NeurIPS 2026 Papers

Automated feature engineering (AFE) enables AI systems to autonomously construct high-utility representations from raw tabular data. However, existing AFE methods rely on statistical heuristics, yielding brittle features that fail under distribution shift. We introduce CAFE, a framework that reformulates AFE as a causally-guided sequential decision process, bridging causal discovery with reinforcement learning-driven feature construction. Phase I learns a sparse directed acyclic graph over features and the target to obtain soft causal priors, grouping features as direct, indirect, or other based on their causal influence with respect to the target. Phase II uses a cascading multi-agent deep Q-learning architecture to select causal groups and transformation operators, with hierarchical reward shaping and causal group-level exploration strategies that favor causally plausible transformations while controlling feature complexity. Across 15 public benchmarks (classification with macro-F1; regression with inverse relative absolute error), CAFE achieves statistically significant gains over the strongest AFE baseline (Wilcoxon p<0.001, large effect), reduces episodes-to-convergence, and delivers competitive time-to-target. Under both controlled covariate shifts and naturally occurring distribution shifts (Folktables), CAFE reduces out-of-distribution performance drop by up to 4x relative to a non-causal multi-agent baseline, and produces more compact feature sets with more stable post-hoc attributions. These findings underscore that causal structure, used as a soft inductive prior rather than a rigid constraint, can substantially improve the robustness and efficiency of automated feature engineering.

Video anomaly detection (VAD) is critical for identifying safety risks in surveillance and autonomous systems. Recent large-model-based VAD methods generate textual rationales and directly output anomaly scores. However, mapping rich semantic descriptions and analysis directly to absolute scalar scores can be poorly grounded and inconsistent, leading to scores that are not well aligned with anomaly severity. With the observation that for humans, it is usually easier to compare things than to rate things directly, we explore whether it also holds for large models in VAD by proposing \textbf{CA-Judge} (\textbf{C}omparative \textbf{A}nomaly \textbf{Judge}), a training-free framework that replaces direct scoring with comparative order inference. CA-Judge treats model-perceived anomaly severity as a latent coordinate, constructs this coordinate from self-comparisons among generated normal/anomalous reference events, and models how anomaly likelihood varies along this coordinate using the generated normal/anomalous reference identities. At test time, CA-Judge avoids exhaustive comparison by maintaining a posterior belief over query severity and selecting a few query--reference comparisons by their expected information about the anomaly status. The resulting comparison trace is aggregated under a Bradley--Terry likelihood and read out through the learned severity--anomaly association to produce the anomaly probability. Without fine-tuning or training data, CA-Judge substantially improves over direct-scoring baselines and surpasses prior training-free state-of-the-art methods on UCF-Crime and XD-Violence.

3D reconstruction could improve colonoscopy by estimating mucosal coverage and alerting clinicians to missed regions during screening. However, algorithm development is limited as no current datasets provide both a realistic in vivo appearance and dense, time-resolved 3D ground truth, especially under non-rigid deformation. We present C3VD-DEFCOL, a framework and dataset for evaluating deformable colonoscopy reconstruction with paired geometry and realistic texture. Starting from C3VD/C3VDv2 colon meshes and camera trajectories, we generate controlled deformations of the colon surface, including peristaltic waves and centerline motion, and render fisheye colonoscopy videos with frame-level depth, surface normals, optical flow, occlusion masks, camera poses, and time-stamped 3D meshes. We then use the rendered geometry, primarily depth, to condition an LTX-2.3-based sim-to-real translation model that produces RGB clips with in vivo-like mucosal color, texture, vasculature, and specular appearance while preserving the underlying 3D scene structure. The resulting dataset contains 110 videos from 12 unique colon mesh geometries, with varying camera trajectories, appearances, and multiple deformation regimes, including three peristaltic severity levels. We evaluate the generated videos using appearance realism, geometric consistency, and temporal consistency metrics, and use the paired ground truth to benchmark the downstream task of pose estimation in deformable 3D reconstruction. Our experiments show how pose estimation error increases with increasing deformation severity, providing a controlled stress test that is not possible with existing in vivo datasets. Overall, C3VD-DEFCOL is designed as a reproducible, quantitative evaluation platform for testing deformable 3D reconstruction algorithms, with the goal of reducing the domain gap between synthetic datasets and in vivo colonoscopy.

Multimodal large language models (MLLMs) are transforming recommender systems into content-aware decision systems. In these systems, multimodal compression and cross-modal fusion are often viewed as robustness-enhancing mechanisms against content manipulation, as compression can filter noisy signals and cross-modal consensus can suppress inconsistent anomalies. We challenge this assumption by showing that the same process can expose a compression-to-consensus (C2C) bias in ranking decisions. Signals that remain salient after compression and receive consistent support across modalities can gain disproportionate influence in candidate comparison, offering an exploitable direction for content manipulation. In this paper, we introduce C3H, a novel inference-time attack that exploits this vulnerability. Rather than relying on noisy perturbations or training-time poisoning, C3H promotes a target item by tailoring its multimodal content to align with the decision criteria implicitly expressed by the recommender. Extensive experiments demonstrate that C3H substantially increases target-item exposure while maintaining overall recommendation utility.

Maintaining character consistency over long-horizon gameplay is a critical bottleneck for deploying Large Language Models (LLMs) as Non-Player Characters. Prevalent discrete Affective State Machines (ASMs) fail due to two paradoxes: (1) the \textit{Cliff-edge Effect}, where threshold-based switching causes jarring narrative discontinuities; and (2) \textit{Auto-regressive Inertia}, where accumulated dialogue history dilutes current state instructions. We propose \textbf{C3 (Causal-Continuous Characterization)}, which couples a continuous state space with \textit{hysteresis-aware dynamics} to dampen emotional volatility, and a \textit{State-Conditioned Memory Rewriting} mechanism that re-contextualizes past events through the lens of the current relationship. On a 100-turn dynamic simulation (\textsc{Chronos-Sim}, 5{,}000 turns per condition), C3 achieves a \textbf{10.5\% absolute SAR improvement} over a strong smoothed-ASM baseline (Welch's $t$-test, $p < 0.0001$, Cohen's $d \approx 3.97$). Cross-model evaluations on \texttt{Llama-3-8B-Instruct} and \texttt{Mistral-v0.2-7B} reproduce the relative gain, indicating the mechanism transfers across backbones although absolute SAR remains backbone-dependent. A human study with 20 game designers (ICC$=0.74$) corroborates higher perceived consistency, and a separate factual audit (Fleiss' $\kappa=0.81$) confirms the Rewriter alters interpretation without corrupting facts. Deployment analysis shows a $\sim$60\% reduction in long-session token consumption.

The rapid growth of generative AI is reshaping the energy footprint of modern computing infrastructure. Hyperscale AI data centers now operate at power levels comparable to large industrial facilities, creating new challenges for grid reliability, energy procurement, and real time demand flexibility. At the same time, their large controllable loads, cooling systems, and on site storage make them potential participants in grid services such as frequency regulation. However, existing evaluation environments do not capture the coupled decision making required for grid interactive operation. We introduce C2G-BENCH, a cyber-physical benchmark for hierarchical reinforcement learning in hyperscale data centers. C2G-BENCH simulates a real-world scenario, where a 250 MW-class hyperscale data center participates in grid frequency regulation while serving realistic AI workload demand under weather-driven ambient conditions. The benchmark explicitly links 15-minute market level orchestration with 5-second physical control, enabling agents to jointly reason over regulation commitments, workload flexibility, cooling, battery dispatch, and facility safety. C2G-BENCH achieves this through two coupled Gymnasium environments: C2GMacroEnv, where a high level controller bids into the regulation market, and C2GFastEnv, where a low level controller actuates the Dynamic Voltage and Frequency Scaling (DVFS), Coolant Distribution Unit (CDU) pump speed, HVAC effort, and Battery dispatch (BESS). The simulator models a 250 MW class two zone facility using seven modular physics engines covering workload, thermal, electrical, battery, grid signal, weather, and market dynamics, driven by real workload and weather traces. We benchmark a family of rule-based, LLM-based & RL-based controllers for both the macro and fast environments, and evaluate them on four stress-test environments in multiple grid markets. Baseline results show that active low level control substantially improves tracking quality: a rule-based macro controller paired with a RL-based low level controller reduces the tracking RMSE from 2,737 kW to 229 kW compared with a macro only configuration, while maintaining full throughput and zero thermal violations. These results demonstrate that C2G-BENCH exposes meaningful cross layer tradeoffs and provides a reusable benchmark for studying safe, hierarchical control of grid interactive AI data centers.

Multimodal Large Language Models (MLLMs) have demonstrated remarkable capabilities, yet they frequently struggle with fine-grained visual perception, often suffering from hallucinations when faced with visually similar but semantically distinct instances. A major underlying cause is the conventional single-sample Supervised Fine-Tuning (SFT) paradigm, which isolates training instances and limits the model's ability to explicitly learn subtle discriminative boundaries. To address this, we propose **C2FT**, a novel **Confuse-then-Contrast Fine-Tuning** framework that enhances fine-grained perception by shifting from independent supervision to joint supervision. Specifically, C2FT assesses the MLLM's internal uncertainty to dynamically mine model-specific semantic confusions, subsequently constructing challenging multi-image groups comprising hard positive and hard negative samples. By interleaving these samples into a unified prompt, our approach forces the MLLM's internal attention mechanisms to cross-reference images and explicitly capture localized visual discrepancies. Extensive experiments on both Fine-Grained Visual Classification (FGVC) and visual question answering (VQA) benchmarks demonstrate the effectiveness of C2FT. For instance, on the Qwen3-VL-4B model, our approach achieves an average improvements of 3.87\% and 4.35\% points over standard SFT and GRPO baselines, respectively.

Token-level distillation assumes a shared categorical support, an assumption violated when teacher and student tokenizers segment the same text differently. Existing cross-tokenizer methods sidestep this mismatch by aligning token spaces, matching likelihoods along tokenizer-specific paths, or filtering shared spans, all of which can approximate the original distributional objective or discard probability mass. We argue that decoded UTF-8 bytes are common observable events shared by every text tokenizer, while token boundaries are tokenizer-specific latent transitions. Building on this view, we introduce \emph{ByteDistill}, which compares the conditional distribution of the next observable byte rather than aligning token identities. Its loss, \emph{Byte-Level Distribution Alignment} (BLDA), projects each native softmax into a 256-way observable next-byte distribution inside aligned decoded chunks, marginalizing token-ending mass as a latent boundary transition along the observed tokenization path. ByteDistill consistently outperforms supervised fine-tuning and prior cross-tokenizer baselines in both off-policy and on-policy settings, with off-policy training improving GSM8K accuracy by $1.75$ percentage points over the current state of the art.

Sparse autoencoders (SAEs) are widely used to decompose language model activations into interpretable features, but recent work finds that many features are not reproducible across random seeds. Intuitively, the more important a feature is, the more reproducible it should be. We test this hypothesis and find a surprising inversion: the 50 most important features by zero-ablation are among the least reproducible across seeds, far below dictionary-wide baselines. We trace this to residual stream anisotropy, a well-studied phenomenon in NLP embeddings that has been documented in transformer representations but not connected to SAE training. In transformer middle layers, a single principal component (PC1) explains 70--99.9\% of activation variance, and its sparse decomposition is non-identifiable: different seeds learn different features that tile the same direction. These features dominate importance rankings, explaining the inversion. We propose an intervention: bypassing PC1 during SAE training, routing it around the SAE and restoring it during inference. Across 9 models from 5 families (124M--70B parameters), this intervention raises top-50 feature recovery to 74--97\%. The recovered features show greater causal consistency under steering ($p < 0.01$ in 8 of 9 models), and shared features produce $4\times$ more accurate logit-lens predictions than orphan features. Our results reconcile recent SAE instability findings with the existence of dense model features: the dominant PC1 direction is real and reproducible, but the individual sparse coordinates used to reconstruct it are not. Once this component is separated, a reproducible core of sparse, semantically coherent SAE features emerges.

As generative video models become increasingly realistic, detecting AI-generated videos requires systems that offer both accuracy and interpretability. However, applying Multimodal Large Language Models (MLLMs) to video forensics is currently limited by outdated datasets, simplistic evaluation protocols, and a reliance on black-box classification. To address these issues, we introduce a comprehensive dataset, benchmark, and baseline model for video forgery detection. First, we present \textbf{GenBuster-200K}, a fair dataset of over 200,000 high-quality videos sourced from state-of-the-art generators, featuring diverse real-world scenarios. Second, we propose \textbf{GenBuster-Bench}, a diagnostic benchmark spanning three progressive tracks (In-Domain, Out-of-Domain, and In-the-Wild) to evaluate models across \textit{domain shifts} and \textit{generational shifts}. It also introduces an MLLM-as-a-Judge protocol to assess the quality of the generated forensic explanations. Finally, we develop \textbf{BusterX}, an MLLM baseline with RL training. Instead of direct binary classification, BusterX formulates detection as a visual reasoning task, where the generated reasoning chain serves as detector itself. Experimental results demonstrate that BusterX outperforms several leading MLLMs (e.g., Qwen3.5, Claude-Sonnet-4.6) in both detection accuracy and rationale quality.

Finite-shot quantum tomography often leaves a posterior over many density matrices compatible with the same measurement record, but fast reconstruction methods collapse this uncertainty to a single state. We frame this setting as amortized posterior sampling and introduce BuresTomFlow, a measurement-conditioned Flow Matching sampler for full-rank density matrices. The method transports base density matrices along Bures-Wasserstein paths on the positive matrix cone, projects the path back to the trace-one state space, and trains the velocity field with a Bures-aligned tangent loss. Thus the sampler is trained in the same fidelity-based geometry used to judge posterior quality. We evaluate BuresTomFlow in dense four-qubit simulations with count, shadow, and mixed measurement records. Against Cholesky, log-density, and generic Riemannian Flow Matching controls at comparable learned-sampler runtime, BuresTomFlow improves Bures distance, credible-ball coverage, and held-out observable prediction. In the main comparison across multiple seeds, mean Bures distance decreases from $0.618$ for the Cholesky baseline to $0.367$ for BuresTomFlow, and mean coverage error decreases from $0.203$ to $0.100$. These results show that the value of Bures geometry is not merely enforcing physical states, but producing calibrated posterior uncertainty when finite-record tomography cannot be reduced to one reconstruction.

Recently, deep neural networks on manifold-valued representations have garnered significant attention across various machine learning applications. One recent focus is the generalization of Euclidean fully connected (FC) and convolutional layers to non-Euclidean geometries. However, previous approaches typically focus on a few selected manifolds and rely on specific properties of the target manifold. In contrast, this work proposes a framework for constructing FC and convolutional layers over computationally tractable Riemannian spaces. This framework incorporates several previous FC layers across different geometries as special cases and is instantiated on ten representative manifolds, including three hyperbolic models, five geometries of the symmetric positive definite (SPD) manifold, and two Grassmannian perspectives. Experiments on different manifolds demonstrate the effectiveness and applicability of our approach.

Most coding-agent benchmarks ask whether generated code behaves correctly. That remains essential, but repository-level engineering is increasingly agent-managed: one agent writes a repository, and later agents inspect, audit, or extend it as working context. In that setting, a generated repository is not only an answer to a task but also a communication artifact for future work. Even when strong agents nearly satisfy the visible behavioral objective, repositories can differ in how clearly they expose the intended behavior and design choices behind that behavior. We introduce Build-and-Find, a protocol for evaluating whether downstream agents can recover those intended choices from generated repositories, and how much inspection that recovery requires. For each task, a builder sees a hidden repository specification and creates a codebase; a finder sees only the codebase and a specification-traced multiple-choice question bank. The protocol separates behavioral correctness from artifact-side recovery and reports recovery accuracy, repeatability, implementation coverage, and inspection effort. Accuracy and stability act as gates: effort is interpreted only when recovery succeeds reliably. Among artifacts from which the same intent can be recovered, lower effort by the same finder suggests that the artifact makes that intent easier to locate. Question-only and spec-only controls quantify generic priors and specification access, while audits separate omitted claims from finder failures and check whether correct answers cite artifact evidence. In the released high-prior task pack, recovery accuracy is near saturation, so inspection effort and finder-specific effects provide the main panel-local comparison. We release the harness, tasks, generated artifacts, records, canonical tables, reports, scripts, metadata, licenses, and evidence audit to support auditable extensions and private task packs.

The design of modern neural architectures has converged through incremental empirical choices, yet the mechanisms governing their training dynamics remain only partially understood. We identify and analyze a negative weight drift induced by the interaction between standard losses and positively biased activation functions. We prove that under MSE or cross-entropy loss, the gradient with respect to positive pre-activations is non-negative in expectation at initialization, driving downstream weights toward negative values during early training. The drift is intrinsic to optimization rather than data, and persists across architectures (MLP, ResNet, ViT, GPT-nano, MP-SENe) and asymmetric activation functions (ReLU, GELU, SiLU). Coupled with ReLU, weight drift produces activation sparsity reaching up to 90\% in GPT-nano. We characterize the sparsity–accuracy tradeoff across 79 configurations and identify a sharp accuracy cliff above $\sim$70\% activation sparsity. While ReLU$^2$ achieves a good sparsity–accuracy tradeoff, it pathologically amplifies identified activation spikes in intermediate transformer layers. Clipping resolves this while preserving the representational benefits of squaring: clipped ReLU$^2$ outperforms its unclipped version, and GELU$^2$ achieves the lowest validation loss on GPT-nano.

Off-policy evaluation (OPE) estimates the value of a target policy from logged data, but limited behavior-policy coverage can force high-variance reweighting or reward-model extrapolation. Counterfactual annotations can add evidence about unobserved actions, yet practical sources, including domain experts and LLMs, may be costly, biased, or noisy. We study budgeted acquisition of such annotations for contextual-bandit OPE. Given source-specific costs and error profiles, we formulate an integer allocation problem over context-action pairs and annotation sources to minimize the component of estimator variance that depends on the annotation plan. We characterize when annotations are valuable through a first-annotation threshold and local annotation-value regimes. For the coupled multi-source problem, we develop a majorization-minimization algorithm with dynamic-programming subroutines that monotonically improves the objective. Experiments in synthetic clinical and LLM-annotated education bandits show that our proposed allocation method reduces the optimized variance component by $20.58$% and $10.77$%, respectively, relative to no-annotation baseline.

Real-world time series typically exhibit non-uniform information density across temporal positions, frequency structures, and forecasting horizons: critical dynamics are concentrated in local segments, different frequency components carry different structural information, and different horizons require different modeling complexity. However, most existing methods still allocate computation approximately uniformly across these dimensions, leading to redundant computation in low-information regions and insufficient modeling of key patterns. To address this, we propose BucpTSF, an information-density-driven structured framework for time series forecasting. Specifically, BucpTSF reorganizes raw sequences into hierarchically heterogeneous multi-level token representations through information-density-aware temporal representation reorganization, captures key cross-token dependencies with low overhead via frequency-selective structured token interaction, and further introduces horizon-conditioned structured extrapolation to adaptively balance dynamic modeling and structural extrapolation bias for different horizons, thereby improving the accuracy and stability of long-term forecasting. Extensive experiments on real-world datasets show that BucpTSF consistently outperforms strong baselines under various forecasting settings, validating the effectiveness of information-density-driven non-uniform computation allocation for time series forecasting.

Existing evaluation methods largely rely on clean, static benchmarks, which can overestimate true model performance by failing to capture the noise and variability inherent in real-world user inputs. This is especially true for language models, which can face human-generated text queries containing mistakes, typos, or alternative ways of phrasing the same question. In this work, we introduce a theoretical framework for quantifying model sensitivity to prompt variants, or brittleness, that can enable us to disentangle data-induced difficulty from prompt-related variability. Using this framework, we design a novel evaluation pipeline, Brittlebench, to holistically evaluate the sensitivity of frontier models. We apply semantics-preserving perturbations to a suite of popular benchmarks, and observe model performance to degrade as much as ~21%. However, these perturbations do not affect all models equally: even a single perturbation alters the relative ranking of models in 63% of cases, impacting conclusions about comparative model performance. Decomposing the total variance of both state-of-the-art open-weight and commercial models, we find that semantics-preserving input perturbations can account for up to half of the performance variance for a given model. Brittlebench highlights the need for more robust evaluations and models, and allows us to systematically understand model brittleness.

A fundamental challenge in science and engineering is the simulation-to-experiment gap. While we often possess prior knowledge of physical laws, these physical laws can be too difficult to solve exactly for complex systems. Such systems are commonly modeled using simulators, which impose computational approximations. Meanwhile, experimental measurements more faithfully represent the real-world, but experimental data typically consists of observations that only partially reflect the system’s full underlying state. We propose a data-driven distribution alignment framework that bridges this simulation-to-experiment gap by pre-training a generative model on fully observed (but imperfect) simulation data, then aligning it with partial (but real) observations of experimental data. While our method is domain-agnostic, we ground our approach in the physical sciences by introducing Adversarial Distribution Alignment (ADA). This method aligns a generative model of atomic positions—initially trained on a simulated Boltzmann distribution—with the distribution of experimental observations. We prove that our method recovers the target observable distribution, even with multiple, potentially correlated observables. We also empirically validate our framework on synthetic, molecular, and experimental protein data, demonstrating that it can align generative models with diverse observables.

Multiple instance learning (MIL) is the dominant framework for whole-slide image analysis in computational pathology, typically combining a frozen patch encoder, a projection layer, and a slide-level aggregator. While encoders and aggregators have been extensively studied, the projection layer remains a largely morphology-only bottleneck. This limits endpoints such as biomarker status and survival, which are governed by a molecular state that is not fully captured by H\&E morphology. We introduce Molecularly Informed Staining Transform (MIST), a plug-in replacement for the MIL projection layer that uses paired spatial transcriptomics only during training to construct virtual molecular stains. MIST clusters gene expression profiles into cross-modal prototypes, anchors them in the frozen foundation model feature space, and uses them to reorganize H\&E patch features along molecularly guided axes. It requires no transcriptomics at inference and can be inserted before standard MIL aggregators. We evaluate MIST across 23 downstream tasks and 8 MIL aggregators. MIST improves 240 of 256 configurations over the standard projection layer, with an average gain of +3.5\%, observed consistently across endpoint types: +5.2\% on survival prediction, +3.3\% on tissue subtyping, and +2.6\% on biomarker prediction. Ablations confirm that gene-derived prototypes are the primary source of the gains, while spatial, biological, and pathological analyses show that cross-modal prototype affinities capture spatially coherent molecular programs from H\&E alone.