_T^{1/3} T }/{ B^{1/3}})$, where $T$ is the number of rounds, $K$ is the number of arms, $B$ is the query budget, and $V_T$ is the variation budget capturing the degree of non-stationarity.

Recent work indicates that low-dimensional dynamics of neural and behavioral data are often preserved across days and subjects. However, extracting these preserved dynamics remains challenging: high-dimensional neural population activity and the recorded neuron populations vary across recording sessions. While existing modeling tools can improve alignment between neural and behavioral data, they often operate on a per-subject basis or discretize behavior into categories, disrupting its natural continuity and failing to capture the underlying dynamics. We introduce $\underline{\text{C}}$ontrastive $\underline{\text{A}}$ligned $\underline{\text{N}}$eural $\underline{\text{D}}$$\underline{\text{Y}}$namics (CANDY), an end‑to‑end framework that aligns neural and behavioral data using rank-based contrastive learning, adapted for continuous behavioral variables, to project neural activity from different sessions onto a shared low-dimensional embedding space. CANDY fits a shared linear dynamical system to the aligned embeddings, enabling an interpretable model of the conserved temporal structure in the latent space. We validate CANDY on synthetic and real-world datasets spanning multiple species, behaviors, and recording modalities. Our results show that CANDY is able to learn aligned latent embeddings and preserved dynamics across neural recording sessions and subjects, and it achieves improved cross-session behavior decoding performance. We further show that the latent linear dynamical system generalizes to new sessions and subjects, achieving comparable or even superior behavior decoding performance to models trained from scratch. These advances enable robust cross‑session behavioral decoding and offer a path towards identifying shared neural dynamics that underlie behavior across individuals and recording conditions. The code and two-photon imaging data of striatal neural activity that we acquired here are available at https://github.com/schnitzer-lab/CANDY-public.git.

Multi-label complementary label learning (MLCLL) is a weakly supervised paradigm that addresses multi-label learning (MLL) tasks using complementary labels (i.e., irrelevant labels) instead of relevant labels. Existing methods typically adopt an unbiased risk estimator (URE) under the assumption that complementary labels follow a uniform distribution. However, this assumption fails in real-world scenarios due to instance-specific annotation biases, making URE-based methods ineffective under such conditions. Furthermore, existing methods underutilize label correlations inherent in MLL. To address these limitations, we propose ComRank, a ranking loss framework for MLCLL, which encourages complementary labels to be ranked lower than non-complementary ones, thereby modeling pairwise label relationships. Theoretically, our surrogate loss ensures Bayes consistency under both uniform and biased cases. Experiments demonstrate the effectiveness of our method in MLCLL tasks. The code is available at https://github.com/JellyJamZhu/ComRank.

The need for training multilingual multi-task speech processing (MSP) models that perform both automatic speech recognition and speech-to-text translation is increasingly evident. However, a significant challenge arises from the conflicts among multiple objectives when using a single model. Multi-objective optimization can address this challenge by facilitating the optimization of multiple conflicting objectives and aligning the gradient updates in a common descent direction. While multi-objective optimization helps avoid conflicting gradient updates, a critical issue is that when there are many objectives, such as in MSP, it is often {\em difficult to find} a common descent direction. This leads to an important question: Is it more effective to separate highly conflicting objectives into different optimization levels or to keep them in a single level? To address this question, this paper investigates three multi-objective MSP formulations, which we refer to as \textbf{objective soup recipes}. These formulations apply multi-objective optimization at different optimization levels to mitigate potential conflicts among all objectives. To keep computation and memory overhead low, we incorporate a lightweight layer‑selection strategy that detects the most conflicting layers and uses only their gradients when computing the conflict‑avoidance direction. We conduct an extensive investigation using the CoVoST v2 dataset for combined multilingual ASR and ST tasks, along with the LibriSpeech and AISHELL-1 datasets for multilingual ASR, to identify highly conflicting objectives and determine the most effective training recipe among the three proposed multi-objective optimization algorithms.

Reward Models (RMs) are crucial to aligning large language models (LLMs), but the degree to which an RM specialized to one task (e.g. writing) generalizes to new tasks (e.g. math) is often not known a priori, often making using only one fixed RM to train LLMs suboptimal. However, optimizing LLMs with multiple RMs simultaneously can incur a prohibitively high computational cost and lead to conflicting signals from different RMs that may degrade performance. To address these challenges, we introduce LASeR (Learning to Adaptively Select Rewards), which frames reward model selection as a multi-armed bandit problem, iteratively and efficiently training LLMs using multiple RMs by selecting the most well-suited RM for each instance. On commonsense and math reasoning tasks, we show that LASeR boosts iterative LLM training, improving the absolute average accuracy of Llama-3-8B over three datasets by $2.67$% over an ensemble of RM scores while also showing superior efficiency (e.g., a $2\times$ speedup). Moreover, on WildChat (open-ended instruction-following tasks), LASeR leads to a $72.69$% AlpacaEval win rate over the RM score ensemble baseline. Extending to long-context generation, LASeR improves by $2.96$ F1 points (avg.) on single-document QA tasks and $2.97$ F1 points on few-shot learning over the RM score ensemble baseline with best-of-$n$ sampling. We include our code in the supplementary.

Meta-learning enables learning systems to adapt quickly to new tasks, similar to humans. Different meta-learning approaches all work under/with the mini-batch episodic training framework. Such framework naturally gives the information about task identity, which can serve as additional supervision for meta-training to improve generalizability. We propose to exploit task identity as additional supervision in meta-training, inspired by the alignment and discrimination ability which is is intrinsic in human's fast learning. This is achieved by contrasting what meta-learners learn, i.e., model representations. The proposed ConML is evaluating and optimizing the contrastive meta-objective under a problem- and learner-agnostic meta-training framework. We demonstrate that ConML integrates seamlessly with existing meta-learners, as well as in-context learning models, and brings significant boost in performance with small implementation cost.

Tables and table-based use cases play a crucial role in many important real-world applications, such as spreadsheets, databases, and computational notebooks, which traditionally require expert-level users like data engineers, data analysts, and database administrators to operate. Although LLMs have shown remarkable progress in working with tables (e.g., in spreadsheet and database copilot scenarios), comprehensive benchmarking of such capabilities remains limited. In contrast to an extensive and growing list of NLP benchmarks, evaluations of table-related tasks are scarce, and narrowly focus on tasks like NL-to-SQL and Table-QA, overlooking the broader spectrum of real-world tasks that professional users face. This gap limits our understanding and model progress in this important area. In this work, we introduce MMTU, a large-scale benchmark with over 28K questions across 25 real-world table tasks, designed to comprehensively evaluate models ability to understand, reason, and manipulate real tables at the expert-level. These tasks are drawn from decades’ worth of computer science research on tabular data, with a focus on complex table tasks faced by professional users. We show that MMTU require a combination of skills -- including table understanding, reasoning, and coding -- that remain challenging for today's frontier models, where even frontier reasoning models like OpenAI GPT-5 and DeepSeek R1 score only around 69% and 57% respectively, suggesting significant room for improvement. We highlight key findings in our evaluation using MMTU and hope this benchmark drives further advances in understanding and developing foundation models for structured data processing and analysis. Our code and data are available at https://github.com/MMTU-Benchmark/MMTU and https://huggingface.co/datasets/MMTU-benchmark/MMTU.