NeurIPS 2025 Papers

In this paper, we consider fully decentralized cooperative multi-agent reinforcement learning, where each agent has access only to the states, its local actions, and the shared rewards. The absence of information about other agents' actions typically leads to the non-stationarity problem during per-agent value function updates, and the relative overgeneralization issue during value function estimation. However, existing works fail to address both issues simultaneously, as they lack the capability to model the agents' joint policy in a fully decentralized setting. To overcome this limitation, we propose a simple yet effective method named Return-Aware Context (RAC). RAC formalizes the dynamically changing task, as locally perceived by each agent, as a contextual Markov Decision Process (MDP), and addresses both non-stationarity and relative overgeneralization through return-aware context modeling. Specifically, the contextual MDP attributes the non-stationary local dynamics of each agent to switches between contexts, each corresponding to a distinct joint policy. Then, based on the assumption that the joint policy changes only between episodes, RAC distinguishes different joint policies by the training episodic return and constructs contexts using discretized episodic return values. Accordingly, RAC learns a context-based value function for each agent to address the non-stationarity issue during value function updates. For value function estimation, an individual optimistic marginal value is constructed to encourage the selection of optimal joint actions, thereby mitigating the relative overgeneralization problem. Experimentally, we evaluate RAC on various cooperative tasks (including matrix game, predator and prey, and SMAC), and its significant performance validates its effectiveness.

Large language model (LLM) based agentic workflows have become a popular paradigm for coordinating multiple specialized agents to solve complex tasks. To improve serving efficiency, existing LLM systems employ prefix caching to reuse key-value (KV) tensors corresponding to agents' fixed prompts, thereby avoiding redundant computation across repeated invocations. However, current systems typically evict KV caches using a Least Recently Used (LRU) policy, which fails to anticipate future agent usage and often discards KV caches shortly before their reuse. This leads to frequent cache misses and substantial recomputation or swap- ping overhead. We present KVFlow, a workflow-aware KV cache management framework tailored for agentic workloads. KVFlow abstracts the agent execution schedule as an Agent Step Graph and assigns each agent a steps-to-execution value that estimates its temporal proximity to future activation. These values guide a fine-grained eviction policy at the KV node level, allowing KVFlow to preserve entries likely to be reused and efficiently manage shared prefixes in tree-structured caches. Moreover, KVFlow introduces a fully overlapped KV prefetching mecha- nism, which proactively loads required tensors from CPU to GPU in background threads for agents scheduled in the next step, thereby avoiding cache miss stalls during generation. Compared to SGLang with hierarchical radix cache, KVFlow achieves up to 1.83× speedup for single workflows with large prompts, and up to 2.19× speedup for scenarios with many concurrent workflows.

Accurately extracting protocol-state machines (PSMs) from the long, densely written Request-for-Comments (RFC) standards that govern Internet‐scale communication remains a bottleneck for automated security analysis and protocol testing. In this paper, we introduce RFC2PSM, the first large-scale dataset that pairs 1,580 pages of cleaned RFC text with 108 manually validated states and 297 transitions covering 14 widely deployed protocols spanning the data-link, transport, session, and application layers. Built on this corpus, we propose PsmBench, a benchmark that (i) feeds chunked RFC to an LLM, (ii) prompts the model to emit a machine-readable PSM, and (iii) scores the output with structure-aware, semantic fuzzy-matching metrics that reward partially correct graphs. A comprehensive baseline study of nine state-of-the-art open and commercial LLMs reveals a persistent state–transition gap: models identify many individual states (up to $0.82$ F1) but struggle to assemble coherent transition graphs ($\leq 0.38$ F1), highlighting challenges in long-context reasoning, alias resolution, and action/event disambiguation. We release the dataset, evaluation code, and all model outputs as open-sourced, providing a fully reproducible starting point for future work on reasoning over technical prose and generating executable graph structures. RFC2PSM and PsmBench aim to catalyze cross-disciplinary progress toward LLMs that can interpret and verify the protocols that keep the Internet safe.

We introduce a novel framework for efficient sampling from complex, unnormalised target distributions by exploiting multiscale dynamics. Traditional score-based sampling methods either rely on learned approximations of the score function or involve computationally expensive nested Markov chain Monte Carlo (MCMC) loops. In contrast, the proposed approach leverages stochastic averaging within a slow-fast system of stochastic differential equations (SDEs) to estimate intermediate scores along a diffusion path without training or inner-loop MCMC. Two algorithms are developed under this framework: *MultALMC*, which uses multiscale annealed Langevin dynamics, and *MultCDiff*, based on multiscale controlled diffusions for the reverse-time Ornstein-Uhlenbeck process. Both overdamped and underdamped variants are considered, with theoretical guarantees of convergence to the desired diffusion path. The framework is extended to handle heavy-tailed target distributions using Student’s t-based noise models and tailored fast-process dynamics. Empirical results across synthetic and real-world benchmarks, including multimodal and high-dimensional distributions, demonstrate that the proposed methods are competitive with existing samplers in terms of accuracy and efficiency, without the need for learned models.

Instruction fine-tuning is crucial in NLP tasks, enhancing pretrained models' instruction-following capabilities and task-specific performance. However, obtaining high-quality fine-tuning data for large models is challenging due to data collection difficulties and high production costs. To address this, we propose MASTER, a novel data augmentation method that enriches original data through interactions among multiple agents with varying cognitive levels. We simulate three pedagogically grounded teaching scenarios, leveraging multi-agent conversations to generate high-quality teacher-student interaction data. Utilizing MASTER, we construct BOOST-QA, a fine-tuning dataset augmented from existing datasets like Orca-Math-200k, ProcQA, and OpenHermes2.5. Experiments show that models fine-tuned with BOOST-QA perform excellently across multiple benchmarks, demonstrating strong multitask generalization. Notably, MASTER significantly improves models' reasoning abilities in complex tasks, providing valuable insights for future research.

Dynamic 3D scene reconstruction from multi-view videos demands representation to model complex deformations at scale. Current Gaussian Splatting based methods often either suffer from significant computation cost due to dense MLP-based modeling or explicit modeling deformation of each Gaussian independently. However, the dynamics of objects within a scene are typically hierarchical and exhibit structural correlations. To leverage these structural priors into the representation, we introduce **TreeSplat**, a **Tree** data structure for deformable Gaussian **Splat**ting. In TreeSplat, as the name suggests, motions of Gaussian are represented hierarchically within a tree. Each node learns coefficients for time-varying basis functions, defining a part of the motion. The full motion for any given Gaussian is then determined by accumulating these transformations along the tree path from its leaf node to the root node. This tree isn't predefined; instead, it is constructed adaptively alongside Gaussian densification, where cloning or splitting a Gaussian correspondingly creates new leaf nodes. One central property of TreeSplat is its mergeability; after optimization during training, the hierarchical motion parameters for each Gaussian can be efficiently consolidated. By performing this merging step before test time, we eliminate the need to traverse the tree explicitly for each Gaussian during rendering. This results in dramatically faster rendering over 200 FPS and compact storage, while maintaining state-of-the-art rendering quality. Experiments on diverse synthetic and real-world datasets validate these advantages.

While conditional diffusion models have achieved remarkable success in various applications, they require abundant data to train from scratch, which is often infeasible in practice. To address this issue, transfer learning has emerged as an essential paradigm in small data regimes. Despite its empirical success, the theoretical underpinnings of transfer learning conditional diffusion models remain unexplored. In this paper, we take the first step towards understanding the sample efficiency of transfer learning conditional diffusion models through the lens of representation learning. Inspired by practical training procedures, we assume that there exists a low-dimensional representation of conditions shared across all tasks. Our analysis shows that with a well-learned representation from source tasks, the sample complexity of target tasks can be reduced substantially. Numerical experiments are also conducted to verify our results.

Differentially private (DP) synthetic data generation is a promising technique for utilizing private datasets that otherwise cannot be exposed for model training or other analytics. While much research literature has focused on generating private unstructured text and image data, in enterprise settings, structured data (e.g., tabular) is more common, often including natural language fields or components. Existing synthetic data evaluation techniques (e.g., FID) struggle to capture the structural properties and correlations of such datasets. In this work, we propose Struct-Bench, a framework and benchmark for evaluating synthetic datasets derived from structured datasets that contain natural language data. The Struct-Bench framework requires users to provide a representation of their dataset structure as a Context-Free Grammar (CFG). Our benchmark comprises 5 real-world and 2 synthetically generated datasets. We show that these datasets demonstrably present a great challenge even for state-of-the-art DP synthetic data generation methods. Struct-Bench provides reference implementations of different metrics and a leaderboard, offering a standardized platform to benchmark and investigate privacy-preserving synthetic data methods. We also present a case study showing how Struct-Bench improves the synthetic data quality of Private Evolution (PE) on structured data. The benchmark and the leaderboard have been publicly made available at https://struct-bench.github.io.

Designing optimal prompts and reasoning processes for large language models (LLMs) on domain-specific tasks is both necessary and challenging in real-world applications. Determining how to integrate domain knowledge, enhance reasoning efficiency, and even provide domain experts with refined knowledge integration hints are particularly crucial yet unresolved tasks. In this research, we propose Evolutionary Graph Optimization for Prompting (EGO-Prompt), an automated framework to designing better prompts, efficient reasoning processes and providing enhanced causal-informed process. EGO-Prompt begins with a general prompt and fault-tolerant initial Semantic Causal Graph (SCG) descriptions, constructed by human experts, which is then automatically refined and optimized to guide LLM reasoning. Recognizing that expert-defined SCGs may be partial or imperfect and that their optimal integration varies across LLMs, EGO-Prompt integrates a novel causal-guided textual gradient process in two steps: first, generating nearly deterministic reasoning guidance from the SCG for each instance, and second, adapting the LLM to effectively utilize the guidance alongside the original input. The iterative optimization algorithm further refines both the SCG and the reasoning mechanism using textual gradients with ground-truth. We tested the framework on real-world public health, transportation and human behavior tasks. EGO-Prompt achieves 7.32\%–12.61\% higher F1 than cutting-edge methods, and allows small models to reach the performence of larger models at under 20\% of the original cost. It also outputs a refined, domain-specific SCG that improves interpretability.

Spectral reconstruction from RGB images often suffers from a metameric dilemma, where distinct spectral distributions map to nearly identical RGB values, making them indistinguishable to current models and leading to unreliable reconstructions. In this paper, we present Diff-Spectra that integrates supervised physics-aware spectral estimation and unsupervised high-fidelity spectral regularization for HSI reconstruction. We first introduce an Adaptive illumiChroma Decoupling (AICD) module to decouple illumination and chrominance information, which learns intrinsic and distinctive feature distributions, thereby mitigating the metameric issue. Then, we incorporate the AICD into a learnable spectral response function (SRF) guided hyperspectral initial estimation mechanism to mimic the physical image formation and thus inject physics-aware reasoning into neural networks, turning an ill-posed problem into a constrained, interpretable task. We also introduce a metameric spectra augmentation method to synthesize comprehensive hyperspectral data to pre-train a Spectral Diffusion Module (SDM), which internalizes the statistical properties of real-world HSI data, enforcing unsupervised high-fidelity regularization on the spectral transitions via inner-loop optimization during inference. Extensive experimental evaluations demonstrate that our Diff-Spectra achieves SOTA performance on both Spectral reconstruction and downstream HSI classification.

Graph Edit Distance (GED) is a fundamental graph similarity metric widely used in various applications. However, computing GED is an NP-hard problem. Recent state-of-the-art hybrid GED solver has shown promising performance by formulating GED as a bipartite graph matching problem, then leveraging a generative diffusion model to predict node matching between two graphs, from which both the GED and its corresponding edit path can be extracted using a traditional algorithm. However, such methods typically rely heavily on ground-truth supervision, where the ground-truth node matchings are often costly to obtain in real-world scenarios. In this paper, we propose GEDRanker, a novel unsupervised GAN-based framework for GED computation. Specifically, GEDRanker consists of a matching-based GED solver and introduces an interpretable preference-aware discriminator. By leveraging preference signals over different node matchings derived from edit path lengths, the discriminator can guide the matching-based solver toward generating high-quality node matching without the need for ground-truth supervision. Extensive experiments on benchmark datasets demonstrate that our GEDRanker enables the matching-based GED solver to achieve near-optimal solution quality without any ground-truth supervision.

Vision Transformers (ViTs) are essential in computer vision but are computationally intensive, too. Model quantization, particularly to low bit-widths like 4-bit, aims to alleviate this difficulty, yet existing Post-Training Quantization (PTQ) and Quantization-Aware Training (QAT) methods exhibit significant limitations. PTQ often incurs substantial accuracy drop, while QAT achieves high accuracy but suffers from prohibitive computational costs, limited generalization to downstream tasks, training instability, and lacking of open-source codebase. To address these challenges, this paper introduces General, Practical, and Lightning Quantization (GPLQ), a novel framework designed for efficient and effective ViT quantization. GPLQ is founded on two key empirical insights: the paramount importance of activation quantization and the necessity of preserving the model's original optimization basin to maintain generalization. Consequently, GPLQ employs a sequential activation-first, weights-later strategy. Stage 1 keeps weights in FP32 while quantizing activations with a feature mimicking loss in only 1 epoch to keep it stay in the same basin, thereby preserving generalization. Stage 2 quantizes weights using a PTQ method. As a result, GPLQ is 100x faster than existing QAT methods, lowers memory footprint to levels even below FP32 training, and achieves 4-bit model performance that is highly competitive with FP32 models in terms of both accuracy on ImageNet and generalization to diverse downstream tasks, including fine-grained visual classification and object detection. We release an easy-to-use open-source toolkit supporting multiple vision tasks at [GPLQ code](https://github.com/wujx2001/GPLQ).

In performative stochastic optimization, decisions can influence the distribution of random parameters, rendering the data-generating process itself decision-dependent. In practice, decision-makers rarely have access to the true distribution map and must instead rely on imperfect surrogate models, which can lead to severely suboptimal solutions under misspecification. Data scarcity or costly collection further exacerbates these challenges in real-world settings. To address these challenges, we propose a distributionally robust framework for performative optimization that explicitly accounts for ambiguity in the decision-dependent distribution. Our framework introduces three modeling paradigms that capture a broad range of applications in machine learning and decision-making under uncertainty. This latter setting has not previously been explored in the performative optimization literature. To tackle the intractability of the resulting nonconvex objectives, we develop an iterative algorithm named repeated robust risk minimization, which alternates between solving a decision-independent distributionally robust optimization problem and updating the ambiguity set based on the previous decision. This decoupling ensures computational tractability at each iteration while enhancing robustness to model uncertainty. We provide reformulations compatible with off-the-shelf solvers and establish theoretical guarantees on convergence and suboptimality. Extensive numerical experiments in strategic classification, revenue management, and portfolio optimization demonstrate significant performance gains over state-of-the-art baselines, highlighting the practical value of our approach.

Inference in both brains and machines can be formalized by optimizing a shared objective: maximizing the evidence lower bound (ELBO) in machine learning, or minimizing variational free energy ($\mathcal{F}$) in neuroscience (ELBO = $-\mathcal{F}$). While this equivalence suggests a unifying framework, it leaves open how inference is implemented in neural systems. Here, we introduce FOND (*Free energy Online Natural-gradient Dynamics*), a framework that derives neural inference dynamics from three principles: (1) natural gradients on $\mathcal{F}$, (2) online belief updating, and (3) iterative refinement. We apply FOND to derive iP-VAE (*iterative Poisson variational autoencoder*), a recurrent spiking neural network that performs variational inference through membrane potential dynamics, replacing amortized encoders with iterative inference updates. Theoretically, iP-VAE yields several desirable features such as emergent normalization via lateral competition, and hardware-efficient integer spike count representations. Empirically, iP-VAE outperforms both standard VAEs and Gaussian-based predictive coding models in sparsity, reconstruction, and biological plausibility, and scales to complex color image datasets such as CelebA. iP-VAE also exhibits strong generalization to out-of-distribution inputs, exceeding hybrid iterative-amortized VAEs. These results demonstrate how deriving inference algorithms from first principles can yield concrete architectures that are simultaneously biologically plausible and empirically effective.

In recent years, multicalibration has emerged as a desirable learning objective for ensuring that a predictor is calibrated across a rich collection of overlapping subpopulations. Existing approaches typically achieve multicalibration by discretizing the predictor's output space and iteratively adjusting its output values. However, this discretization approach departs from the standard empirical risk minimization (ERM) pipeline, introduces rounding error and an additional sensitive hyperparameter, and may distort the predictor’s outputs in ways that hinder downstream decision-making. In this work, we propose a discretization-free multicalibration method that directly optimizes an empirical risk objective over an ensemble of depth-two decision trees. Our ERM approach can be implemented using off-the-shelf tree ensemble learning methods such as LightGBM. Our algorithm provably achieves multicalibration, provided that the data distribution satisfies a technical condition we term as loss saturation. Across multiple datasets, our empirical evaluation shows that this condition is always met in practice. Our discretization-free algorithm consistently matches or outperforms existing multicalibration approaches—even when evaluated using a discretization-based multicalibration metric that shares its discretization granularity with the baselines. Code to replicate the results in this work is available at https://github.com/hjenryin/Discretization-free-MC.

Adaptive cooperation in multi-agent reinforcement learning (MARL) requires policies to express homogeneous, specialised, or mixed behaviours, yet achieving this adaptivity remains a critical challenge. While parameter sharing (PS) is standard for efficient learning, it notoriously suppresses the behavioural diversity required for specialisation. This failure is largely due to cross-agent gradient interference, a problem we find is surprisingly exacerbated by the common practice of *coupling agent IDs with observations*. Existing remedies typically add complexity through altered objectives, manual preset diversity levels, or sequential updates -- raising a fundamental question: *can shared policies adapt without these intricacies?* We propose a solution built on a key insight: an agent-conditioned hypernetwork can generate agent-specific parameters and *decouple* observation- and agent-conditioned gradients, directly countering the interference from coupling agent IDs with observations. Our resulting method, **HyperMARL**, avoids the complexities of prior work and empirically reduces policy gradient variance. Across diverse MARL benchmarks (22 scenarios, up to 30 agents), HyperMARL achieves performance competitive with six key baselines while preserving behavioural diversity comparable to non-parameter sharing methods, establishing it as a versatile and principled approach for adaptive MARL. The code is publicly available at https://github.com/KaleabTessera/HyperMARL.

We introduce FreshStack, a holistic framework for automatically building information retrieval (IR) evaluation benchmarks by incorporating challenging questions and answers. FreshStack conducts the following steps: (1) automatic corpus collection from code and technical documentation, (2) nugget generation from community-asked questions and answers, and (3) nugget-level support, retrieving documents using a fusion of retrieval techniques and hybrid architectures. We use FreshStack to build five datasets on fast-growing, recent, and niche domains to ensure the tasks are sufficiently challenging. On FreshStack, existing retrieval models, when applied out-of-the-box, significantly underperform oracle approaches on all five domains, denoting plenty of headroom to improve IR quality. In addition, we identify cases where rerankers do not improve first-stage retrieval accuracy (two out of five domains) and oracle context helps an LLM generator generate a high-quality RAG answer. We hope FreshStack will facilitate future work toward constructing realistic, scalable, and uncontaminated IR and RAG evaluation benchmarks.

Conventional wisdom dictates that small batch sizes make language model pretraining and fine-tuning unstable, motivating gradient accumulation, which trades off the number of optimizer steps for a proportional increase in batch size. While it is common to decrease the learning rate for smaller batch sizes, other hyperparameters are often held fixed. In this work, we revisit small batch sizes all the way down to batch size one, and we propose a rule for scaling Adam hyperparameters to small batch sizes. In particular, rather than holding the decay rate of the second moment fixed across batch sizes, we propose to hold its half-life fixed in terms of tokens. We find that small batch sizes (1) train stably, (2) are consistently more robust to hyperparameter choices, (3) achieve equal or better per-FLOP performance than larger batch sizes, and (4) notably enable stable language model training with vanilla SGD, even without momentum, despite storing no optimizer state. Building on these results, we provide practical recommendations for selecting a batch size and setting optimizer hyperparameters. We further recommend against gradient accumulation unless training on multiple devices with multiple model replicas. Finally, we show that a small batch size combined with an optimizer with a small state size can provide the performance benefits of full fine-tuning while maintaining a similar memory footprint to LoRA.

The increasing utilization of graph databases across various fields stems from their capacity to represent intricate interconnections. Nonetheless, exploiting the full capabilities of graph databases continues to be a significant hurdle, largely because of the inherent difficulty in translating natural language into Cypher. Recognizing the critical role of schema selection in database query generation and drawing inspiration from recent progress in reasoning-augmented approaches trained through reinforcement learning to enhance inference capabilities and generalization, we introduce Cypher-RI, a specialized framework for the Text-to-Cypher task. Distinct from conventional approaches, our methodology seamlessly integrates schema selection within the Cypher generation pipeline, conceptualizing it as a critical element in the reasoning process. The schema selection mechanism is guided by textual context, with its outcomes recursively shaping subsequent inference processes. Impressively, our 7B-parameter model, trained through this RL paradigm, demonstrates superior performance compared to baselines, exhibiting a 9.41\% accuracy improvement over GPT-4o on CypherBench. These results underscore the effectiveness of our proposed reinforcement learning framework, which integrates schema selection to enhance both the accuracy and reasoning capabilities in Text-to-Cypher tasks.