NeurIPS 2025 Papers

Large Vision-Language Action (VLA) models have shown significant potential for embodied AI. However, their predominant training via supervised fine-tuning (SFT) limits generalization due to susceptibility to compounding errors under distribution shifts. Reinforcement learning (RL) offers a path to overcome these limitations by optimizing for task objectives via trial-and-error, yet a systematic understanding of its specific generalization benefits for VLAs compared to SFT is lacking. To address this, our study introduces a comprehensive benchmark for evaluating VLA generalization and systematically investigates the impact of RL fine-tuning across diverse visual, semantic, and execution dimensions. Our extensive experiments reveal that RL fine-tuning, particularly with PPO, significantly enhances generalization in semantic understanding and execution robustness over SFT, while maintaining comparable visual robustness. We identify PPO as a more effective RL algorithm for VLAs than LLM-derived methods like DPO and GRPO. We also develop a simple recipe for efficient PPO training on VLAs, and demonstrate its practical utility for improving VLA generalization. The project page is at https://rlvla.github.io

Large Language Models (LLMs) have demonstrated remarkable instruction-following capabilities across various applications. However, their performance in multilingual settings lacks systematic investigation, with existing evaluations lacking fine-grained constraint analysis across diverse linguistic contexts. We introduce **XIFBench**, a comprehensive constraint-based benchmark for evaluating multilingual instruction-following abilities of LLMs, comprising 558 instructions with 0-5 additional constraints across five categories (*Content*, *Style*, *Situation*, *Format*, and *Numerical*) in six languages spanning different resource levels. To support reliable and consistent cross-lingual evaluation, we implement three methodological innovations: cultural accessibility annotation, constraint-level translation validation, and requirement-based evaluation using English requirements as semantic anchors across languages. Extensive experiments with various LLMs not only quantify performance disparities across resource levels but also provide detailed insights into how language resources, constraint categories, instruction complexity, and cultural specificity influence multilingual instruction-following. Our code and data are available at https://github.com/zhenyuli801/XIFBench.

Training effective AI agents for multi-turn interactions requires high-quality data that captures realistic human-agent dynamics, yet such data is scarce and expensive to collect manually. We introduce APIGen-MT, a two-phase framework that generates verifiable and diverse multi-turn agent data. In the first phase, our agentic pipeline produces detailed task blueprints with ground-truth actions, leveraging a committee of LLM reviewers and iterative feedback loops. These blueprints are then transformed into complete interaction trajectories through simulated human-agent interplay. We train a family of models---the xLAM-2-fc-r series with sizes ranging from 1B to 70B parameters. Our models outperform frontier models such as GPT-4o and Claude 3.5 on $\tau$-bench and BFCL benchmarks, with the smaller models surpassing their larger counterparts, particularly in multi-turn settings, while maintaining superior consistency across multiple trials. Comprehensive experiments demonstrate that our verified blueprint-to-details approach yields high-quality training data, enabling the development of more reliable, efficient, and capable agents. We open-source both the synthetic data collected and the trained xLAM-2-fc-r models to advance research in AI agents. Dataset: https://huggingface.co/datasets/Salesforce/APIGen-MT-5k & Models: https://huggingface.co/collections/Salesforce/xlam-2-67ef5be12949d8dcdae354c4

Recently, mobile manipulation has attracted increasing attention for enabling language-conditioned robotic control in household tasks. However, existing methods still face challenges in coordinating mobile base and manipulator, primarily due to two limitations. On the one hand, they fail to explicitly model the influence of the mobile base on manipulator control, which easily leads to error accumulation under high degrees of freedom. On the other hand, they treat the entire mobile manipulation process with the same visual observation modality (e.g., either all 2D or all 3D), overlooking the distinct multimodal perception requirements at different stages during mobile manipulation. To address this, we propose the Adaptive Coordination Diffusion Transformer (AC-DiT), which enhances mobile base and manipulator coordination for end-to-end mobile manipulation. First, since the motion of the mobile base directly influences the manipulator's actions, we introduce a mobility-to-body conditioning mechanism that guides the model to first extract base motion representations, which are then used as context prior for predicting whole-body actions. This enables whole-body control that accounts for the potential impact of the mobile base’s motion. Second, to meet the perception requirements at different stages of mobile manipulation, we design a perception-aware multimodal conditioning strategy that dynamically adjusts the fusion weights between various 2D visual images and 3D point clouds, yielding visual features tailored to the current perceptual needs. This allows the model to, for example, adaptively rely more on 2D inputs when semantic information is crucial for action prediction, while placing greater emphasis on 3D geometric information when precise spatial understanding is required. We empirically validate AC-DiT through extensive experiments on both simulated and real-world mobile manipulation tasks, demonstrating superior performance compared to existing methods.

Balancing helpfulness and safety (harmlessness) is a critical challenge in aligning large language models (LLMs). Current approaches often decouple these two objectives, training separate preference models for helpfulness and safety, while framing safety as a constraint within a constrained Markov Decision Process (CMDP) framework. This paper identifies a potential issue when using the widely adopted expected safety constraints for LLM safety alignment, termed "safety compensation'', where the constraints are satisfied on expectation, but individual prompts may trade off safety, resulting in some responses being overly restrictive while others remain unsafe. To address this issue, we propose **Rectified Policy Optimization (RePO)**, which replaces the expected safety constraint with critical safety constraints imposed on every prompt. At the core of RePO is a policy update mechanism driven by rectified policy gradients, which penalizes the strict safety violation of every prompt, thereby enhancing safety across nearly all prompts. Our experiments demonstrate that RePO outperforms strong baseline methods and significantly enhances LLM safety alignment.

In recent years, data poisoning attacks have been increasingly designed to appear harmless and even beneficial, often with the intention of verifying dataset ownership or safeguarding private data from unauthorized use. However, these developments have the potential to cause misunderstandings and conflicts, as data poisoning has traditionally been regarded as a security threat to machine learning systems. To address this issue, it is imperative for harmless poisoning generators to claim ownership of their generated datasets, enabling users to identify potential poisoning to prevent misuse. In this paper, we propose the deployment of watermarking schemes as a solution to this challenge. We introduce two provable and practical watermarking approaches for data poisoning: post-poisoning watermarking and poisoning-concurrent watermarking. Our analyses demonstrate that when the watermarking length is $\Theta(\sqrt{d}/\epsilon_w)$ for post-poisoning watermarking, and falls within the range of $\Theta(1/\epsilon_w^2)$ to $O(\sqrt{d}/\epsilon_p)$ for poisoning-concurrent watermarking, the watermarked poisoning dataset provably ensures both watermarking detectability and poisoning utility, certifying the practicality of watermarking under data poisoning attacks. We validate our theoretical findings through experiments on several attacks, models, and datasets.

Monitoring chemical processes is essential to prevent catastrophic failures, optimize costs and profits, and ensure the safety of employees and the environment. A key component of modern monitoring systems is the automated detection of anomalies in sensor data over time, called time series, enabling partial automation of plant operation and adding additional layers of supervision to crucial components. The development of anomaly detection methods in this domain is challenging, since real chemical process data are usually proprietary, and simulated data are generally not a sufficient replacement. In this paper, we present NoBOOM, the first collection of datasets for anomaly detection in real-world chemical process data, including labeled data from a running process at our industry partner BASF SE — one of the world’s leading chemical companies — and several chemical processes run in laboratory‑scale and pilot‑scale plants. While we are not able to share every detail about the industrial process, for the laboratory‑ and pilot‑scale plants, we provide comprehensive information on plant configuration, process operation, and, in particular, anomaly events, enabling a differentiated analysis of anomaly detection methods. To demonstrate the complexity of the benchmark, we analyze the data with regard to common issues of time-series anomaly detection (TSAD) benchmarks, including potential triviality and bias.

We revisit the problem of secure aggregation of high-dimensional vectors in a two-server system such as Prio. These systems are typically used to aggregate vectors such as gradients in private federated learning, where the aggregate itself is protected via noise addition to ensure differential privacy. Existing approaches require communication scaling with the dimensionality, and thus limit the dimensionality of vectors one can efficiently process in this setup. We propose PREAMBLE: {\bf Pr}ivate {\bf E}fficient {\bf A}ggregation {\bf M}echanism via {\bf BL}ock-sparse {\bf E}uclidean Vectors. PREAMBLE builds on an extension of distributed point functions that enables communication- and computation-efficient aggregation of {\em block-sparse vectors}, which are sparse vectors where the non-zero entries occur in a small number of clusters of consecutive coordinates. We show that these block-sparse DPFs can be combined with random sampling and privacy amplification by sampling results, to allow asymptotically optimal privacy-utility trade-offs for vector aggregation, at a fraction of the communication cost. When coupled with recent advances in numerical privacy accounting, our approach incurs a negligible overhead in noise variance, compared to the Gaussian mechanism used with Prio.

The use of large language models (LLMs) to simulate human behavior has gained significant attention, particularly through personas that approximate individual characteristics. Persona-based simulations hold promise for transforming disciplines that rely on population-level feedback, including social science, economic analysis, marketing research, and business operations. Traditional methods to collect realistic persona data face significant challenges. They are prohibitively expensive and logistically challenging due to privacy constraints, and often fail to capture multi-dimensional attributes, particularly subjective qualities. Consequently, synthetic persona generation with LLMs offers a scalable, cost-effective alternative. However, current approaches rely on ad hoc and heuristic generation techniques that do not guarantee methodological rigor or simulation precision, resulting in systematic biases in downstream tasks. Through extensive large-scale experiments including presidential election forecasts and general opinion surveys of the U.S. population, we reveal that these biases can lead to significant deviations from real-world outcomes. Based on the experimental results, this position paper argues that **a rigorous and systematic science of persona generation is needed to ensure the reliability of LLM-driven simulations of human behavior.** We call for not only methodological innovations and empirical foundations but also interdisciplinary organizational and institutional support for the development of this field. To support further research and development in this area, we have open-sourced approximately one million generated personas, available for public access and analysis.

The interpretation of genomic sequences is crucial for understanding biological processes. To handle the growing volume of DNA sequence data, Genomic Foundation Models (GFMs) have been developed by adapting architectures and training paradigms from Large Language Models (LLMs). Despite their remarkable performance in DNA sequence classification tasks, there remains a lack of systematic understanding regarding the training and task-adaptation processes of GFMs. Moreover, existing GFMs cannot achieve state-of-the-art performance on both short and long-context tasks and lacks multimodal abilities. By revisiting pre-training architectures and post-training techniques, we propose **Omni-DNA**, a family of models spanning 20M to 1.1B parameters that supports sequence understanding, long-context genomic reasoning, and natural-language annotation. **Omni-DNA** establishes new state-of-the-art results on 18 of 26 evaluations drawn from Nucleotide Transformer and Genomic Benchmarks. When jointly fine-tuning on biologically related tasks, **Omni-DNA** consistently outperform existing models and demonstrate multi-tasking abilities. To enable processing of arbitrary sequence lengths, we introduce **SEQPACK**—an adaptive compression operator that packs historical tokens into a learned synopsis using a position-aware learnable sampling mechanism, enabling transformer-based models to process ultra-long sequences with minimal memory and computational requirements. Our approach demonstrates superior performance on enhancer-target interaction tasks, capturing distant regulatory interactions at the 450kbp range more effectively than existing models. Finally, we present a new dataset termed **seq2func**, enabling Omni-DNA to generate accurate and functionally meaningful interpretations of DNA sequences, unlocking new possibilities for genomic analysis and discovery.

Traditional photography composition approaches are dominated by 2D cropping-based methods. However, these methods fall short when scenes contain poorly arranged subjects. Professional photographers often employ perspective adjustment as a form of 3D recomposition, modifying the projected 2D relationships between subjects while maintaining their actual spatial positions to achieve better compositional balance. Inspired by this artistic practice, we propose photography perspective composition (PPC), extending beyond traditional cropping-based methods. However, implementing the PPC faces significant challenges: the scarcity of perspective transformation datasets and undefined assessment criteria for perspective quality. To address these challenges, we present three key contributions: (1) An automated framework for building PPC datasets through expert photographs. (2) A video generation approach that demonstrates the transformation process from less favorable to aesthetically enhanced perspectives. (3) A perspective quality assessment (PQA) model constructed based on human performance. Our approach is concise and requires no additional prompt instructions or camera trajectories, helping and guiding ordinary users to enhance their composition skills.

Training vision language models (VLMs) aims to align visual representations from a vision encoder with the textual representations of a pretrained large language model (LLM). However, many VLMs exhibit reduced factual recall performance compared to their LLM backbones, raising the question of how effective multimodal fine-tuning is at extending existing mechanisms within the LLM to visual inputs. We argue that factual recall based on visual inputs requires VLMs to solve a two-hop problem: (1) forming entity representations from visual inputs, and (2) recalling associated factual knowledge based on these entity representations. By benchmarking 14 VLMs with various architectures (LLaVA, Native, Cross-Attention), sizes (7B-124B parameters), and training setups on factual recall tasks against their original LLM backbone models, we find that 11 of 14 models exhibit factual recall degradation. We select three models exhibiting high- and two models with low performance degradation, and use attribution patching, activation patching, and probing to show that degraded VLMs struggle to use the existing factual recall circuit of their LLM backbone, because they resolve the first hop too late in the computation. In contrast, high-performing VLMs resolve entity representations early enough to reuse the existing factual recall mechanism. Finally, we demonstrate two methods to recover performance: patching entity representations from the LLM backbone into the VLM, and prompting with chain-of-thought reasoning. Our results highlight that the speed of early entity resolution critically determines how effective VLMs are in using preexisting LLM mechanisms. More broadly, our work illustrates how mechanistic analysis can explain and unveil systematic failures in multimodal alignment.

Human brain achieves dynamic stability-plasticity balance through synaptic homeostasis, a self-regulatory mechanism that stabilizes critical memory traces while preserving optimal learning capacities. Inspired by this biological principle, we propose SPICED: a neuromorphic framework that integrates the synaptic homeostasis mechanism for unsupervised continual EEG decoding, particularly addressing practical scenarios where new individuals with inter-individual variability emerge continually. SPICED comprises a novel synaptic network that enables dynamic expansion during continual adaptation through three bio-inspired neural mechanisms: (1) critical memory reactivation, which mimics brain functional specificity, selectively activates task-relevant memories to facilitate adaptation; (2) synaptic consolidation, which strengthens these reactivated critical memory traces and enhances their replay prioritizations for further adaptations and (3) synaptic renormalization, which are periodically triggered to weaken global memory traces to preserve learning capacities. The interplay within synaptic homeostasis dynamically strengthens task-discriminative memory traces and weakens detrimental memories. By integrating these mechanisms with continual learning system, SPICED preferentially replays task-discriminative memory traces that exhibit strong associations with newly emerging individuals, thereby achieving robust adaptations. Meanwhile, SPICED effectively mitigates catastrophic forgetting by suppressing the replay prioritization of detrimental memories during long-term continual learning. Validated on three EEG datasets, SPICED show its effectiveness. More importantly, SPICED bridges biological neural mechanisms and artificial intelligence through synaptic homeostasis, providing insights into the broader applicability of bio-inspired principles.

Event-based structured light (SL) systems have attracted increasing attention for their potential in high-performance 3D measurement. Despite the inherent HDR capability of event cameras, reflective and absorptive surfaces still cause event cluttering and absence, which produce overexposed and underexposed regions that degrade the reconstruction quality. In this work, we present the first HDR 3D measurement framework specifically designed for event-based SL systems. First, we introduce a multi-contrast HDR coding strategy that facilitates imaging of areas with different reflectance. Second, to alleviate inter-frame interference caused by overexposed and underexposed areas, we propose a universal confidence-driven stereo matching strategy. Specifically, we estimate a confidence map as the fusion weight for features via an energy-guided confidence estimation. Further, we propose the confidence propagation volume, an innovative cost volume that offers both effective suppression of inter-frame interference and strong representation capability. Third, we contribute an event-based SL simulator and propose the first event-based HDR SL dataset. We also collect a real-world benchmarking dataset with ground truth. We validate the effectiveness of our method with the proposed confidence-driven strategy on both synthetic and real-world datasets. Experimental results demonstrate that our proposed HDR framework enables accurate 3D measurement even under extreme conditions.

Recent progress in 3D object generation has greatly improved both the quality and efficiency. However, most existing methods generate a single mesh with all parts fused together, which limits the ability to edit or manipulate individual parts. A key challenge is that different objects may have a varying number of parts. To address this, we propose a new end-to-end framework for part-level 3D object generation. Given a single input image, our method generates high-quality 3D objects with an arbitrary number of complete and semantically meaningful parts. We introduce a dual volume packing strategy that organizes all parts into two complementary volumes, allowing for the creation of complete and interleaved parts that assemble into the final object. Experiments show that our model achieves better quality, diversity, and generalization than previous image-based part-level generation methods. Our project page is at \url{https://research.nvidia.com/labs/dir/partpacker/}.

Autonomous AI agents that can follow instructions and perform complex multi-step tasks have tremendous potential to boost human productivity. However, to perform many of these tasks, the agents need access to personal information from their users, raising the question of whether they are capable of using it appropriately. In this work, we introduce a new benchmark **AgentDAM** that measures if AI web-navigation agents follow the privacy principle of *"data minimization"*. For the purposes of our benchmark, data minimization means that the agent uses a piece of potentially sensitive information only if it is "necessary" to complete a particular task. Our benchmark simulates realistic web interaction scenarios end-to-end and is adaptable to all existing web navigation agents. We use AgentDAM to evaluate how well AI agents built on top of GPT-4, Llama-3 and Claude can limit processing of potentially private information, and show that they are prone to inadvertent use of unnecessary sensitive information. We also propose a prompting-based defense that reduces information leakage, and demonstrate that our end-to-end benchmarking provides a more realistic measure than probing LLMs about privacy. Our results highlight that further research is needed to develop AI agents that can prioritize data minimization at inference time. We open source our benchmark at: https://github.com/facebookresearch/ai-agent-privacy

Discrete diffusion models offer a flexible, controllable approach to structured sequence generation, yet they still lag behind causal language models in expressive power. A key limitation lies in their reliance on the Markovian assumption, which restricts each step to condition only on the current state, leading to potential uncorrectable error accumulation. In this paper, We introduce CaDDi, a discrete diffusion model that conditions on the entire generative trajectory, thereby lifting the Markov constraint and allowing the model to revisit and improve past states. By unifying sequential (causal) and temporal (diffusion) reasoning in a single non‑Markovian transformer, CaDDi also treats standard causal language models as a special case and permits the direct reuse of pretrained LLM weights with no architectural changes. Empirically, CaDDi outperforms state‑of‑the‑art discrete diffusion baselines on natural‑language benchmarks, substantially narrowing the remaining gap to large autoregressive transformers.

Large Language Models (LLMs) are being explored for applications in scientific research, including their capabilities to synthesize literature, answer research questions, generate research ideas, and even conduct computational experiments. Ultimately, our goal is for these to help scientists derive novel scientific insights. In many areas of science, such insights often arise from processing and visualizing data to understand its patterns. However, evaluating whether an LLM-mediated scientific workflow produces outputs conveying the correct scientific insights is challenging to evaluate and has not been addressed in past work. We introduce AstroVisBench, the first benchmark for both scientific computing and visualization in the astronomy domain. AstroVisBench judges a language model’s ability to both (1) create astronomy-specific workflows to process and analyze data and (2) visualize the results of these workflows through complex plots. Our evaluation of visualizations uses a novel LLM-as-a-judge workflow, which is validated against annotation by five professional astronomers. Using AstroVisBench we present an evaluation of state-of-the-art language models, showing a significant gap in their ability to engage in astronomy research as useful assistants. This evaluation provides a strong end-to-end evaluation for AI scientists that offers a path forward for the development of visualization-based workflows, which are central to a broad range of domains from physics to biology.

We introduce the Region Encoder Network (REN), a fast and effective model for generating region-based image representations using point prompts. Recent methods combine class-agnostic segmenters (e.g., SAM) with patch-based image encoders (e.g., DINO) to produce compact and effective region representations, but they suffer from high computational cost due to the segmentation step. REN bypasses this bottleneck using a lightweight module that directly generates region tokens, enabling 60x faster token generation with 35x less memory, while also improving token quality. It uses a few cross-attention blocks that take point prompts as queries and features from a patch-based image encoder as keys and values to produce region tokens that correspond to the prompted objects. We train REN with three popular encoders—DINO, DINOv2, and OpenCLIP—and show that it can be extended to other encoders without dedicated training. We evaluate REN on semantic segmentation and retrieval tasks, where it consistently outperforms the original encoders in both performance and compactness, and matches or exceeds SAM-based region methods while being significantly faster. Notably, REN achieves state-of-the-art results on the challenging Ego4D VQ2D benchmark and outperforms proprietary LMMs on Visual Haystacks' single-needle challenge. The code and pretrained models are available at https://github.com/savya08/ren.