NeurIPS 2026 Papers

We argue that cognitive control, the capacity to hold multiple divergent candidate responses and reconcile them through categorical inhibition rather than soft blending, should be a first-class computational object in transformer-based language models. We develop this claim at two levels. At the \emph{signal level}, interference paradigms reveal a characteristic failure signature: rather than adjudicating between competing responses, current transformer-based language models blend them, and accuracy collapses as interference scales. At the \emph{goal level}, sustained agency requires maintaining concurrent goals, enforcing priority during signal level adjudication, and revising the goal set as evidence accumulates about whether those goals are being successfully pursued; current systems approximate this only through external scaffolding. The architectural gap revealed at the signal level is the same gap that will block transformers from scaling to the kind of goal modulated and self regulating cognition that long-horizon autonomous agents require. We define concrete criteria for what counts as architecturally explicit cognitive control, show that neither the dominant training pipeline nor reasoning-token training implements it, and commit to predictions that distinguish our position from scale-based optimism.

Text-to-image diffusion models fail on semantically rare attribute--object compositions, but it is unclear whether this brittleness arises during denoising or is already present in conditioning representations. We study this question at three levels of granularity. First, using a controlled stress protocol across SD1.5, SDXL, and SD3, we show that a Compositional Stress Index derived from text-side non-factorization residuals separates anchor from stress prompts with perfect discrimination (AUC${=}1.0$) cross-model, establishing a robust pre-denoising risk axis. Second, embedding-level low-rank repair reduces representation residuals without any positive semantic improvement, while a downstream cross-attention intervention succeeds on the same subset---a site-sensitive boundary between diagnosis and control. Third, we decompose this boundary at block-group resolution under three intervention modalities. Diagnostic accessibility peaks at the deep encoder (D2) across all conditions, but token-level intervention of either sign produces no measurable improvement at any block. Only broad cross-attention ablation moves semantics at D2, while decoder blocks respond to selective boost and collapse under ablation. Together, these results establish that compositional defects are diagnosable before denoising, but \textbf{where to look is not how to fix}: the diagnostic site and the control site require different intervention modalities, and even at the correct site, the wrong modality is futile.

The efficiency and fidelity of 3D Gaussian Splatting critically depend on adaptive density control, which grows and suppresses Gaussian primitives throughout optimization. However, the standard pipeline still relies on heuristic view-space gradients, scale thresholds, and global opacity resets, rather than explicitly assessing the loss-reducing effect of each split, clone, or opacity update. This mismatch can allocate primitives to already-sufficient regions, miss under-reconstructed structures, and destabilize training through blanket opacity changes. In this work, we recast adaptive density control as a collection of local loss-reduction decisions. We derive efficient Taylor-based per-Gaussian surrogates that predict the objective change caused by three primitive-level operations: splitting, cloning, and opacity decay. This yields unified loss-aware framework with three components. First, we extend second-order splitting to a normalized joint position-scale space, enabling children to adapt not only their locations but also their anisotropic extents. Second, we formulate duplicate cloning as a surrogate minimization problem, selecting only Gaussians whose duplicated contribution is predicted to improve reconstruction and assigning each child a principled opacity. Third, we replace global opacity reset with selective soft opacity decay that weakens harmful primitives while preserving useful ones. Across standard 3DGS benchmarks, our method produces more targeted primitive allocation and achieves improved rendering quality with comparable or smaller Gaussian budgets compared to baselines.

Do large language models represent psychological constructs internally, or do they only generate outputs that resemble psychometric structure? We address this question using a cross-persona paradigm across 14 models (Llama3 and Qwen2.5, 0.5B--14B, base and instruct). Given responses on one psychological scale from real individuals ($N = 272$), models predict responses on six other scales. At the behavioral level, LLMs reproduce human cross-scale correlation structure and systematically amplify it, with model-generated correlations exceeding human estimates even after correcting for measurement attenuation. We then examine whether this structure is reflected internally. Contrastive direction analysis reveals an organized geometry in activation space aligned with psychometric relationships. This structure emerges in large instruct models but is not observed in base models. Across models, geometry strength predicts behavioral amplification ($r = 0.68$, $p = 0.008$; partial $r = 0.69$ controlling for log size), independent of model size. To relate internal structure to output behavior, a matched activation-probing paradigm shows that representational amplification is less variable than behavioral amplification ($1.38$--$1.77$ vs $0.53$--$1.53$). A synthetic control with known ground-truth structure shows that this range survives subtraction of a ridge-probe baseline ($\approx 1.3$), with adjusted slopes still predicting behavior at $r = 0.88$. We term the systematic representation-to-behavior gap \emph{readout attenuation}. Together, these findings suggest that LLMs encode structured representations aligned with psychological constructs, while differences in output primarily reflect how these representations are read out.

A key strength of diffusion models lies in their flexibility, since their outputs can be controlled at sampling time through guidance. However, beyond simple cases such as conditional sampling, the target distribution is often left implicit, defined only through a sampling rule or a heuristic energy function. To address this, we propose Jeffrey guidance, a principled framework that extends diffusion-model control to applications beyond what standard guidance can express. It leverages Jeffrey’s rule of conditioning to update marginal distributions towards a prescribed target, preserving the conditional structure and minimally perturbing the joint distribution. We first demonstrate Jeffrey guidance by targeting a prescribed embedding distribution. With Inception embeddings as the target, this leads to substantial reductions in FID on both CIFAR-10 and FFHQ. We further apply Jeffrey guidance to fairness on CelebA-HQ, updating an unconditional diffusion model to enforce independence between attributes.

Humanoid whole-body loco-manipulation requires scalable and physically reliable demonstrations, yet existing human-object interaction datasets lack direct contact supervision and egocentric observability, often resulting in physically inconsistent interactions. We present DeXHOI, a multimodal egocentric HOI benchmark for whole-body dexterous interaction, capturing synchronized body motion, object trajectories, tactile sensing, and egocentric RGB-D across diverse spatial layouts and long-horizon activities. To improve physical consistency, we introduce a tactile-guided refinement pipeline that leverages measured pressure to correct hand-object interactions while preserving global motion fidelity. We further establish an Egocentric HOI Recovery Benchmark for contact-aware reconstruction from first-person observations, along with EgoTacti, a tactile prediction model. Together, EgoTacti provides a scalable and physically grounded foundation for learning humanoid whole-body dexterous interaction, pointing toward low-cost egocentric data collection for embodied intelligence.

Spiking neural networks (SNNs) offer a compelling path to energy-efficient, biologically plausible computation. Yet, a persistent gap remains: fully spiking models are difficult to train and incompatible with standard deep learning frameworks, while ANN-to-SNN conversion and hybrid approaches sacrifice either biological plausibility or deployment efficiency. We address this gap by introducing SpikeCore, a recurrent spiking unit designed to be natively compatible with ANN training frameworks, differentiable, gradient-friendly, and composable within standard pipelines, while retaining event-driven, spike-based computation. SpikeCore achieves this through learnable membrane decay and spike-history-dependent threshold adaptation, enabling rich temporal dynamics without surrogate gradient instability. Building on SpikeCore, we present a full spatio-temporal pipeline comprising: a two-stage learnable temporal delay hierarchy motivated by differential retinal ganglion latency and axonal conduction delays; a Multi-Branch Squeeze-Excitation (MBSE) backbone that extracts spatial features at multiple scales with channel-wise recalibration, using $11.8\times$ fewer parameters than equivalent residual architectures; and a Parameter-Free Temporal Gain Control (PFTGC) module that modulates each channel's temporal response without added parameters. Evaluated across five diverse benchmarks, the proposed architecture converges $5$--$6\times$ faster and requires ${\approx}3\times$ fewer timesteps than competitive baselines while matching or surpassing state-of-the-art models on accuracy.

The feed-forward propagation through neural networks is crucial for training and inference. To trade off efficiency and accuracy, it has been proposed to consider intermediate outputs of hidden states as data-dependent regularization during training and, ultimately, as early-exits during inference. In this work, we analyze the joint optimization of loss functions computed for the output layer and intermediate layers using optimal control theory. We prove that ResNets trained in this joint fashion exhibit the so-called turnpike phenomenon, this implies that the number of layers required to reach a prescribed loss level is bounded from above independently of the overall network depth. This induces a similarity property across the internal representations of early-exiting networks of varying depths. We connect the turnpike to the overthinking phenomenon, and provide theoretical insights of the success of early-exit criteria derived from softmax probabilities, specifically entropy and max-probability. In a range of experiments, we demonstrate the existence and effect of turnpikes and the resulting similarity in the early-exit performance for ResNets of varying depths.

Certifying authorities for safety-critical software audit recovered trace matrices module by module, not on average. Every existing distribution-free calibration delivers a marginal certificate, and we show the gap to the auditor's quantity is structural and unbounded. Under within-block exchangeability with cross-block heterogeneity in the conditional null score distribution, every procedure controlling marginal $\mathrm{FDR}$ at level $\alpha$ on a bipartite trace matrix with $K$ source modules admits worst-module false discovery proportion at least $1{-}\alpha{-}O(1/K)$ with probability at least $1{-}2/K$. We name this the block catastrophe and close it with $\mathrm{CERTRA}$, which wraps any black-box scorer, computes a within-block rank-indicator conformal $e$-value, and applies $e$-$\mathrm{BH}$ separately within each source module. The construction yields finite-sample per-module $\mathrm{FDR}$ control under arbitrary intra-module dependence, with matching power up to a rank-quantization factor. We release IndTrace-2026, a 12{,}406-link composite across aerospace, automotive, and medical-device specifications with professional ASIL/DAL/SIL annotation (Cohen's $\kappa{=}0.86$). $\mathrm{CERTRA}$ holds worst-block $\mathrm{FDP}$ at 0.118 where split-conformal drifts to 0.413, retains 0.741 of true links versus 0.572 for the strongest fair-operating-point baseline, and reduces audit deficiency 3.3-fold. Per-module distribution-free calibration is what an industrial audit can sign.

The *synthetic control method* (SCM) estimates causal effects in panel data with a single-treated unit by constructing a counterfactual outcome as a weighted combination of untreated control units that matches the pre-treatment trajectory. In this paper, we introduce the **targeted synthetic control (TSC)** method, a new two-stage estimator that directly estimates the counterfactual outcome. Specifically, our TSC method (1) yields a targeted debiasing estimator, in the sense that the targeted updating refines the initial weights to produce more stable weights; and (2) ensures that the final counterfactual estimation is a convex combination of observed control outcomes to enable direct interpretation of the synthetic control weights. TSC is flexible and can be instantiated with arbitrary machine learning models. Methodologically, TSC starts from an initial set of synthetic-control weights via a one-dimensional targeted update through the weight-tilting submodel, which calibrates the weights to reduce bias of weights estimation arising from pre-treatment fit. Furthermore, TSC avoids key shortcomings of existing methods (e.g., the augmented SCM), which can produce unbounded counterfactual estimates. Across extensive synthetic and real-world experiments, TSC consistently improves estimation accuracy over state-of-the-art SCM baselines.

Model distillation can transfer not only intended capabilities but also hidden traits from the teacher. A teacher biased by a system prompt can generate semantically clean training data, such as numeric sequences, that still causes a downstream student to inherit the hidden preference, a phenomenon known as $\textit{subliminal learning}$. Prior work establishes this phenomenon, but the transfer mechanism remains unclear, making targeted mitigation difficult. We propose and validate $\textbf{trait-direction drift}$ as the mechanism underlying subliminal learning: $\textbf{(1)}$ on the teacher side, biased teacher samples retain a measurable preference gap toward the target trait even when their semantic content is unrelated to that trait; $\textbf{(2)}$ on the student side, under a low-rank logit-linear approximation, samples with larger preference gaps induce stronger trait-aligned updates during supervised fine-tuning, and these updates accumulate over training. Guided by this mechanism, we propose $\textbf{probe-space corridor}$, a regularizer that constrains drift along a calibrated trait readout during distillation. The method substantially reduces hidden-trait transfer while preserving task performance: for example, it lowers malicious-response transfer from 29.55% to 6.47% with low main-task accuracy cost, and consistently suppresses animal-preference transfer across the main Qwen-7B preference settings. These results provide both a validated mechanistic explanation of subliminal learning and a targeted recipe for more controllable distillation.

While large reasoning models (LRMs) can solve complex tasks via generating extended Chain-of-Thought traces (Long-CoT), longer traces often increase confidence without increasing correctness. In this work, we formalize such failure via a \emph{self-conditioned evidence} model: generated tokens provide information about the model's current working hypothesis rather than the ground-truth answer. This yields a simple diagnostic principle: entropy, fluency, and answer stability can certify within-trajectory self-consistency while remaining causally disconnected from correctness. Building on these findings, we propose \textbf{Faithful Pandora Inference (FPI)}, a minimal training-free controller that keeps the base LRM frozen and fits only a small calibration map on held-out data. At each checkpoint, FPI estimates calibrated correctness $r_t$ and the marginal evidence gain $m_t$ of continuing the current trajectory. It stops only when $r_t$ reaches a user-specified reliability target; if $r_t$ remains insufficient and $m_t$ falls below cost, it reallocates the remaining budget to a diversified restart. Across six reasoning benchmarks and multiple DeepSeek-R1 distilled models, FPI improves the accuracy-efficiency frontier, reduces expected calibration error, and improves selective accuracy on high-confidence outputs. Ablations on confidently wrong trajectories show that calibrated stopping and marginal-gain restarts address different failure modes and that the gains are not explained by additional samples alone.

Primitive splatting has become a powerful representation for efficient differentiable rendering, but existing density-control strategies remain largely heuristic and often suffer from uncontrolled primitive growth. In this work, we present \textbf{Moment-Preserving Density Control (MPDC)}, a general theoretical framework for densification in primitive splatting. Unlike conventional density-control strategies that \emph{select first and split later}, MPDC follows a \emph{split-then-select} paradigm. Our key insight is that a principled split should first preserve the current rendering, rather than immediately perturb the image. By matching local moments, MPDC decomposes a primitive into multiple offspring while keeping the rendered output, and hence the loss, unchanged up to higher-order error. Although such a split does not directly reduce the loss, it changes the parameter space: a saddle point in the original parameterization may no longer remain a saddle point after splitting. This provides a distinct saddle-escaping mechanism from prior approaches that move primitives along negative-curvature directions. Building on this view, we derive closed-form splitting rules for different primitive parameterizations and then select only primitives whose moment-preserving splits reduce an upper bound of the loss through a splitting-matrix criterion. Experiments on three diverse primitive splatting methods show that MPDC substantially reduces primitive counts without sacrificing rendering quality, while improving memory efficiency and rendering speed. Code will be released upon publication.

Language models are increasingly used in text generation, decision support, and automated interaction, where their behavior must be controlled in a localized, reversible, and selective way. Existing methods either update shared parameters or constrain generation through external interfaces, leaving open whether frozen models contain compact internal control. We introduce sparse internal control, a framework for steering target behavior by applying inference-time interventions to a small set of internal model nodes. The framework formulates control through local target controllability, where candidate site–direction pairs are characterized by their behavioral effects under target, preservation, and energy constraints. This yields two complementary selectors: Driver-OMP, which selects nodes that sparsely reconstruct a desired behavioral displacement, and Coverage-Driver, which favors nodes whose effects cover prompts stably. We further propose a six-axis control-evaluation protocol measuring reachability, signed reversal, dose response, feedback controllability, off-target preservation, and held-out reuse. Across IOI, MMLU MCQA, and sentiment steering on models from GPT-2 small to Qwen3-4B, sparse driver sets act as executable actuators: they steer target behavior, respond monotonically to dose, support closed-loop control, and preserve unrelated next-token behavior. On refusal steering, the same protocol exposes a boundary case that passes all evaluated axes except strict reachability. Held-out reuse separates prompt-specific controls requiring refitted strengths from population-level controls that transfer as frozen interventions. These results show that model control can move beyond output-side steering: mechanistic internal variables can be selected, certified, and reused as localized control handles for frozen language models.

Reinforcement Learning (RL) algorithms often rely on data collected from a target environment, which can be expensive and noisy. Although simulators can generate large amounts of additional data, standard simulation mixing approaches—which directly combine simulated and real experience—can degrade performance due to simulator inaccuracies, especially when they bias the solution away from the optimal policy. We propose a conservative simulation-aided learning method based on control variates, where simulated data is used solely to reduce variance in the learning process rather than directly shaping the learning target. We develop our approach for Policy-Evaluation, which is a fundamental building block of RL algorithms. Then we combine it within the actor-critic framework towards policy improvement, as well as extend it towards offline RL scenarios. Experiments on multiple MuJoCo benchmark tasks demonstrate improvements over standard simulation mixing approaches, highlighting control variates as an effective tool for simulation-aided reinforcement learning.

A universal controller for any robot morphology would greatly improve computational and data efficiency. Steps have been made towards such multi-robot control by utilizing *contextual* information about the properties of individual robots and exploiting their modular structure in the architecture of deep reinforcement learning agents. When the robots have highly dissimilar morphologies, however, this becomes a challenging problem, especially when the agent must generalize to new, unseen robots. In this paper, we posit that contextual features are often only partially available, but that they can be recovered through modular interactions. This can allow for better multi-robot control and generalization to contexts that are not seen during training. To this extent, we implement a transformer-based architecture with shared modular recurrence and evaluate its (generalization) performance on a large set of MuJoCo robots. The results show a substantial improvement in zero-shot generalization performance on robots with unseen dynamics, kinematics, and topologies, in four different environments.

As AI agents become increasingly autonomous and capable, ensuring their security against vulnerabilities such as prompt injection becomes critical. This paper explores the use of information-flow control (IFC) to provide security guarantees for AI agents. We present a formal model to reason about the security and expressiveness of agent planners. Using this model, we characterize the class of properties enforceable by dynamic taint-tracking and construct a taxonomy of tasks to evaluate security and utility trade-offs of planner designs. Informed by this exploration, we present Fides, a planner that tracks confidentiality and integrity labels, deterministically enforces security policies, and introduces novel primitives for selectively hiding and revealing information. Its evaluation on AgentDojo and WASP demonstrates that this approach enables us to complete a broad range of tasks with security guarantees.

Building humanoid robots that perform generalizable whole-body loco-manipulation in the real world remains a fundamental challenge: existing approaches either rely on heavy task-specific reward engineering, rigidly replay reference motions that fail to generalize, or depend on costly teleoperation that limits scalability. While human videos capture diverse human behaviors, the motion priors inferred from them are inherently imperfect, suffering from occlusion, contact artifacts, and retargeting errors that render them unsuitable for direct policy learning. To this end, we present SUGAR, a data-driven framework that converts diverse human videos into deployable humanoid loco-manipulation skills, without any task-specific reward engineering or reference-motion conditioning at inference. SUGAR proceeds in three coupled stages: First, a fully automated pipeline extracts scalable kinematic interaction priors including human-object motion trajectories and contact labels from diverse human videos. Second, a privileged physics-based refiner utilizes a unified mimic-style reward and a progressive state pool to transform imperfect kinematic interaction priors into physically feasible, high-fidelity skills. Third, the refined skills are distilled into a hierarchical policy that comprises a task-guided planner and a command tracker for autonomous task execution. We evaluate our method on six representative loco-manipulation tasks in both simulation and real-world humanoid hardware. SUGAR substantially outperforms reference-tracking baselines, and its performance scales clearly with the amount of human video data. It also achieves zero-shot real-world transfer with reliable closed-loop execution, autonomous failure recovery and stable long-horizon performance under external perturbations. Project Page: https://sugar-humanoid.github.io/

Continual learning suffers from catastrophic forgetting, leading to the loss of past knowledge and unstable adaptation across task streams. Existing approaches typically address this issue through replay buffers, architectural changes, or additional regularization terms. We argue that forgetting is a consequence of unstable learning dynamics, which may benefit from an optimizer-level solution. The main challenge is that learning dynamics are time-dependent, as demonstrated by empirical phenomena such as the stability gap. First, we formally prove that, for distance-based classifiers, bounded parameter drift guarantees bounded margin erosion under a local score-stability condition. Based on this insight, we propose STABLE, a primal–dual optimizer with adaptive drift control. STABLE enforces a drift budget by increasing the dual variable when the budget is violated and decreasing it otherwise. This results in high stability without penalizing plasticity. Experiments on long-horizon unimodal streams and CLIP-based multimodal streams show that vanilla PEFT baselines trained with STABLE match or outperform specialized continual learning methods. In the multimodal setting, STABLE further preserves zero-shot transfer, matching or exceeding dedicated multimodal continual learning methods.