NeurIPS 2026 论文列表

Language model agents are becoming proficient executors at isolated, short-horizon tasks such as software engineering and customer service. Yet real-world challenges require a combination of sophisticated skills that remain largely untested in agents: (1) navigating long horizons amid uncertainty; (2) acquiring information in noisy environments; (3) adapting to a changing world; (4) orchestrating multiple moving parts toward a coherent goal. We introduce CEO-Bench, which evaluates these capabilities together by simulating a representative real-world task: operating a startup for 500 days. Given diverse and realistic company management tools, business databases, and social media, an agent needs to design pricing strategies, allocate operating budgets, analyze business data, respond to unexpected competitor moves, and more. Our evaluation shows that most state-of-the-art models struggle to succeed in this environment, and only one model (GPT-5.5) finishes the simulation above its $1M starting balance. CEO-Bench reimagines the role of AI in the future, shifting from solving isolated tasks to driving sustained, adaptive progress over time. We open source full code and trajectory.

Convolutional neural networks (CNNs) are widely used for time-series classification, but their deployment in critical domains requires understanding the temporal and spectral patterns that drive their predictions. Concept extraction (CE) methods identify such patterns by analyzing representations within the models' latent space. However, existing time-series CE methods have three limitations: they operate only in the time domain and overlook frequency features, predefine the number of concepts, and produce localizations misaligned with the regions the model uses. We address these limitations by proposing CENDRe, a concept extraction method for CNNs. It first discovers concepts by clustering per-timestep latent representations in two stages, where silhouette-guided aggregation selects the number of concepts automatically. Then, it localizes each concept through gradients of a presence score that contrasts the latent representations with their prototypes, producing masks that concentrate on the regions driving the concept. These gradients, propagated through a differentiable invertible mapping of the input such as a Fourier transform, yield localizations for the same concepts in the frequency domain. Finally, each concept receives a relevance score that quantifies its contribution to each class. On synthetic benchmarks, CENDRe achieves representation correctness comparable to state-of-the-art CE methods and substantially higher importance correctness. On real bearing-fault data, CENDRe extracts the frequency bands driving the model's predictions, located in regions commonly inspected for fault diagnosis, producing evidence to assess the model that time-domain CE methods cannot.

Can we trust evaluation scores to capture an LLM's true real-world performance? *Certifiable* evaluation answers this question by providing guarantee for LLM evaluation. In particular, existing methods sequentially curate evaluation samples and keep updating confidence intervals (CIs) that cover the true performance with high probability (e.g., 95%) until some conditions are satisfied, e.g., the CI width reaches a target precision. However, existing methods are not generally anytime-valid: the claimed coverage (e.g., 95%) may fail when CIs are repeatedly updated and used to decide when to stop, leaving a *gap* between theoretical rigor and practice. This paper bridges this gap by proposing **Celeus**, a **C**ertifiable framework for **E**fficient **L**LM evaluation, which leverages ***E***-processes to build anytime-valid CIs. Concretely, we propose signals that combine two ingredients: (i) **U**ncertainty-guided sampling to select informative samples for evaluation, and (ii) **S**urrogate-assisted approximations for unevaluated samples. We prove that such signals remain unbiased for the evaluation score conditional on the past, enabling statistically grounded and anytime-valid $e$-process CIs. More importantly, the two ingredients reduce estimation variance and help reach the target precision with fewer evaluated samples. We also prove that CIs obtained by **Celeus** can shrink at a near-parametric rate up to logarithmic factors and analyze the oracle variance-optimal sampling rule that motivates the empirical uncertainty-guided one. Experiments show that **Celeus** reaches the target precision using 54-62% fewer evaluated samples than baselines, while preserving anytime-valid coverage.

Despite growing interest in on-device LLM deployment, large-scale Mixture-of-Experts (MoE) models remain impractical on resource-constrained devices, as sparse expert activation still requires all expert weights to be memory-resident. To address this, we identify three key observations for on-device MoE inference: (a) cache hit rates are fundamentally bounded even under Belady’s optimal policy; (b) these bounds are insufficient for low-latency on-device inference; and (c) prior cache-aware routing based on binary cache presence can induce routing instability. Based on these insights, we propose CECAR (Cache & Expert Co-Aware Routing), a MoE inference system that jointly considers expert routing and cache management. Unlike prior cache-aware routing methods, CECAR uses an ML-based cache policy to predict the expected reuse distance of each expert and prioritizes routing and execution toward experts with smaller predicted reuse distances. On the Qwen3-30B-A3B model, CECAR achieves 12.5 tokens/s on a consumer-grade GPU with only 16 GB of VRAM, where non-resident experts are fetched from SSD, delivering a 4.5× speedup over conventional MoE inference and exceeding the decode speed of baselines that keep the entire model fully resident in DRAM.

Reflectance confocal microscopy (RCM) provides noninvasive, cellular-resolution “optical biopsies” of human skin {\em in vivo} by acquiring en-face images at successive depths, forming a sparse $z$-stack. Due to optical limitations, these stacks are anisotropic 3D volumes with lateral resolution ($0.5\mu m$) $\sim$6 times higher compared to axial resolution, which is defined by the optical sectioning ($3\mu m$), limiting the interpretation of tissue. Our goal is to provide continuous-depth visualization by interpolating intermediate sections and making the 3D volume isotropic. Such a representation permits arbitrary-direction sectioning, including histopathology-like cross-sectional examination, without requiring per-patient optimization. To that end, we introduce the first RCM-specific novel-view synthesis (NVS) approach CD-RCM: a feedforward model that predicts realistic, unseen depths from sparsely sampled RCM stacks. Classical neural rendering methods focus on reconstruction from surface-level multi-view observations. In contrast to surface-level camera views, RCM can acquire optically sectioned en-face images of tissue beyond the surface up to $200 \mu m$. However, during visualization of the RCM stacks, observations of the shallower sections (towards the surface) obscure the deeper ones. This unique axial imaging geometry and layer-dependent anatomical organization motivated our development of a tailored architectural and training framework that explicitly accounts for RCM’s depth-resolved, occlusive imaging physics. Experiments demonstrate that CD-RCM achieves high-fidelity novel-view synthesis with sub-second inference time.

Many survival problems are inherently time-dependent: a biomarker may be predictive soon after diagnosis but irrelevant years later, or a treatment effect may diminish, reverse, or interact with follow-up duration. However, tree-based survival models commonly used in practice typically assign each individual a fixed, time-independent risk score. Neural non-proportional-hazards models can capture dynamic effects, but are often more challenging to tune, scale, and interpret than boosted trees. We present CCTimeBoost, a gradient-boosted tree framework for learning continuous-time, time-varying relative risk. The method augments covariates with time features, trains boosted trees on sampled risk sets using a group-softmax objective, and estimates a full-risk-set baseline hazard to generate survival curves. We show that this objective has a likelihood interpretation: for finite sampled risk sets, it matches a nested case-control likelihood, and in the full-risk-set limit, it converges to the Cox partial likelihood. Across eight right-censored survival benchmarks, CCTimeBoost delivers state-of-the-art discrimination performance relative to tree and neural baselines and is competitive in calibration. In controlled simulations, it accurately recovers diminishing effects, crossing hazards, and rule-based temporal interactions, also behaving as a proportional model when proportional hazards are present. These results indicate that non-proportional survival modeling can be incorporated into standard boosted-tree methods.

Set-level visual difference discovery is increasingly important for dataset auditing and for understanding model behavior under distribution shift, yet manually inspecting thousands of images is impractical. This motivates the emerging task of *set difference captioning*, i.e., describing concepts that are more often true for one image set than another. Different from existing approaches which heavily rely on large language models (LLMs), incurring substantial computational cost in terms of time and tokens, we introduce **CCDiff**, a lightweight, statistically grounded, and training-free framework for set difference captioning that significantly reduces LLM dependence during core difference discovery, achieving over **$2\times$** speedup and reducing token usage to zero. CCDiff operates in three stages: constructing a shared pool of candidate concepts from domain vocabularies or lightweight captioning models, filtering candidates to obtain set-specific concept pools, and performing inverse canonical correlation analysis (CCA) across sets to identify low-correlation directions that correspond to set differences. Candidate concepts are then ranked using a CCDiff score that balances inverse correlation with within-set representativeness. We evaluate CCDiff on VisDiffBench and further validate it on additional domains including MetaShift and MIMIC-CXR. Across benchmarks and real-world settings, CCDiff matches or exceeds prior LLM-based approaches while being substantially more efficient and scalable.

Training a single 3D Gaussian Splatting (3DGS) scene routinely requires tens of gigabytes of GPU memory, placing standard 3DGS training beyond many low-memory GPU budgets commonly found on laptops and edge devices. Existing low-memory approaches mainly follow two directions: block-wise training and frustum-culling-based host offloading. Block-wise methods train spatial blocks independently and merge them afterward. However, because 3DGS renders each image through depth-ordered alpha compositing over all visible Gaussians, independent block training reduces optimization to block-local objectives, thereby losing full-image loss supervision. Frustum-culling-based host offloading retains full-image loss supervision by jointly rendering all visible Gaussians while uploading only the view-visible subset to the GPU. However, its GPU memory consumption still scales with the size of the visible set, which can exceed the memory budget of low-end GPUs for dense scenes or wide-coverage views. We propose CC-GS, a low-memory block-wise 3DGS training framework based on CPU-GPU block-wise context compositing, featuring three key designs. $\textbf{i) Block-wise context compositing}$: CC-GS partitions Gaussians into capacity-bounded blocks and represents non-optimized visible blocks using cached per-pixel color and residual transmittance, with depth used only for ordering, enabling full-image loss supervision while processing one capacity-bounded block at a time. $\textbf{ii) Two-pass block-wise training}$: each training view is processed through a lightweight context pass for caching compositing context and a differentiable block-optimization pass for updating individual blocks. $\textbf{iii) Asynchronous CPU-GPU execution}$: CC-GS overlaps block gathering, host-device transfers, rendering, backpropagation, gradient offloading, and CPU-side updates through an asynchronous CPU-GPU pipeline. Experiments on standard 3DGS benchmarks show that CC-GS keeps the measured peak GPU memory below 2GB for all evaluated scenes while matching the rendering quality of standard 3DGS. Compared with standard 3DGS, CC-GS, incurs at most a $1.54\times$ dataset-level training slowdown.

Vision-Language Models (VLMs) have achieved strong performance on general multimodal reasoning, yet remain challenged in integrating nonlocal visual information to support semantically underdetermined visual reasoning. We describe this challenge as Fragmented Visual Reasoning. To this end, we propose Credit Assignment for Visual Evidence (CAVE), a structured process-reward method based on GRPO for interleaved visual reasoning. Specifically, CAVE evaluates the contribution of intermediate steps at the action level via three complementary reasoning process signals: belief update, evidence acquisition, and adaptive focus control, thereby guiding the model to optimize each reasoning action and learn more reliable visual reasoning strategies. Meanwhile, we construct TRACER-Bench, which covers four nonlocal and semantically confusable reasoning dimensions and provides key intermediate evidence to supervise reasoning paths. Experiments demonstrate that CAVE substantially improves performance on tasks requiring fragmented visual evidence integration, covering both public benchmarks and our newly introduced TRACER-Bench, while retaining competitive performance on general multimodal evaluations. Further analyses reveal that CAVE effectively improves the visual reasoning capacity and exhibits stronger robustness under longer-range and deeper cross-region dependencies. Our code and the proposed benchmark are available at: https://anonymous.4open.science/r/Anonymous_code-2D8B/.

While Speculative Decoding (SD) has become an essential lossless acceleration technique for Large Language Models, its direct application to Multimodal Large Language Models (MLLMs) is hindered by the intricate visual dependencies present in cross-modal generation. Existing SD methods enforce a uniform global alignment between the draft and target models, which proves ineffective in multimodal settings due to a structured distribution bias: deviations of the draft model are systematically concentrated on a sparse subset of tokens that demand fine-grained visual comprehension. To overcome this limitation, we reformulate the training of multimodal draft models as an acceptance-oriented critical-token alignment problem and introduce CATS, a novel two-stage training framework. CATS employs two complementary selection mechanisms to pinpoint critical tokens:hard tokens that exhibit high rejection probabilities, and vision-critical tokens whose prediction fundamentally depends on visual semantics. Following an initial global alignment phase, CATS performs sparse, targeted refinement exclusively on the union of these identified token sets. Extensive experiments across three representative benchmarks using multiple LLaVA and Qwen2.5-VL target models demonstrate that our approach consistently improves the token acceptance rate over strong baselines such as EAGLE-2 and supervised fine-tuning. Ultimately, CATS substantially accelerates MLLM inference, achieving speedups of up to 3.05×.

Many decision-facing stochastic systems are observed through aggregate distributions rather than scalar trajectories: queue occupancies, mobility shares, public-health mixtures, generation-source shares, ecological compositions, and air-quality severity profiles all live on the probability simplex and evolve over time. We study causal (time-respecting online) forecasting for these distribution-valued time series and argue that the transition operator itself should be structured around the simplex. We introduce CAST (Causal Anchored Simplex Transport), a successor-local operator that (i) retrieves empirical successors from causal context, (ii) stabilizes them with a persistence anchor, and (iii) applies a bounded local stochastic transport on ordered supports; every stage preserves the simplex by construction. We identify a structural failure mode, latent transition-kernel aliasing, where similar observed distributions evolve differently under different contextual regimes, and prove that any forecaster depending only on an aliased summary incurs an irreducible weighted Jensen--Shannon excess-risk lower bound, while the CAST hypothesis class contains the regime-aware Bayes successor; for ordered supports an additional Pinsker separation holds whenever the transported successor lies outside the no-transport anchor hull. On a suite of eleven public and simulated benchmarks spanning ecology, energy, diet, mortality, employment, air quality, severe weather, mobility, and $G/G/1$, $G_t/G/1$ queue occupancy, CAST achieves the best average rank on both one-step KL (1.27) and autoregressive rollout JSD (1.91), winning 8/11 sections on each metric against a broad statistical, compositional, recurrent, convolutional, Transformer, and modern time-series baseline set, and top-2 on all 11 sections for offline KL. Component ablations and a controlled synthetic aliasing experiment corroborate the theory.

Cascade attacks in LLM multi-agent systems (MAS) arise when adversarial influence propagates across agents, leading to escalated system-level failures through complex interaction dynamics. Detecting such cascades is challenging, as their signals are distributed and tightly coupled across interaction channels, often appearing locally benign while unfolding either within a single turn or gradually across multiple turns. Existing defenses, being largely local and text-centric, fail to capture the cross-channel, temporally coordinated structure of cascade propagation. We present CASPIAN, the first approach to provide a unified, cross-channel causal view of cascade behavior in LLM-MAS through online monitoring of dynamic influence propagation across agents. CASPIAN models multi-agent interactions as a unified, dynamic causal influence matrix across channels, estimated efficiently via a late-interaction conditional transfer entropy (LI-CTE) formulation, enabling the detection of cascade onset from emergent system-level structure rather than isolated anomalies. It then performs online causal attribution, identifying the origin, bridge, and amplifier agents driving the cascade and reconstructing its principal propagation pathways, capabilities not supported by existing methods. Across diverse multi-agent frameworks and benchmarks, CASPIAN consistently outperforms semantic guardrails, LLM-based judges, and graph-based anomaly detectors in both detection accuracy and early cascade identification, while operating with negligible additional latency. These results demonstrate that unified cross-channel causal modeling is essential for reliably detecting and understanding cascade failures in LLM multi-agent systems.

Quantifying variability in a target population relative to a reference population is central to many scientific and clinical problems (e.g., diseased vs. healthy). Yet, without paired data and in the presence of heterogeneous target variation, existing methods struggle to separate multiple modes of target-specific variation. We propose CASL-VAE, a deep contrastive latent variable model that learns structured latent generative factors from unpaired data. CASL-VAE factorizes variation into continuous common latent factors shared across populations and hierarchical salient latent factors that model target-specific heterogeneity as discrete subtypes and continuous within-subtype variation. Using variational inference, we show how approximate joint likelihood optimization over reference and target domains can be performed using unpaired data, providing a principled basis for paired-sample generation and cross-domain analysis. We validate CASL-VAE on semi-synthetic neuroimaging data, demonstrating improved subtype recovery and paired-sample generation compared to baseline clustering and generative models. We also validate its ability to reveal biologically plausible heterogeneity in Alzheimer's disease. Code will be released upon acceptance.

Many studies argue that the success of sharpness-aware minimization (SAM) is due to the implicit regularization of the gradient norm or the eigenvalues of the Hessian. However, they cannot fully explain why the flatness-based measures do not always correlate well with generalization. For instance, we can easily construct a counterexample by only optimizing with hard examples. In this paper, we propose to resolve such a contradiction from the perspective of anti-overfitting. First, we show that the sharpness at the mini-batch level approximately equals the sum of the gradient norm and the Kullback–Leibler (KL) divergence between the predictive distributions evaluated at the current point and its adversary. Second, we show that the \emph{gradient consistency}, which approximately quantifies the inner product between the mini-batch gradient and the per-example gradient, decreases consistently during training. This result explains why SAM is less likely to suffer from overfitting. At last, building on these insights, we further introduce CASAM, an algorithm that reweights each example according to gradient consistency to enhance the generalization performance. Extensive experiments on popular benchmarks such as ImageNet-1K and Clothing1M corroborate its efficacy.

Video detectors are usually evaluated on observational splits, where target events co-occur with environmental and capture factors. A high AUC under this setup can reflect either the event itself or its surrounding conditions, and standard reporting cannot tell them apart. This work addresses the entanglement with CARVE, a counterfactual video editing protocol. For each source clip, CARVE generates a matched quartet $(V^0, V^E, V^A, V^{AE})$ that holds camera pose, layout, and surrounding traffic fixed while independently intervening on environment and event. A two-phase generator first produces a layout-conditioned reference image, then performs reference-guided video editing; a three-layer reference / objective / VLM-panel protocol filters the output. Each quartet supports thresholded and continuous diagnostics for false-positive purity, event faithfulness, environment-induced drift, and held-out factor compositions. Thresholded scores are paired with continuous ones so that a conservative detector cannot suppress scores everywhere and look robust. The same audit drives training: CGAA-IC samples weak factors weighted by measured brittleness and adds quartet-level purity, consistency, and faithfulness losses; the inference graph is unchanged. On CCTV accident detection, the audit shows that clean AUC does not predict robustness rankings: a VideoMAE detector with $91.4\%$ AUC on TAD scores only $72.3\%$ CPS on the night slice of CARVE-Q. CGAA-IC raises purity from $84.2\%$ to $89.9\%$ and cross-dataset average AUC from $80.3\%$ to $84.5\%$.

Reinforcement learning with verifiable rewards (RLVR) has become the dominant paradigm for post-training large language models (LLMs) on complex reasoning and multi-step agentic tasks. However, the standard scalar reward signal is sparse and uninformative: it conveys whether a trajectory succeeded but not which reasoning steps failed or how to correct them, severely limiting sample efficiency on long-horizon tasks where successful trajectories are rare. We propose \textbf{CARD} (\textbf{C}ritique-\textbf{A}ugmented \textbf{R}einforcement \textbf{D}istillation), a training framework that amplifies the learning signal from a \texttt{verify\_answer} tool within the GRPO objective via two complementary signals. First, a token-level self-distillation reweighting measures how much each token's generation was shaped by the expert critique: tokens that merely echo the critique text are masked, while critique-driven revision tokens are amplified. Second, a rejection-sampling distillation loss maximizes the likelihood of critique-free trajectories for responses where the critique provably improved the answer, directly accelerating the internalization of critique-guided behavior. Together, these two signals distill critique-augmented reasoning into the model's base policy without requiring additional rollouts or separate reward models. Experiments on four long-form deep search benchmarks and three multi-hop question-answering benchmarks demonstrate consistent improvements over the GRPO baseline. On the deep search benchmarks, CARD improves average performance by 40\% over the Qwen3-8B base model and by 6\% over GRPO; with Qwen3-14B, the gains are 46\% and 7\%, respectively. These results establish expert critique as an effective and scalable source of rich text supervision for reinforcement learning of long-horizon research agents. Furthermore, we show that \texttt{verify\_answer} can serve as a test-time self-revision scaffold: at inference, the tool returns only task rubric criteria---without expert feedback---prompting the model to reflect on its draft and self-revise, yielding consistent performance gains over the no-VA baseline.

Text-to-SQL systems are typically evaluated by query-level execution correctness, but this terminal signal provides little guidance about which intermediate SQL decision caused success or failure. Token-level dense supervision is also ill-suited: SQL tokens do not align with complete semantic decisions, can penalize execution-equivalent queries, and are difficult to label reliably at scale. We therefore propose CAPER, which automatically derives clause-level supervision via counterfactual intervention on the SQL abstract syntax tree (AST), enabling root-cause error localization for reward modeling; the resulting data is used to train CAPER-9B, a lightweight Clause-PRM that provides clause-boundary feedback for policy optimization and candidate verification. Experiments on BIRD and Spider show that clause-aligned supervision not only improves execution accuracy, achieving up to a 15.3\% relative EX improvement over GPT-5.4, but also strengthens failure-localization capability, reaching 84.53\% accuracy and 90.60\% MRR on held-out failures.

Layout generation for real-world facilities is a challenging problem, requiring reasoning over irregular site boundaries, heterogeneous orientations, access-aware placements, and motion-planning feasibility. Yet, most existing layout benchmarks in the generative AI space target simpler placements over rectangular domains and rely on distributional metrics such as FID and IoU that reward conformity to dataset priors, thus discounting design innovation. Motivated by these gaps, we introduce $\textbf{ALPS-Bench}$, a benchmark of $1,000$ professionally annotated real-world facility layouts paired with an instance-specific scoring protocol grounded in a structured design manual. As a strong baseline for $\textbf{ALPS-Bench}$, we propose $\texttt{CANDO}$, a training-free multi-agent framework in which specialized agents iteratively refine layouts through a verification-grounded loop, concentrating reasoning on strategic spatial decisions. We demonstrate that $\texttt{CANDO}$ surpasses state-of-the-art baselines on the widely adopted $\textbf{PubLayNet}$ and $\textbf{RICO}$ benchmarks by a significant margin, establishing cooperative agentic design as a broadly effective recipe for constraint-aware layout synthesis. Code and benchmark will be released upon acceptance.

Memory facilitates question answering over long videos by extracting and retrieving facts to fit within the limited context windows of multimodal LLMs (MLLMs). Existing solutions typically extract independent memory entries from fixed-length video clips and thus cannot capture high-level semantics that need to be summarized over extended time periods, such as character traits and relations. Moreover, they rely solely on similarity-based retrieval and may fail to retrieve the fine-grained details required for question answering. To tackle these problems, we propose CAM, featuring continuous extraction for high-level semantics and \textit{adaptive querying} for fine-grained details. In particular, CAM stores the entities and relations extracted from video clips in a knowledge graph. To capture the high-level semantics of each entity or relation, CAM summarizes the local subgraph of the target entity or relation once the subgraph reaches a predefined size. To retrieve the fine-grained details required for question answering, CAM supports multiple search methods, such as knowledge graph traversal, video re-watching, and audio listening, and utilizes a planner-executor-verifier pipeline to adaptively compose these search methods according to question intent. Evaluations on three benchmarks show that CAM outperforms SOTA baselines and improves their accuracy by up to 23 percentage points.