NeurIPS 2025 Papers

Large Language Models (LLMs) have demonstrated impressive capabilities as intelligent agents capable of solving complex problems. However, effective planning in scenarios involving dependencies between API or tool calls-particularly in multi-turn conversations-remains a significant challenge. To address this, we introduce T1, a tool-augmented, multi-domain, multi-turn conversational dataset specifically designed to capture and manage inter-tool dependencies across diverse domains. T1 enables rigorous evaluation of agents' ability to coordinate tool use across nine distinct domains (4 single domain and 5 multi-domain) with the help of an integrated caching mechanism for both short- and long-term memory, while supporting dynamic replanning-such as deciding whether to recompute or reuse cached results. Beyond facilitating research on tool use and planning, T1 also serves as a benchmark for evaluating the performance of open-weight and proprietary large language models. We present results powered by T1-Agent highlighting their ability to plan and reason in complex, tool-dependent scenarios.

How to integrate and verify spatial intelligence in foundation models remains an open challenge. Current practice often proxies Visual-Spatial Intelligence (VSI) with purely textual prompts and VQA-style scoring, which obscures geometry, invites linguistic shortcuts, and weakens attribution to genuinely spatial skills. We introduce Spatial Intelligence Grid (SIG): a structured, grid-based schema that explicitly encodes object layouts, inter-object relations, and physically grounded priors. As a complementary channel to text, SIG provides a faithful, compositional representation of scene structure for foundation-model reasoning. Building on SIG, we derive SIG-informed evaluation metrics that quantify a model’s intrinsic VSI, which separates spatial capability from language priors. In few-shot in-context learning with state-of-the-art multimodal LLMs (e.g. GPT- and Gemini-family models), SIG yields consistently larger, more stable, and more comprehensive gains across all VSI metrics compared to VQA-only representations, indicating its promise as a data-labeling and training schema for learning VSI. We also release SIGBench, a benchmark of 1.4K driving frames annotated with ground-truth SIG labels and human gaze traces, supporting both grid-based machine VSI tasks and attention-driven, human-like VSI tasks in autonomous-driving scenarios.

We introduce \emph{SLED}, an alternative approach to speech language modeling by encoding speech waveforms into sequences of continuous latent representations and modeling them autoregressively using an energy distance objective. The energy distance offers an analytical measure of the distributional gap by contrasting simulated and target samples, enabling efficient training to capture the underlying continuous autoregressive distribution. By bypassing reliance on residual vector quantization, SLED avoids discretization errors and eliminates the need for the complicated hierarchical architectures common in existing speech language models. It simplifies the overall modeling pipeline while preserving the richness of speech information and maintaining inference efficiency. Empirical results demonstrate that SLED achieves strong performance in both zero-shot and streaming speech synthesis, showing its potential for broader applications in general-purpose speech language models. Demos and code are available at \url{https://github.com/ictnlp/SLED-TTS}.

Recent advances in Audio-Language Models (ALMs) have significantly improved multimodal understanding capabilities. However, the introduction of the audio modality also brings new and unique vulnerability vectors. Previous studies have proposed jailbreak attacks that specifically target ALMs, revealing that defenses directly transferred from traditional audio adversarial attacks or text-based Large Language Model (LLM) jailbreaks are largely ineffective against these ALM-specific threats. To address this issue, we propose ALMGuard, the first defense framework tailored to ALMs. Based on the assumption that safety-aligned shortcuts naturally exist in ALMs, we design a method to identify universal Shortcut Activation Perturbations (SAPs) that serve as triggers that activate the safety shortcuts to safeguard ALMs at inference time. To better sift out effective triggers while preserving the model’s utility on benign tasks, we further propose Mel-Gradient Sparse Mask (M-GSM), which restricts perturbations to Mel-frequency bins that are sensitive to jailbreaks but insensitive to speech understanding. Both theoretical analyses and empirical results demonstrate the robustness of our method against both seen and unseen attacks. Overall, \MethodName reduces the average success rate of advanced ALM-specific jailbreak attacks to 4.6\% across four models, while maintaining comparable utility on benign benchmarks, establishing it as the new state of the art. Our code and data are available at https://github.com/WeifeiJin/ALMGuard.

Reinforcement learning (RL) has recently demonstrated strong potential in enhancing the reasoning capabilities of large language models (LLMs). Particularly, the "Zero" reinforcement learning introduced by Deepseek-R1-Zero, enables direct RL training of base LLMs without relying on an intermediate supervised fine-tuning stage. Despite these advancements, current works for LLM reasoning mainly focus on mathematical and coding domains, largely due to data abundance and the ease of answer verification. This limits the applicability and generalization of such models to broader domains, where questions often have diverse answer representations, and data is more scarce. In this paper, we propose General-Reasoner, a novel training framework designed to enhance LLM reasoning capabilities across diverse domains. Our key contributions include: (1) constructing a large-scale, high-quality dataset of questions with verifiable answers curated by web crawling, covering a wide range of disciplines; and (2) developing a generative model-based answer verifier, which replaces traditional rule-based verification with the capability of chain-of-thought and context-awareness. We train a series of models and evaluate them on a wide range of datasets covering wide domains like physics, chemistry, finance, electronics etc. Our comprehensive evaluation across these 12 benchmarks (e.g. MMLU-Pro, GPQA, SuperGPQA, TheoremQA, BBEH and MATH AMC) demonstrates that General-Reasoner outperforms existing baseline methods, achieving robust and generalizable reasoning performance while maintaining superior effectiveness in mathematical reasoning tasks.

Robust verbal confidence generated by large language models (LLMs) is crucial for the deployment of LLMs to help ensure transparency, trust, and safety in many applications, including those involving human-AI interactions. In this paper, we present the first comprehensive study on the robustness of verbal confidence under adversarial attacks. We introduce attack frameworks targeting verbal confidence scores through both perturbation and jailbreak-based methods, and demonstrate that these attacks can significantly impair verbal confidence estimates and lead to frequent answer changes. We examine a variety of prompting strategies, model sizes, and application domains, revealing that current verbal confidence is vulnerable and that commonly used defence techniques are largely ineffective or counterproductive. Our findings underscore the need to design robust mechanisms for confidence expression in LLMs, as even subtle semantic-preserving modifications can lead to misleading confidence in responses.

Existing molecular machine learning force fields (MLFFs) generally focus on the learning of atoms, molecules, and simple quantum chemical properties (such as energy and force), but ignore the importance of electron density (ED) $\rho(r)$ in accurately understanding molecular force fields (MFFs). ED describes the probability of finding electrons at specific locations around atoms or molecules, which uniquely determines all ground state properties (such as energy, molecular structure, etc.) of interactive multi-particle systems according to the Hohenberg-Kohn theorem. However, the calculation of ED relies on the time-consuming first-principles density functional theory (DFT), which leads to the lack of large-scale ED data and limits its application in MLFFs. In this paper, we introduce EDBench, a large-scale, high-quality dataset of ED designed to advance learning-based research at the electronic scale. Built upon the PCQM4Mv2, EDBench provides accurate ED data, covering 3.3 million molecules. To comprehensively evaluate the ability of models to understand and utilize electronic information, we design a suite of ED-centric benchmark tasks spanning prediction, retrieval, and generation. Our evaluation of several state-of-the-art methods demonstrates that learning from EDBench is not only feasible but also achieves high accuracy. Moreover, we show that learning-based methods can efficiently calculate ED with comparable precision while significantly reducing the computational cost relative to traditional DFT calculations. All data and benchmarks from EDBench will be freely available, laying a robust foundation for ED-driven drug discovery and materials science.

Data fusion techniques integrate information from heterogeneous data sources to improve learning, generalization, and decision-making across data sciences. In causal inference, these methods leverage rich observational data to improve causal effect estimation, while maintaining the trustworthiness of randomized controlled trials. Existing approaches often relax the strong "no unobserved confounding" assumption by instead assuming exchangeability of counterfactual outcomes across data sources. However, when both assumptions simultaneously fail—a common scenario in practice—current methods cannot identify or estimate causal effects. We address this limitation by proposing a novel partial identification framework that enables researchers to answer key questions such as: *Is the causal effect positive/negative?* and *How severe must assumption violations be to overturn this conclusion?* Our approach introduces interpretable sensitivity parameters that quantify assumption violations and derives corresponding causal effect bounds. We develop doubly robust estimators for these bounds and operationalize breakdown frontier analysis to understand how causal conclusions change as assumption violations increase. We apply our framework to the Project STAR study, which investigates the effect of classroom size on students’ third-grade standardized test performance. Our analysis reveals that the Project STAR results are robust to simultaneous violations of key assumptions, both on average and across various subgroups of interest. This strengthens confidence in the study's conclusions despite potential unmeasured biases in the data.

Sparse autoencoders (SAEs) are designed to extract interpretable features from language models by enforcing a sparsity constraint. Ideally, training an SAE would yield latents that are both sparse and semantically meaningful. However, many SAE latents activate frequently (i.e., are *dense*), raising concerns that they may be undesirable artifacts of the training procedure. In this work, we systematically investigate the geometry, function, and origin of dense latents and show that they are not only persistent but often reflect meaningful model representations. We first demonstrate that dense latents tend to form antipodal pairs that reconstruct specific directions in the residual stream, and that ablating their subspace suppresses the emergence of new dense features in retrained SAEs---suggesting that high density features are an intrinsic property of the residual space. We then introduce a taxonomy of dense latents, identifying classes tied to position tracking, context binding, entropy regulation, letter-specific output signals, part-of-speech, and principal component reconstruction. Finally, we analyze how these features evolve across layers, revealing a shift from structural features in early layers, to semantic features in mid layers, and final to output-oriented signals in the last layers of the model. Our findings indicate that dense latents serve functional roles in language model computation and should not be dismissed as training noise.

A common problem in LLM evaluation is how to choose a subset of metrics from a full suite of possible metrics. Subset selection is usually done for efficiency or interpretability reasons, and the goal is often to select a "representative" subset of metrics. However, "representative" is rarely clearly defined. In this work, we use ideas from social choice theory to formalize two notions of representation for the selection of a subset of evaluation metrics. We first introduce *positional representation*, which guarantees every alternative is sufficiently represented at every position cutoff. We then introduce *positional proportionality*, which guarantees no alternative is proportionally over- or under-represented by more than a small error at any position. We prove upper and lower bounds on the smallest number of metrics needed to guarantee either of these properties in the worst case. We also study a generalized form of each property that allows for additional input on groups of metrics that must be represented. Finally, we tie theory to practice through real-world case studies on both LLM evaluation and hospital quality evaluation.

Routing Problems are central to many real-world applications, yet remain challenging due to their (NP-)hard nature. Amongst existing approaches, heuristics often offer the best trade-off between quality and scalability, making them suitable for industrial use. While Reinforcement Learning (RL) offers a flexible framework for designing heuristics, its adoption over handcrafted heuristics remains incomplete. Existing learned methods still lack the ability to adapt to specific instances and fully leverage the available computational budget. Current best methods either rely on a collection of pre-trained policies, or on RL fine-tuning; hence failing to fully utilize newly available information within the constraints of the budget. In response, we present MEMENTO, an approach that leverages memory to improve the search of neural solvers at inference. MEMENTO updates the action distribution dynamically based on the outcome of previous decisions. We validate its effectiveness on Traveling Salesman and Capacitated Vehicle Routing problems, demonstrating its superiority over tree-search and policy-gradient fine-tuning; and showing that it can be zero-shot combined with diversity-based solvers. We successfully train all RL auto-regressive solvers on large instances, and verify MEMENTO's scalability and data-efficiency: pushing the state-of-the-art on 11 out of 12 evaluated tasks.

Neural video compression (NVC) has made significant progress in recent years, while neural B-frame video compression (NBVC) remains underexplored compared to P-frame compression. NBVC can adopt bi-directional reference frames for better compression performance. However, NBVC's hierarchical coding may complicate continuous temporal prediction, especially at some hierarchical levels with a large frame span, which could cause the contribution of the two reference frames to be unbalanced. To optimize reference information utilization, we propose a novel NBVC method, termed Bi-directional Reference Harmonization Video Compression (BRHVC), with the proposed Bi-directional Motion Converge (BMC) and Bi-directional Contextual Fusion (BCF). BMC converges multiple optical flows in motion compression, leading to more accurate motion compensation on a larger scale. Then BCF explicitly models the weights of reference contexts under the guidance of motion compensation accuracy. With more efficient motions and contexts, BRHVC can effectively harmonize bi-directional references. Experimental results indicate that our BRHVC outperforms previous state-of-the-art NVC methods, even surpassing the traditional coding, VTM-RA (under random access configuration), on the HEVC datasets. The source code will be released. The source code is released at https://github.com/kwai/NVC.

Sign Language Translation (SLT) aims to map sign language videos to spoken language text. A common approach relies on gloss annotations as an intermediate representation, decomposing SLT into two sub-tasks: video-to-gloss recognition and gloss-to-text translation. While effective, this paradigm depends on expert-annotated gloss labels, which are costly and rarely available in existing datasets, limiting its scalability. To address this challenge, we propose a gloss-free pseudo gloss generation framework that eliminates the need for human-annotated glosses while preserving the structured intermediate representation. Specifically, we prompt a Large Language Model (LLM) with a few example text-gloss pairs using in-context learning to produce draft sign glosses from spoken language text. To enhance the correspondence between LLM-generated pseudo glosses and the sign sequences in video, we correct the ordering in the pseudo glosses for better alignment via a weakly supervised learning process. This reordering facilitates the incorporation of auxiliary alignment objectives, and allows for the use of efficient supervision via a Connectionist Temporal Classification (CTC) loss. We train our SLT model—consisting of a vision encoder and a translator—through a three-stage pipeline, which progressively narrows the modality gap between sign language and spoken language. Despite its simplicity, our approach outperforms previous state-of-the-art gloss-free frameworks on two SLT benchmarks and achieves competitive results compared to gloss-based methods.

Video Temporal Grounding (VTG) aims to localize temporal segments in long, untrimmed videos that align with a given natural language query. This task typically comprises two subtasks: \textit{Moment Retrieval (MR)} and \textit{Highlight Detection (HD)}. While recent advances have been progressed by powerful pretrained vision-language models such as CLIP and InternVideo2, existing approaches commonly treat all text tokens uniformly during cross-modal attention, disregarding their distinct semantic roles. To validate the limitations of this approach, we conduct controlled experiments demonstrating that VTG models overly rely on [EOS]-driven global semantics while failing to effectively utilize word-level signals, which limits their ability to achieve fine-grained temporal alignment. Motivated by this limitation, we propose DualGround, a dual-branch architecture that explicitly separates global and local semantics by routing the [EOS] token through a sentence-level path and clustering word tokens into phrase-level units for localized grounding. Our method introduces (1) token-role-aware cross modal interaction strategies that align video features with sentence-level and phrase-level semantics in a structurally disentangled manner, and (2) a joint modeling framework that not only improves global sentence-level alignment but also enhances fine-grained temporal grounding by leveraging structured phrase-aware context. This design allows the model to capture both coarse and localized semantics, enabling more expressive and context-aware video grounding. DualGround achieves state-of-the-art performance on both Moment Retrieval and Highlight Detection tasks across QVHighlights and Charades-STA benchmarks, demonstrating the effectiveness of disentangled semantic modeling in video-language alignment.

In this work, we consider the problem of Bayesian Optimization (BO) under reward model uncertainty—that is, when the underlying distribution type of the reward is unknown and potentially intractable to specify. This challenge is particularly evident in many modern applications, where the reward distribution is highly ill-behaved, often non-stationary, multi-modal, or heavy-tailed. In such settings, classical Gaussian Process (GP)-based BO methods often fail due to their strong modeling assumptions. To address this challenge, we propose a novel surrogate model, the infinity-Gaussian Process ($\infty$-GP), which represents a sequential spatial Dirichlet Process mixture with a GP baseline. The $\infty$-GP quantifies both value uncertainty and model uncertainty, enabling more flexible modeling of complex reward structures. Combined with Thompson Sampling, the $\infty$-GP facilitates principled exploration and exploitation in the distributional space of reward models. Theoretically, we prove that the $\infty$-GP surrogate model can approximate a broad class of reward distributions by effectively exploring the distribution space, achieving near-minimax-optimal posterior contraction rates. Empirically, our method outperforms state-of-the-art approaches in various challenging scenarios, including highly non-stationary and heavy-tailed reward settings where classical GP-based BO often fails.

We present Perceive Anything Model (PAM), a conceptually straightforward and efficient framework for comprehensive region-level visual understanding in images and videos. Our approach extends the powerful segmentation model SAM 2 by integrating Large Language Models (LLMs), enabling simultaneous object segmentation with the generation of diverse, region-specific semantic outputs, including categories, label definition, functional explanations, and detailed captions. A key component, Semantic Perceiver, is introduced to efficiently transform SAM 2's rich visual features, which inherently carry general vision, localization, and semantic priors into multi-modal tokens for LLM comprehension. To support robust multi-granularity understanding, we also develop a dedicated data refinement and augmentation pipeline, yielding a high-quality dataset of 1.5M image and 0.6M video region-semantic annotations, including novel region-level streaming video caption data. PAM is designed for lightweightness and efficiency, while also demonstrates strong performance across a diverse range of region understanding tasks. It runs 1.2$-$2.4$\times$ faster and consumes less GPU memory than prior approaches, offering a practical solution for real-world applications. We believe that our effective approach will serve as a strong baseline for future research in region-level visual understanding.

Neural rendering techniques, including NeRF and Gaussian Splatting (GS), rely on photometric consistency to produce high-quality reconstructions. However, in real-world driving scenarios, it is challenging to guarantee perfect photometric consistency in acquired images. Appearance codes have been widely used to address this issue, but their modeling capability is limited, as a single code is applied to the entire image. Recently, the bilateral grid was introduced to perform pixel-wise color mapping, but it is difficult to optimize and constrain effectively. In this paper, we propose a novel multi-scale bilateral grid that unifies appearance codes and bilateral grids. We demonstrate that this approach significantly improves geometric accuracy in dynamic, decoupled autonomous driving scene reconstruction, outperforming both appearance codes and bilateral grids. This is crucial for autonomous driving, where accurate geometry is important for obstacle avoidance and control. Our method shows strong results across four datasets: Waymo, NuScenes, Argoverse, and PandaSet. We further demonstrate that the improvement in geometry is driven by the multi-scale bilateral grid, which effectively reduces floaters caused by photometric inconsistency.

Large language models (LLMs) have achieved impressive success, but their high memory requirements present challenges for long-context token generation. In this paper we study the streaming complexity of attention approximation, a key computational primitive underlying token generation. Our main contribution is BalanceKV, a streaming algorithm for $\epsilon$-approximating attention computations based on geometric process for selecting a balanced collection of Key and Value tokens as per Banaszczyk's vector balancing theory. We complement our algorithm with space lower bounds for streaming attention computation. Besides strong theoretical guarantees, BalanceKV exhibits empirically validated performance improvements over existing methods, both for attention approximation and end-to-end performance on various long context benchmarks.

In-context learning (ICL) is one of the key capabilities contributing to the great success of LLMs. At test time, ICL is known to operate in the two modes: task recognition and task learning. In this paper, we investigate the emergence and dynamics of the two modes of ICL during pretraining. To provide an analytical understanding of the learning dynamics of the ICL abilities, we investigate the in-context random linear regression problem with a simple linear-attention-based transformer, and define and disentangle the strengths of the task recognition and task learning abilities stored in the transformer model’s parameters. We show that, during the pretraining phase, the model first learns the task learning and the task recognition abilities together in the beginning, but it (a) gradually forgets the task recognition ability to recall the priorly learned tasks and (b) relies more on the given context in the later phase, which we call (a) \textit{prior forgetting} and (b) \textit{in-context overfitting}, respectively.