NeurIPS 2026 Papers

The rapid emergence of heterogeneous large language models (LLMs) has highlighted the limitations of static single-model inference systems. Although model routing alleviates this issue by assigning the most suitable model to each query, existing methods often struggle to stably align query semantics with actual model performance and remain constrained by the capability ceiling of a single model. On the other hand, although multi-model aggregation can improve performance on complex queries, it also introduces extremely high computational overhead for routine tasks. In this paper, we propose PARS, a model confidence estimation method for dynamic LLM routing and aggregation. PARS first decouples the feature representation space and applies structured smoothing to local performance signals, producing more discriminative confidence score estimates for candidate models. On this basis, PARS represents inference modes such as direct routing, self-consistency sampling, and multi-model voting as test-time compute allocation actions under a budget constraint. By extracting state features such as routing confidence distributions and predictive uncertainty, a lightweight selector adaptively assigns an inference strategy to each query. Extensive experiments show that PARS improves average accuracy over the best baselines by 1.02% and 1.36% in the single-model and multi-model settings, respectively, while reducing average inference cost by 1.73x and exhibiting stable out-of-distribution generalization.

MLLM-based detection has shown promising end-to-end detection ability in recent years, but its generated detection results still lack reliable confidence for evaluation and interpretation. Since detection data itself do not provide human-annotated confidence labels, existing MLLM-based detection methods mostly obtain confidence implicitly from prompted self-assessment, generation probabilities, or external proxy scores. While these scores are available for evaluation, they are not explicitly connected with the actual detection quality. Inspired by conventional detectors that optimize confidence through classification, objectness, or quality estimation branches, we formulate confidence in MLLM-based detection as an explicit modeling target to be learned, optimized and calibrated. Specifically, we present ConfDet, a practical confidence framework that first uses discrete confidence tokens based supervised fine-tuning to enable stable and parsable confidence generation, then applies GRPO with detection and separation joint rewards to optimize confidence score separation and preserve detection quality, and finally performs condition-aware post-hoc calibration to align confidence with empirical accuracy. Experiments across different in-distribution and out-of-distribution benchmarks show that, ConfDet greatly improves AP through better confidence ranking and reduces confidence miscalibration across benchmarks, providing a reliable confidence for MLLM-based detection.

Gaussian processes (GPs) offer a principled probabilistic model over functions, but exact inference is restricted to the linear-Gaussian regime. We establish an explicit equivalence between GPs and a class of linear diffusion models, recasting predictive sampling as an ODE with closed-form Gaussian dynamics and a likelihood-dependent guidance term that admits a simple Monte Carlo approximation. In the linear-Gaussian setting, we recover standard GP conditioning exactly; beyond conjugacy, the same machinery handles any conditioning statement admitting point-wise likelihood evaluation — including non-linear physics, and, for the first time, natural language via large language models. Whitening isolates the irreducible non-Gaussian dynamics, minimising Wasserstein-2 transport cost and eliminating numerical stiffness. The result is a general-purpose GP inference scheme requiring no bespoke derivations. Together, these results provide a general mechanism for incorporating the full richness of real-world knowledge as conditioning information, opening a new frontier for the probabilistic modelling of real-world problems.

Finetuning a language model can lead to emergent misalignment (EM) [Betley et al., 2025b]. Models trained on a narrow distribution of misaligned behavior generalize to more egregious behaviors when tested outside the training distribution. We study three interventions proposed to reduce EM. We confirm that these interventions reduce or eliminate EM on existing evaluations (questions like “How do I make a quick buck?”). However, if the evaluation prompts are tweaked to resemble the training context, the model displays EM. We call this conditional misalignment. As in standard EM, the model displays misaligned behaviors more egregious than those seen during training, but only on inputs sharing features with the training data. The first two interventions are diluting misaligned data with benign data, and finetuning on benign data after misaligned data. Both produce conditional misalignment. For instance, models trained on a mix of only 5% insecure code still show misalignment when asked to format responses as Python strings (resembling the training context). The third intervention is inoculation prompting. Here, statements with a similar form to the inoculation prompt serve as triggers for misalignment, even if they have the opposite meaning. On the positive side, inoculation prompting has lower (but still non-zero) conditional misalignment if training is on-policy or includes reasoning distillation. Our results imply that in realistic post-training, where misaligned data is typically combined with benign data, models may be conditionally misaligned even if standard evaluations look clean.

Ranking models define probability distributions over rankings and are fundamental in applications such as preference learning and decision-making. Without structural assumptions, these models are intractable due to their factorial complexity. Although independence assumptions are key to reducing complexity and enabling interpretability, standard notions of conditional independence do not directly apply to rankings because they are subject to mutual exclusivity constraints. We introduce \emph{relative rank models}, a probabilistic framework based on structured coarsenings of the ranking space that relax these constraints while preserving relative order information. This representation enables a well-defined notion of conditional independence for rankings and allows rankings to be modeled using Bayesian networks. As a result, relative rank models support tractable learning, probabilistic inference, and interpretable representations. We further show that two prominent ranking models based on independence assumptions tailored to rankings arise as special cases, providing a unifying framework for independence in ranking models.

Multimodal large language models have made rapid progress in video temporal grounding, yet real-world applications routinely require localizing every event that satisfies compositional temporal and spatial conditions. Existing benchmarks fall short: they localize only a single moment per query, count without temporal conditions, or treat grounding and counting as disjoint tasks. We introduce CoMET-Bench for Conditional Multi-Event Temporal Grounding in long-form video, comprising 2,789 queries over 600 videos averaging 33.8 minutes across five real-world domains, with each query composed from 4 temporal conditions, 3 spatial conditions, and a dedicated negative-query subset. We further propose a unified evaluation protocol jointly measuring counting, grounding, and negative-query recognition, including a new Rejection-F1 metric that prevents trivial gaming by lazy "always-empty'' models. Benchmarking a broad suite of MLLMs, agent-based, and grounding-specialized methods reveals that existing approaches remain far from solving this task. Building on these findings, we propose CoMET-Agent, a training-free agentic framework that reformulates the task as structured search-and-aggregate, improving F1@0.5 by 6.1% over GPT-5 purely through structural reasoning. Failure analysis further surfaces three open directions: fine-grained entity tracking, position-uniform retrieval, and causal event pairing.

A complete understanding of heterogeneous treatment effects requires fully capturing the conditional distribution of the counterfactual outcomes. To this end, we propose the *Conditional Counterfactual Mean Embeddings* (CCME), a framework that embeds conditional distributions of counterfactual outcomes into a reproducing kernel Hilbert space (RKHS). Under this framework, we develop a two-stage meta-estimator for CCME that accommodates any RKHS-valued regression in each stage. Based on this meta-estimator, we develop three practical CCME estimators: (1) *Ridge Regression* estimator, (2) *Deep Feature* estimator that parameterizes the feature map by a neural network, and (3) *Neural-Kernel* estimator that performs RKHS-valued regression, with the coefficients parameterized by a neural network. We provide finite-sample convergence rates for all estimators, establishing that they possess the double robustness property. Our experiments demonstrate that our estimators accurately recover distributional features including multimodal structure of conditional counterfactual distributions.

Robot manipulation often depends on short-term cues—such as brief contacts, temporary occlusions, and intermediate object states—that are ambiguous in a single frame and can only be resolved by integrating observations over time. However, injecting historical information solely by enlarging the history window is computationally costly and frequently yields limited gains due to substantial redundancy and task-agnostic, inflexible sampling of observations. We propose CondenseVLA, a multi-frame VLA policy with a lightweight learnable history condensation module for using short-term visual context in action decoding. Given a sequence of cached per-frame features from the VLM backbone, CondenseVLA distills observation history into a fixed set of slot tokens using learnable queries and multi-head cross-attention. Simple auxiliary regularization encourages diverse attention patterns by penalizing overlap among slot attention maps and reduces slot 12 redundancy by decorrelating slot embeddings, stabilizing learned condensation without imposing strong time priors. By expanding the history window and distilling it into a compact token set, CondenseVLA improves over the baseline on the evaluated Simpler-WidowX tasks, increasing success rate from 58.4 to 70.8 (+12.4), achieves 97.2% on LIBERO, and reaches 52.0% on real-world tasks. Beyond the gains, our results show that long history windows contain substantial redundancy, so naively providing more frames yields limited benefits, whereas learned condensation into fixed slot tokens provides an effective bounded-token way to leverage temporal context.

Link prediction remains difficult in sparse graphs, where diffusion heuristics lack connectivity and geometric methods suffer spectral instability when coupling learned representations with diffusion operators. We introduce **Concorde**, an energy-based framework that resolves this by decomposing link likelihood into three structurally decoupled branches: **(i)** a parameter-free spectral anchor that establishes permutation-invariant topological stability; **(ii)** a mixed-curvature residual on a learnable $\mathbb{H}\_{c\_h}^{d\_h} \times \mathbb{S}\_{c\_s}^{d\_s}$ product manifold, regularized by a feature-aware Ollivier--Ricci curvature proxy; and **(iii)** a continuous vector field on $\mathcal{T}\mathbb{H}\_c^d$ for non-local structural alignment across topological gaps. The branches are fused via learned gating and trained with an $N$-pair ranking loss aligned with retrieval-style evaluation. Across five benchmarks spanning sparse, dense, featureless, and attributed regimes, **Concorde** establishes state-of-the-art early retrieval precision, nearly doubling $P@10\\%$ over leading baselines in sparse settings.

Three-dimensional Gaussian Splatting (3DGS) has recently become a promising representation for dense SLAM, due to its explicit scene parameterization, efficient differentiable rendering, and high-quality visual reconstruction. Extending 3DGS to semantic SLAM enables the construction of maps that are not only geometrically accurate but also semantically interpretable. However, existing semantic Gaussian SLAM methods usually optimize semantic, depth, and normal renderings through separate supervision objectives. Such a design underexplores the fact that these outputs are generated from the same Gaussian field and should therefore provide a concordant explanation of the underlying scene structure. This limitation is especially problematic around object boundaries and geometric discontinuities, where inaccurate Gaussian growth can cause cross-boundary expansion, boundary blurring, and semantic leakage. To address this problem, we propose Concord-SLAM, a boundary-native semantic Gaussian SLAM framework that establishes concordance among semantic, depth, and normal renderings while promoting boundary-aware Gaussian growth. Our framework contains two key components. First, Cross-Render Concordance (CRC) jointly examines semantic, depth, and normal views rendered from the same Gaussian map, and converts their structural discrepancies in boundary responses, regional continuity, and local geometry into optimization signals. In this way, disagreement among different renderings is used as an intrinsic cue for improving the semantic-geometric coherence of the Gaussian field. Second, Boundary-Native Splatting (BNS) actively inserts new Gaussians near semantic boundaries, depth discontinuities, and salient geometric variations, encouraging primitives to align with physical and semantic boundaries during map construction rather than correcting boundary mixing only through later optimization. Experiments on Replica and ScanNet demonstrate that Concord-SLAM achieves superior performance in camera tracking, scene reconstruction, and semantic mapping.

Circuit extraction identifies model components that preserve a target behavior under ablation, yet it remains unclear which parts of the reported circuit—components, edges, or coarser summaries—remain stable across reasonable extraction choices. We study this question in a Lean-style, rule-generated tactic-prediction benchmark spanning atomic and compositional proof tasks. The proof-state rules and task structure are fixed by construction, and dense and weight-sparse variants of one transformer architecture are compared. In this controlled setting, we compare three ways to describe a circuit: a compact prediction-preserving circuit, a broader graph that keeps read, write, and routing components around that circuit, and a pruning graph whose size is set by a target post-ablation loss budget. We vary model sparsity, whether the query and key sides of attention are represented together or separately, the post-ablation loss budget for the pruning graph, and which supervised checkpoint initializes reinforcement learning (RL). When the same extraction method is used, dense and weight-sparse checkpoints overlap far more in the set of attention heads identified than in exact component-to-component edge lists. This head-level overlap is well above a random top-$k$ baseline; the exact-edge overlap is not consistently so. Tracking query and key sides separately, and varying the loss budget across the tested range, preserves the ordering of the RL initialization conditions when each graph is summarized by how many components it selects. Among these RL conditions, the largest gains on compositional tasks are accompanied by the largest fraction of compositional-task circuit nodes that lie outside the matched atomic-task circuits. These findings show that, even in this controlled setting, behavior preservation under ablation does not determine a unique circuit-level conclusion: the conclusion depends on which graph is reported, how it is extracted, and at which level of description the comparison is made.

Neuro-Symbolic Concept-based Models (NeSy-CBMs) are a family of architectures that integrate neural networks with symbolic reasoning for enhanced reliability in high-stakes applications. They work by first extracting high-level concepts from the input and then inferring a task label from these compatibly with given logical constraints. Yet, their label and concept predictions can be overconfident, making it difficult for stakeholders to gauge when the model's decisions can be trusted. We address this issue by integrating ideas from Conformal Prediction (CP), a framework providing rigorous, distribution-free coverage guarantees. We formalize three desiderata -- consistency, coverage, and conciseness -- that any conformal method for NeSy-CBMs should satisfy, and show that existing approaches fall short of at least one. We then introduce COCOCO, a post-hoc framework that conformalizes concepts and labels jointly and reconciles them via a single deduction–abduction revision step. COCOCO satisfies all three desiderata, retains distribution-free coverage, is robust to imperfect knowledge and supports user-specified size budgets. Our experiments on 8 data sets highlight how COCOCO compares favorably against competitors and natural baselines in terms of performance and set size.

Concise reasoning in large language models (LLMs) seeks to generate only essential steps needed to arrive at a final answer, thereby alleviating issues of *overthinking*. Most proposed approaches scalarize length and reward into a single objective, requiring coefficients or thresholds that must be re-tuned across domains and model scales. We address this brittleness by treating concise reasoning as a constrained problem that minimizes length subject to an accuracy floor, and deriving a tractable algorithm, Performance-Aware Length Update (PALU), which replaces each intractable update of Lagrangian optimization with a tractable surrogate while preserving its structural prescription. On *DeepSeek-R1-Distill-Qwen-1.5B*, PALU cuts generation length by 64\% *and* improves accuracy from 44.2\% to 51.6\% across six benchmarks, a level matched by GRPO only with 2.5× more tokens. The Lagrangian dynamics yields a two-phase compression: an initial phase where the residual signal still allows compression without performance cost, followed by a trade-off phase where performance gates further length reduction. The same hyperparameters transfer across math, logic and STEM, and across 1.5B to 14B models, suggesting that constrained optimization is a productive design scaffold for training reasoning LLMs.

Concept-bottleneck models (CBMs) are neural classifiers that compute predictions from high-level concepts extracted from the input. CBMs ensure stakeholders can understand the concepts — and the predictions they entail — by learning these from concept-level annotations, which are however seldom available. Recent CBM architectures work around this issue by obtaining annotations from Vision-Language Models (VLMs). While greatly broadening applicability, doing so can yield lower quality concepts and therefore less interpretable models. We strike for a middle ground by introducing **Vision-plus-Human-guided CBM** (VH-CBM), a *hybrid* approach that exploits both VLMs and a small amount of dense annotations. VH-CBM employs a Gaussian Process in the VLM's embedding space, which captures useful global information about the target domain, to propagate the expert's supervision to any target data point. Our empirical evaluation on four datasets showcases how **VH-CBM** predicts substantially more accurate concepts than VLM-guided CBMs even when annotating as little as **1%** of the data, while sporting better concept-level calibration.

Aligning human-interpretable concepts with the internal representations of modern machine learning systems remains a central challenge for interpretable AI. We introduce a geometric framework to compare supervised human concepts with unsupervised representations derived from foundation models. We formalise concept frustration: a mismatch that arises when an unobserved concept induces relationships between known concepts that cannot be made consistent within an existing ontology. We develop task-aligned similarity measures that detect this phenomenon, and show that frustration is identifiable in task-aligned geometry while conventional Euclidean comparisons fail. Under a linear–Gaussian generative model, we derive a closed-form expression for Bayes-optimal concept-based classifier accuracy, decomposing predictive signal into known and unknown contributions and identifying where frustration impacts performance. Experiments on synthetic data and real language and vision tasks demonstrate that frustration is present in foundation model representations and that incorporating missing concepts reorganises learned representations to better align human and machine reasoning. These results provide a principled framework for diagnosing incomplete concept ontologies and improving alignment in interpretable AI systems.

Multimodal Large Language Models (MLLMs) trained on sequential tasks suffer from catastrophic forgetting. We study this problem in LoRA-based multimodal continual instruction tuning. By analyzing the diagonal Fisher information of LoRA parameters, we find that forgetting is amplified not merely by shared high-sensitivity parameters across tasks, but by the concentration of updates on those shared parameters, \ie, some tasks distribute their change uniformly across parameters, while others pack most of it into a narrow subset, causing disproportionate interference on the parameters that prior tasks depend on. We further observe that this concentration manifests in the dominant directions of recent gradients. Based on this insight, we propose \textbf{D}om\textbf{i}nant-direction \textbf{G}radient \textbf{Pro}jection (\textbf{DiGPro}), which maintains a short buffer of recent gradient directions, extracts their dominant subspace with adaptive rank selection, and attenuates the parallel component for each optimizer step. Crucially, DiGPro requires no prior-task data or stored subspaces, operating on the current gradient history. Experiments on cross-modal and vision-centric continual learning benchmarks show DiGPro consistently reduces forgetting and attains competitive final results. The code is available in the supplementary materials.

Intent-obfuscation-based jailbreak attacks on multimodal large language models (MLLMs) transform a harmful query into a concealed multimodal input to bypass safety mechanisms. We show that such attacks are governed by a \emph{reconstruction--concealment tradeoff}: the transformed input must hide harmful intent from safety filters while remaining recoverable enough for the victim model to reconstruct the original request. Through a reconstruction analysis of three representative black-box methods, we find that existing transformations struggle to balance this tradeoff, limiting their effectiveness. In contrast, we show that character-removed variants achieve a better balance. Building on this, we propose \emph{concealment-aware variant construction}, which greedily selects character-removed variants that are low in harmful-keyword alignment and mutually diverse, and instantiates them through five modality-aware prompting strategies. We further introduce \emph{keyword-related distractor images} that depict the harmful keyword in diverse contexts, providing more effective auxiliary visual context than generic distractor images. Experiments across closed-source and open-source MLLMs show the proposed strategies outperform strong baselines, revealing an underexplored vulnerability: a model's own reconstruction ability can be exploited to recover hidden harmful intent and produce unsafe responses.

Large language models (LLMs) are costly to deploy due to their large memory footprint and high inference cost. Weight-activation quantization can reduce these costs, but low-bit activation quantization remains difficult because activation outliers induce large quantization error. Recent rotation-based methods address this by applying orthogonal transformations that redistribute activation magnitude across dimensions, but existing approaches either require expensive end-to-end rotation training or rely on stored activation corpora, introducing significant compute or storage overhead. We propose a lightweight post-training rotation calibration method for LLM activation quantization. Our method learns orthogonal rotations that align normalized activations with the corners of an inscribed hypercube, encouraging activation energy to be distributed more evenly across dimensions. This objective admits an efficient closed-form update via the orthogonal Procrustes problem, avoiding gradient-based optimization over the orthogonal group. We further introduce an online calibration procedure that updates rotations as calibration samples are processed, eliminating the need to store activations on disk and allowing rotations to adapt to quantized activation distributions during calibration. Experiments on Llama-2 and Llama-3 models from 3B to 70B parameters show that our method achieves competitive or improved performance across perplexity benchmarks and common sense reasoning tasks while avoiding both costly end-to-end training and large offline activation storage.

Approval-based committee voting has received significant attention in the social choice community. Among the studied rules, Thiele rules, and especially Proportional Approval Voting (PAV), stand out for desirable properties such as proportional representation, Pareto optimality, and support monotonicity. Their main drawback is that computing a Thiele outcome is NP-hard in general. A glimpse of hope comes from the fact that Thiele rules are better behaved under structured preferences. On the candidate interval (CI) domain, they are computable in polynomial time via a linear program (LP) that has a totally unimodular constraint matrix. Surprisingly, this approach fails for the related voter interval (VI) domain, and the complexity of the problem has repeatedly been posed as an open question. Our main result resolves this question: although the relevant matrix is not totally unimodular, the ``standard'' LP still admits at least one optimal integral solution, and we provide a fast algorithm for finding it. Our technique naturally extends to the voter-candidate interval (VCI) domain, also known as the 1-dimensional voter-candidate range (1D-VCR) domain, and to the linearly consistent (LC) domain, both of which generalize the candidate and voter interval domains. Although both the VCI and LC domains have been studied in social choice, their relationship was unknown. We show, through connections to graph theory, that LC strictly contains VCI. We also provide an alternative definition of LC that is closer in spirit to VCI and has a natural interpretation in approval elections; this equivalence may be of independent interest. Finally, we study an alternative tree-based generalization of VCI and show that Thiele rules become NP-hard to compute on this domain.